<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mattia Rüfenacht on Medium]]></title>
        <description><![CDATA[Stories by Mattia Rüfenacht on Medium]]></description>
        <link>https://medium.com/@mattiasuisse?source=rss-37f341b69068------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*HIBxQlsdSpi4nPoC3k2kng.jpeg</url>
            <title>Stories by Mattia Rüfenacht on Medium</title>
            <link>https://medium.com/@mattiasuisse?source=rss-37f341b69068------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 23:39:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mattiasuisse/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[enlogen IT and Parashift bring automated document extraction to SMBs]]></title>
            <link>https://medium.com/trea-ventures/enlogen-it-and-parashift-bring-automated-document-extraction-to-smbs-c719e944ca33?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/c719e944ca33</guid>
            <category><![CDATA[serverless]]></category>
            <category><![CDATA[document-capture]]></category>
            <category><![CDATA[microsoft-power-automate]]></category>
            <category><![CDATA[process-optimization]]></category>
            <category><![CDATA[process-automation]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Tue, 02 Mar 2021 11:21:19 GMT</pubDate>
            <atom:updated>2021-03-02T11:21:19.684Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/650/1*TswODqIWAtea92TfS-MoFg.png" /></figure><p>Sissach, Switzerland 01.03.2021– enlogen IT, a young technology consulting and integration company from Berlin, and Parashift, a provider of machine learning-based software for the extraction of any document, have just announced a solution partnership. The partnership aims to enable or improve end-to-end automation in small and medium-sized businesses.</p><p>Together, enlogen IT and Parashift offer solutions that enable organizations to create more comprehensive and robust process automation, enabling them to reduce operational inefficiencies through the benefits of automated document capture and downstream automation.</p><p>“SMEs have rarely had access to economical solutions to process documents in an automated way, which severely limited the automation of downstream processes. Together with the competent team at enlogen IT, we can now change this and open up significant opportunities for end-to-end automation for them as well”, said Alain Veuve, CEO of Parashift.</p><p>“Parashift’s OCR solution fits perfectly with our “serverless” concept. Combined with Parashift’s strong team, we are able to iterate quickly and deliver great projects”, said Tristan Figge, CEO of enlogen IT.</p><p>The companies have already successfully implemented the first customer projects in the real estate sector. The resulting impact reduces the need for manual interactions in document-based tasks and creates valuable capacity to focus more on customers and partners. Further projects and use cases, which will be made possible with established technologies and Parashift’s versatile document extraction platform, are in development.</p><p>About enlogen IT:</p><p>enlogen IT GmbH, founded in 2021, is a technology consulting and integration company for small and medium-sized enterprises. The company is located in Berlin. Among enlogen’s core competencies is the optimization of business processes through customized IT solutions based on the “serverless” paradigm and the principle of “clever ideas brilliantly implemented”.</p><p>About Parashift:</p><p>Parashift AG, founded in 2018, is a deep-tech company focused on the autonomization of document extraction through machine learning. Parashift is domiciled near Basel, Switzerland, and employs about 30 people in machine learning engineering and development. Forbes named Parashift as one of 30 promising AI startups in Europe.</p><p>enlogen IT GmbH<br>Tristan Figge<br>CEO<br>Pappelallee 78/79<br>DE — 10437 Berlin</p><p><a href="/service/mailto:tfigge@enlogen.de">tfigge@enlogen.de</a></p><p><a href="https://enlogen.de/">https://enlogen.de/</a></p><p>Parashift AG<br>Mattia Rüfenacht<br>Business Development &amp; Press Manager<br>Hauptstrasse 134<br>CH — 4450 Sissach</p><p>Phone: +41 61 508 77 77<br><a href="/service/mailto:mattia.ruefenacht@parashift.io">mattia.ruefenacht@parashift.io</a></p><p><a href="http://www.parashift.io/">www.parashift.io</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c719e944ca33" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/enlogen-it-and-parashift-bring-automated-document-extraction-to-smbs-c719e944ca33">enlogen IT and Parashift bring automated document extraction to SMBs</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Handwritten text recognition in 2020]]></title>
            <link>https://mattiasuisse.medium.com/handwritten-text-recognition-in-2020-634984e29ac7?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/634984e29ac7</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[document-capture]]></category>
            <category><![CDATA[ocr]]></category>
            <category><![CDATA[icr]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Fri, 22 Jan 2021 09:04:51 GMT</pubDate>
            <atom:updated>2021-01-22T09:04:51.104Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/711/0*XtGxkcmsC35GM-9n.jpg" /></figure><p>The Optical Character Recognition (OCR) respectively Document Capture software market size is expected to be <a href="https://blog.parashift.io/blog/the-market-for-document-extraction-software-is-gaining-ever-more-relevance">USD 12.6 billion by 2027 with a year on year growth of approx. 10%</a>. The increasing demand for document extraction software solutions is mainly due to compliance initiatives, the digitalization of document management and intentions to optimize operational costs. But also because of the increasingly mobile workforce and the often associated need for document processing via mobile devices. Although OCR has been considered a solved problem by a wide range of people, there are yet many challenges to solve. Handwriting Recognition or Handwritten Text Recognition (HTR) being one of them. The high variance in handwriting styles across people and poor quality of the handwritten text compared to machine-printed text pose significant obstacles in converting it to machine-readable text formats. Nevertheless, it’s a crucial problem to solve for multiple industries like healthcare, insurance and banking and therefore, Parashift’s R&amp;D activities in full swing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/562/0*mJpsCxku41fopP9v.png" /></figure><p>Source: <a href="https://www.researchgate.net/figure/Handwriting-sample-from-for-NIST-SD19-a-Handwritten-sample-form-b-Images-of_fig1_267957644">ResearchGate</a></p><p>Relatively recent advancements in Deep Learning such as the advent of transformer architectures have fast-tracked our progress in solving handwritten text recognition. In the industry, we usually speak of Intelligent Character Recognition (ICR) when talking about recognizing handwritten text. This due to the fact that the algorithms needed to solve ICR need much more “intelligence” than solving generic OCR.</p><p>In this article, we will be learning about the task of handwritten text recognition, its intricacies and how we can solve it using different deep learning techniques.</p><h3>Main challenges in recognizing handwritten text</h3><ul><li>Tremendous variability and ambiguity of strokes from person to person</li><li>Inconsistency in writing style</li><li>Poor quality of the source document due to degradation over time</li><li>Oftentimes text is not written in a strictly straight line</li><li>Cursive handwriting makes separation and recognition of characters even more challenging</li><li>Text in handwriting can have a variable rotation</li><li>Collecting a high-quality dataset with labels is relatively expensive</li></ul><h3>Use cases</h3><h4>Healthcare and pharmaceuticals</h4><p>One of the main pain points in the healthcare and pharmaceutical industry is the digitalization of patient prescriptions. For example, Roche is handling an incredible amount of medical PDFs every day. Also the patient enrollment and other form digitalization cases are requiring a reliable solution to recognize and extract handwritten text. So, by adding handwriting recognition to their operations and business processes, healthcare and pharmaceutical institutions can significantly improve their process efficiency and customer experiences.</p><h4>Insurance</h4><p>Looking at the insurance sector, a large insurance company can easily receive a couple of millions of documents on a single day and a delay in processing, for example, claims can impact the company in a very meaningful way. The claim documents which can contain various different handwriting styles are regularly processed purely manually. Hence, the automation of claims processing has a major bottleneck here.</p><h4>Banking</h4><p>I know it may be hard to believe but people actually still write cheques on a regular basis and so these kinds of documents still play a major role in most non-cash transactions. In some countries more than in others of course. Having said that, in many developing countries, the current cheque processing procedure requires a bank employee to read and manually enter the information present on a cheque and also verify the entries like, for example, signature and date. As in most cases, these cheques are being centralized for processing, it’s quite common that a large number of such cheques have to be processed day in and day out. In other words, here again, a powerful handwriting text recognition system could save significant hours of human work and bring about dramatic cost savings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/388/0*hzzAwHnBfPE_1TIm" /><figcaption>Source: <a href="https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.651.6458&amp;rep=rep1&amp;type=pdf">The Pennsylvania State University</a></figcaption></figure><h3>Online Libraries</h3><p>Another area where HTR comes into play is in online libraries. Cause here, incredible amounts of historical writings are being digitized by uploading scans, creating access for people all around the world. However, because of the huge amount of data, such endeavors are only truly useful when one can identify the text on these scans, index them and make queries, etc.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/567/0*1oKhfnwurQDig8fu.jpg" /><figcaption>Source: <a href="https://www.akg-images.de/archive/Sachsenspiegel-2UMDHUCSHGJ.html">akg-images</a></figcaption></figure><h3>Methods</h3><p>Broadly speaking, handwriting recognition methods can be classified into two different types:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/655/0*WSF39VOnVCw844z-.png" /></figure><p>When we do<strong> online handwriting </strong>recognition, the methods involve a digital pen and through that, we have access to the stroke information and pen location while text is being written. The picture above is a very good illustration of this and should also indicate that for these writings, we usually have a good amount of information which makes it a whole lot easier to classify characters with higher accuracy.</p><p><strong>Offline methods</strong>, on the other hand, involve recognizing text once it’s written down and based on this information only. Thus, we have much more scarce features to use for making predictions. Even worse, we might also have to deal with background noise emanating from the paper.</p><p>If in real-world scenarios it wouldn’t so hard to properly scale the digital pen approach, the task of recognizing handwritten text wouldn’t be that big of a deal anymore. But as it is just the way it is, we have to find solutions that help us make sense of the information we have provided by non-digital pens. That’s why we now will look at various techniques to solve the problem of recognizing offline written text.</p><h3>Some of the techniques</h3><p>The first approaches to solving handwriting recognition involved Machine Learning methods such as well know <a href="https://en.wikipedia.org/wiki/Hidden_Markov_model#:~:text=Hidden%20Markov%20models%20are%20generative,emission%20probabilities)%2C%20is%20modeled.">Hidden Markov Models</a> (HMM), <a href="https://en.wikipedia.org/wiki/Support_vector_machine">Support Vector Machines</a> and the like. These approaches focused on feature extraction after an initial pre-processing of the text. These features could for example be things like loops, inflection points, ratios of some sort and other aspects of a single character. Once extracted, the features are fed to a classifier such as the HMM with the aim of getting good results. However, such approaches are heavily relying on manual work for feature extraction akka not scalable and are pretty limited in terms of learning capacity. As a result, the performance of these kinds of Machine Learning models is rather limited.</p><p>But there is good news: as touched upon at the beginning, Deep Learning has opened up a completely new spectrum of possibilities to tackle the existing problems and brought about tremendous improvements in accuracy for handwriting recognition. In the following, I will go into some of the most prominent research in this particular field.</p><h4>Multi-dimensional Recurrent Neural Networks</h4><p><a href="https://en.wikipedia.org/wiki/Recurrent_neural_network">Recurrent Neural Networks</a> (RNN) or <a href="https://en.wikipedia.org/wiki/Long_short-term_memory">Long Short-Term Memory</a> (LSTM), which is a specific RNN architecture, are known to be able to deal with sequential data to identify temporal patterns and generate results. At least the most common configurations of these kinds of architectures. Nevertheless, there’s a significant limitation: they can only handle 1D data and as a consequence, these architectures can’t be applied to image data directly. To address this problem, the authors of the paper “<a href="https://sci-hub.tw/10.1007/978-1-4471-4072-6_12">CITlab ARGUS for Arabic Handwriting</a>” proposed a multi-dimensional RNN/LSTM structure as visualized below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*YK23uw5T-VO3OJhC.png" /></figure><p>So, what is the difference between a traditional RNN and the proposed multi-dimensional RNN… Usually, in general RNNs, hidden layer i receives state from a previous hidden layer in time i-1. But in cases of multi-dimensional RNNs, let’s take the example of a 2-dimensional RNN structure, the hidden layer (i, j) receives states from multiple previous hidden layers, being (i-1, j) and (i, j-1). Therefore, such architectures capture context from both height and width in an image which is pivotal for getting a clear understanding of the local region. You can further extend that to also get information from future layers. So, this is pretty similar to how Bidirectional Long Short-Term Memory (BI-LSTM) networks receive information from both t-1 and t+1. To get back to our hypothetical 2-dimensional RNN structure, here we would now be able to receive information in both directions as well, e.g. (i-1, j), (i, j-1), (i+1, j), (i, j+1) and thus, capturing context in all directions just as with the BI-LSTM.</p><p>What you can see in the figure below is basically the entire multi-dimensional RNN structure. So, what we do is simply replace the RNN block with an LTSM block. The network’s input is divided into blocks of size 3x4 which is fed into the multi-dimensional RNN layers respectively the network’s hierarchical structure of multi-dimensional RNN layers followed by feed-forward layers in tandem. Feed-forward layers being classical artificial neural networks. The final output is converted into a 1D vector and is passed to a Connectionist Temporal Classification (CTC) function to generate an output.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/599/0*graJiQdx8ufjt3mS.png" /></figure><p>CTC is an algorithm used to deal with tasks such as speech recognition and handwriting recognition. Both of them are tasks where just the input data and the output transcription is available and no alignment details are provided. Meaning, how a particular region in audio for speech or a particular region in images for handwriting is aligned to a specific character. The issue here is that simple heuristics such as giving each character the same area won’t work as the amount of space each character takes varies greatly in handwriting styles. A good resource for more information on how CTC works can be seen <a href="https://distill.pub/2017/ctc/">here</a>.</p><p>When we decode the output of a CTC using simple heuristics of the highest probability for each position, we might get results which might not make lots of sense in real-world cases. That’s why we might have to use another kind of decoder which is more promising to improve our results. In order to get a better understanding of the decision-making, let’s consider the different types of decoding there are:</p><ul><li><strong>Best-path decoding</strong> is the generic decoding we have implicitly discussed so far. In this type, at each position, we take the output of the model and simply find the result with the highest probability.</li><li>When shooting for <strong>beam search decoding</strong>, the <a href="https://en.wikipedia.org/wiki/Beam_search">beam search</a> suggests keeping multiple output paths with the highest probabilities and thereby expands the chain with new outputs while dropping paths with lower probabilities to keep the beam size constant.</li><li>Beam search usually provides more accurate results than grid search. But there is still more room to get truly meaningful results. A way to aim for greater performance is to use a <strong>language model alongside beam search</strong> using both probabilities from the model and the language model (which, for example, scores character-sequences according to probabilities) to generate the final results.</li></ul><p>More details about generating accurate decoding results can be consulted in this <a href="https://towardsdatascience.com/word-beam-search-a-ctc-decoding-algorithm-b051d28f3d2e">article.</a></p><h4>Encoder-decoder and attention networks</h4><p><a href="https://en.wikipedia.org/wiki/Seq2seq">Seq2Seq models</a> (sequence to sequence) having encoder-decoder networks have recently been popular for solving the tasks of speech recognition or also machine translation and have been applied to solve the use case of HTR by deploying an additional attention mechanism. Let’s take a closer look at some recent research in this area.</p><p>Scan, attend and read</p><p>In the paper “<a href="https://arxiv.org/abs/1604.03286">Scan, Attend and Read: End-to-End Handwritten Paragraph Recognition with MDLSTM Attention”</a> the authors propose the usage of an attention-based model for end-to-end multi-line handwriting recognition. The main contribution of the research is the automatic transcription of text without having to segment it into lines as a pre-processing step and as a result, can simply scan an entire page and provide results directly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*v1hiawMjeYE6huGK.png" /></figure><p>Scan, attend and read (SAR) uses a multi-dimensional long-term short-term memory (MDLSTM)-based architecture similar to the one we already discussed before with one difference in the final layer. After the last linear layer, meaning the final sum block in the figure above, the feature maps are collapsed in the vertical dimension and a softmax function is applied to get the outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/599/0*2qiM-dqGszxymmRP.png" /></figure><p>The SAR architecture consists of an MDLSTM architecture which acts as the feature extractor. The final collapsing module with an implemented softmax output and CTC loss is replaced by an attention module and an LSTM decoder. The attention model applied is a hybrid combination of content-based attention and location-based attention. What this really means can be read in the paper we will discuss next. In general, attention mechanisms let a model directly look at, and draw from, the state at any earlier point, allowing it to focus on the most relevant encoded features at each decoding step. The decoder LSTM modules, on the other hand, take the previous state, previous attention map and the encoder features in order to generate an output character and the state vector for the next prediction.</p><p>Convolve, attend and spell</p><p>The paper “<a href="http://www.cvc.uab.es/~marcal/pdfs/GCPR18.pdf">Convolve, Attend and Spell: An Attention-based Sequence-to-Sequence Model for Handwritten Word Recognition</a>” proposes an attention-based sequence-to-sequence model. The architecture proposed has three main parts: an encoder, consisting of a <a href="https://en.wikipedia.org/wiki/Convolutional_neural_network">convolutional neural net</a> (CNN) and a bi-directional <a href="https://en.wikipedia.org/wiki/Gated_recurrent_unit">GRU</a>, an attention mechanism devoted to focusing on the pertinent features and a decoder formed by a one-directional GRU, able to spell the corresponding word, character by character.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/599/0*_ZPs0pz6jWpjWLRm.png" /></figure><p>The encoder uses a CNN to extract low-level visual features. Specifically, a pre-trained VGG-19-BN (basically a 19 layers deep network, batch normalized) architecture is used as a feature extractor. The input image is thereby converted into feature map X which is then reshaped into X’ by splitting all channels column-wise and combining them to get the sequential information. X’ is further converted to H by using a bi-directional GRU. GRU is a neural network similar to LSTM in nature and can capture temporal information.</p><p>Further, an attention model is applied while predicting the output from the decoder, an RNN decoding one character at each time step, thus spelling the whole word. The authors discuss two different types of attention mechanisms they explored.</p><ul><li><strong>Content-based attention:</strong> The idea behind this methodology is to find the similarity between the current hidden state of the decoder and the feature map from the encoder. We can find the most correlated feature vectors in the feature map of the encoder, which can be used to predict the current character at the current time step. A more intuitive way of looking at how attention mechanisms work can be found <a href="https://jalammar.github.io/visualizing-neural-machine-translation-mechanics-of-seq2seq-models-with-attention/">here</a>.</li><li><strong>Location-based attention:</strong> The main disadvantage of location-based mechanisms is that there is an implicit assumption that the location information is embedded in the output of the encoder. Otherwise, would be no way to differentiate between character outputs which are repeated from the decoder. But let’s look at an example to make things more clear. Let’s assume we have the word Parashift, in which the character a is repeated twice. Without location information, the decoder cannot predict them as separate characters. To address this problem, the current character and its alignment are predicted by using both the encoder output and previous alignment. If you are curious about how the methodology works in greater detail, you can dig deeper <a href="https://arxiv.org/abs/1506.07503">here</a>.</li></ul><p>The decoder is a one-directional multi-layered GRU. At each time step t it receives input from previous time step and the context vector from the attention module. Multinomial Decoding and label smoothing are explored at training to improve generalization capability.</p><h4>Transformer models</h4><p>So, encoder-decoder networks have been astonishingly good in handwriting recognition. Nevertheless, these kinds of networks yet do have a significant bottleneck in the training process due to the LSTM layers involved. Therefore, training can’t run in parallel. Not too long ago, however, transformer models have proven to be very promising and actually replaced LSTM structures in solving some language-related tasks. Thus, the question arises, how can we apply these to our handwriting recognition tasks?</p><p>Pay attention to what you read</p><p>In the paper “<a href="https://arxiv.org/abs/2005.13044">Pay Attention to What You Read: Non-recurrent Handwritten Text-Line Recognition</a>”, the authors introduce a non-recurrent approach by instead using a transformer-based architecture and multi-headed self-attention layers at both visual and textual stages. Thereby, they can learn both character recognition as well as language-related dependencies of the character sequences to be decoded. Since the language knowledge is embedded into the model itself, there is no need for any additional post-processing step using a language model. Or in other words, no vocabulary is necessary to be able to predict outputs. In order to achieve this, text encoding happens at a character and not a word level. As the transformer architecture allows parallelized training for every region or character, the training process is much simpler.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/598/0*9wTlQdfQJgRV2uLR.png" /></figure><p>The network architecture proposed in the paper consists of two major components. A visual encoder, extracting relevant features and applying multi-headed visual self-attention on different character locations and a text transcriber, which takes text input, encodes the text, applies multi-headed language self-attention and mutation attention on both visual and textual features.</p><p><strong>Visual encoder</strong></p><p>The authors use the ResNet50 as a backbone convolutional architecture to extract the features (visualized in the figure above). The 3-dimensional feature representation from ResNet50 is passed to a Temporal Encoding (TE) module which reshapes to 2D by keeping the same width and thus, the shape of (f x h, w). This is fed into a fully connected layer to reduce the shape to (f, w) and the resultant output is Fc’, which can be seen as a w-length sequence of visual vectors. In addition, a positional encoding TE is added to Fc’ to retain the position information. The output is passed through a fully connected layer to get the final feature representation with shape (f , w). The final output is passed through a multi-headed attention module with 8 heads to get a visually rich feature map. To learn more about transformer models, click <a href="http://jalammar.github.io/illustrated-transformer/">here</a>.</p><p><strong>Text transcriber</strong></p><p>The second part of the proposed methodology is a text transcriber. There, the text is passed through an encoder which generates character-level embeddings. These embeddings are combined with temporal location similar to the way in the visual encoder, using a TE module. The values are then passed to a multi-head language self-attention module, which is again similar to the module applied in the visual encoder. The text features generated along the visual features from the visual encoder are passed to a mutual-attention module whose sole task is to align and combine the learned features from both images and the text inputs. Finally, the output is passed through a softmax function.</p><h4>Handwriting text generation</h4><p>So get back to a more general level again, let’s look at handwriting text generation. As you may guess, this is all about generating synthetic handwritten text or in other words, augmenting existing datasets, which are scarce. To properly go about such kinds of meaningful data augmentations, we can make use of <a href="https://en.wikipedia.org/wiki/Generative_adversarial_network">generative adversarial networks</a>.</p><p><strong>ScrabbleGAN</strong></p><p><a href="https://arxiv.org/abs/2003.10557">ScrabbleGAN</a> is a proposed method that follows a semi-supervised approach to synthesize handwritten text images that are versatile in both style and lexicon. The approach relies on a novel generative model which can generate images of varying length. The generator also can manipulate the resulting text style which allows us to decide whether the text has to be cursive or say how thick or thin the stroke should be.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*I1niluaxVZS6Gizq.png" /></figure><p>The architecture consists of a fully convolutional generator based on BigGAN. For each character in the input, a corresponding filter is selected from a filter-bank before all the values are concatenated together and then multiplied by a noise vector z that controls the generated text style. As illustrated in the figure above, the regions generated for each individual character overlap, helping in generating connected recursive text as well as allowing the flexibility of different character sizes. As an example, “m” takes up quite a bit of space while the “e”s not so much. To keep the same style for the entire word or also sentence, the noise vector z is kept constant for all characters.</p><p>To classify whether the generated style of an image is fake or real, a convolutional discriminator based on a BigGAN architecture is implemented. As the discriminator is not relying on character-level annotations, it is not based on a class conditional GAN. The advantage of this is that there is no need for labeled data and so even unlabeled data from unseen corpora can be used for training the discriminator. Along with the discriminator a text recognizer R is trained to classify if the text generated is real or if it’s gibberish. The recognizer is based on a convolutional recurrent neural network (CRNN) architecture. The text generated in the output of R is then compared with the input text given to the generator and a corresponding penalty is added to a loss function.</p><p>Some of the outputs generated by ScrabbleGAN are displayed here:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*dklBJkYJJs-vqZfQ.png" /></figure><h3>Summary</h3><p>We have experienced many fundamental breakthroughs in the past years and yet, despite all this progress in the underlying technologies, HTR is far from being a solved problem. However, given recent trends on how fast we can move forward, this might change relatively soon. And just to mention that as well… Already today we can provide companies meaningful technology that can help reduce manual efforts significantly.</p><p>If you’re curious to learn more about Parashift’s functionalities in that regard, feel free to reach out. We are happy to give you a more in-depth introduction to what our extraction engine is capable of dealing with.</p><p><a href="https://parashift.io/">&gt;&gt;&gt;&gt; https://parashift.io/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=634984e29ac7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Parashift teams up with UiPath to bring superior document intelligence into RPA projects]]></title>
            <link>https://medium.com/trea-ventures/parashift-teams-up-with-uipath-to-bring-superior-document-intelligence-into-rpa-projects-55d95e222044?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/55d95e222044</guid>
            <category><![CDATA[uipath]]></category>
            <category><![CDATA[ocr]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[rap]]></category>
            <category><![CDATA[document-capture]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Fri, 22 Jan 2021 08:47:36 GMT</pubDate>
            <atom:updated>2021-01-22T08:47:36.994Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/711/0*zNbuLhaeJL5fK7iB.png" /></figure><p><em>Robotic Process Automation is a powerful technology to automate tedious and repetitive tasks. To unleash its full potential, however, accurate, fast and cheap document extraction is indispensable. The partnership of Parashift with industry leader UiPath is therefore set to drive further adoption.</em></p><p><strong>Sissach, Basel-Landschaft (Switzerland), 22.09.2020</strong> — Parashift, a provider of modern, machine learning-powered document extraction software, has just announced an advanced technology partnership with UiPath, the leading enterprise Robotic Process Automation (RPA) software company, meant to improve end-to-end automation across industries worldwide. Together, Parashift and UiPath offer a solution that enables more comprehensive process automations and empowers businesses to increase operational efficiency, bringing together the benefits of RPA and Document Capture to drive meaningful business outcomes.</p><p><em>“Businesses who are desperately looking for ways to reduce costs, increase productivity, process continuity and system efficiency are often turning to RPA. But the technology is only as good as the data provided. Since, according to Gartner, over 80% of corporate data is unstructured, solving the extraction problem is absolutely crucial. We are therefore very pleased to be working with UiPath to help drive digitalization forward,” said Alain Veuve, CEO of Parashift.</em></p><p>UiPath has a vision to deliver A Robot for Every Person, one where companies enable every employee to use, create, and benefit from the transformative power of automation to liberate the boundless potential of people. Only UiPath offers an end-to-end platform for hyperautomation, combining the leading <a href="https://www.uipath.com/rpa/robotic-process-automation">Robotic Process Automation</a> (RPA) solution with a full suite of capabilities that enable every organization to scale digital business operations at unprecedented speed. The company has already automated millions of repetitive tasks for over 65% of the Fortune 500 and 8 of the Fortune 10.</p><p>Parashift’s versatile document extraction platform can be integrated in UiPath RPA projects using the <a href="https://connect.uipath.com/marketplace/components/parashift-activities">Parashift Activity Pack</a>. By combining these two best-of-breed software components, businesses can implement more diverse and robust business process automations. The resulting impact significantly reduces or even completely eliminates the necessity for manual interactions in document based tasks. So, the integration of the Parashift Platform extends the value of UiPath’s RPA technology by providing a cheaper and more resilient way to convert structured and unstructured data into machine-readable data, fueling the possibilities for hyperautomation across the organization.</p><p>To learn more about UiPath, please visit <a href="https://www.uipath.com/">www.uipath.com</a>.</p><p><strong>About Parashift</strong></p><p>Parashift AG, founded in 2018, is a deep-tech company focusing on the autonomization of document extraction through machine learning. Parashift is based near Basel (Switzerland) and employs around 20 people in the field of machine learning engineering and development. Forbes named Parashift 2019 as one of 30 promising AI startups in Europe.</p><p><a href="https://parashift.io/">https://parashift.io/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=55d95e222044" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/parashift-teams-up-with-uipath-to-bring-superior-document-intelligence-into-rpa-projects-55d95e222044">Parashift teams up with UiPath to bring superior document intelligence into RPA projects</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Even with legacy software can be digitized without a major project]]></title>
            <link>https://medium.com/trea-ventures/even-with-legacy-software-can-be-digitized-without-a-major-project-bd1d46b54539?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd1d46b54539</guid>
            <category><![CDATA[rap]]></category>
            <category><![CDATA[legacy-systems]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Thu, 21 Jan 2021 17:04:21 GMT</pubDate>
            <atom:updated>2021-01-21T17:04:21.497Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/712/0*_FbkQH6jHnHu2UBE.jpg" /></figure><p>In the course of digitalization, customer requirements have changed significantly. Everything must be available and retrievable immediately. Whenever possible, things should be in realtime. Speed and adaptability to individual preferences are other essential trends. In order to keep pace with these developments and the needs of customers, companies must live through these changes and be at the cutting edge of technology.</p><p>This is easier for young companies whose offerings are based on modern technology and the latest software and hardware. For larger companies, on the other hand, especially those that have been in business for several decades and often work with various legacy systems as a result, this represents a major challenge. Oftentimes, their hardware and software setups are connected with immense investments that they have had to make in order to gain or maintain competitive advantages. However, since information technologies have relatively short innovation cycles, the cost grind downs on these assets are considerable. In other words, a technology acquired yesterday will be available tomorrow at a significant discount. Ongoing investment calculations of previously acquired technology assets therefore clearly tend to hinder the ongoing digital transformation. Simply increasing the budget because of it is for most executives associated with a lot of company policy and laborious persuasion. The question therefore arises: How can companies with legacy systems still take action and take decisive measures in the technological change?</p><h4>What legacy systems are and why they still exist</h4><p>Companies started to implement the first computers and mainframes (so-called <a href="https://www.it-business.de/was-ist-ein-mainframe-a-919119/">mainframe computers</a>) and build software on them as early as the 1970s and 1980s. On the commercialized hardware at that time, people were able to quickly achieve very large performance advances with iterations that better met the needs of institutional customers. Driven by the spirit of invention and the speed of momentum, much more complex and powerful systems could quickly be created. The largest supplier in the mainframe market has for decades been the American company IBM. The AS-400 system introduced by IBM in 1988 was widely used and characterized by its stability and reliability. Well, hang on. Even today, some large insurance companies and banks still have software that runs on these AS-400 machines. One of the reasons for this is that replacing these outdated systems is extremely time and cost-intensive. An upgrade can take months, if not years, depending on the scope of the systems and the necessary change projects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/239/0*rZng544doQqthNzp.png" /><figcaption>First edition of an AS-400 from IBM (source: <a href="https://commons.wikimedia.org/wiki/File:IBM_AS-400.JPG)">Wikimedia</a>)</figcaption></figure><h3>Difficulties with interfaces</h3><p>A major problem of such legacy systems is that creating interfaces is extremely difficult. This is because software from that time has no standardized interfaces for capturing or reading data. As a consequence, the connection of a modern API is an extraordinary challenge.</p><h4>Missing talent</h4><p>Another problem of legacy systems is that they are based on old programming languages. This means that newly trained IT specialists often have no use for these systems. They simply lack an understanding of the languages and concepts used. And the learned modern programming languages such as Java, Python, Ruby on Rails, etc. are not really helpful to make derivations and be able to deal with the old frameworks.</p><p>In addition, legacy systems in most cases have no technical specifications and no documentation, which makes it difficult to create an alternative with identical functionalities. An additional key aspect is that employees are accustomed to the old systems, which means that a new system would also require training, which in turn has a negative impact on capacity and monetary resources.</p><h4>RPA as a solution</h4><p>The problems described underpin the challenge that many companies are struggling with and often lead the management to believe that a complete upgrade to modern systems is not the solution. But good news: With Robotic Process Automation (RPA), legacy systems can be upgraded with relatively little effort.</p><p>When using RPA, the system is not operated via interfaces, which as mentioned is one of the weak points of these systems, but via a virtual robot that can operate the user interface. Just like you do. The robot can therefore imitate the clicks and sequences that are otherwise made by employees.</p><p>Years of development have resulted in the user interfaces of AS-400 machines working cleanly. For example, if a human or a robot enters an amount without a unit, an error message is displayed. The process then runs under an exception, which in the end usually has to be viewed by a human. In this way, incorrect entries can be almost completely prevented. If you organize these transactions via a direct interface instead, such error messages are handled differently and errors occur more quickly. As you can see, implementing RPA is not only easier and cheaper than setting up an interface to legacy systems, but also brings additional benefits.</p><h4>What advantages RPA has to offer</h4><p>Since robots can imitate humans to execute almost all repetitive and rule-based processes, RPA is relatively easily scalable and can also be adapted to different requirements with practically no additional infrastructure.</p><p>RPA is also extremely constructive when it comes to safety. Compared to other approaches to avoid a complete replacement of legacy systems, such as outsourcing or offshoring, RPA offers a higher level of security. Cause with RPA, companies are still in full possession of sensitive company data and have control over all access authorizations.</p><p>It is obvious that RPA offers a number of advantages for improving legacy systems. But while implementation is fast and relatively inexpensive, successful adoption of this technology requires careful planning. In addition to identifying opportunities and deciding which processes to automate, choosing the right RPA vendor to guide you and your organization along the path of automation is critical to the hoped-for success of the automation initiative.</p><h4>With RPA into the future</h4><p>With our trusted partner and RPA specialist <a href="http://esgroup.ch/"><strong>ESGroup</strong></a>, you have an ideal partner in the process of your RPA initiative and can build upon a team with many years of experience in software development. If you want to take the initiative and make your company fit for the next upcoming years, the best way to get in touch is to contact the team directly by email at <a href="/service/mailto:zurich@esgroup.ch">zurich@esgroup.ch</a> or call +41 42 321 99 82 for a non-binding exchange. Good luck!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd1d46b54539" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/even-with-legacy-software-can-be-digitized-without-a-major-project-bd1d46b54539">Even with legacy software can be digitized without a major project</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Parashift announces US market expansion]]></title>
            <link>https://medium.com/trea-ventures/parashift-announces-us-market-expansion-22fec2993f2b?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/22fec2993f2b</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[document-management]]></category>
            <category><![CDATA[ocr]]></category>
            <category><![CDATA[document-extraction]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Thu, 21 Jan 2021 17:01:49 GMT</pubDate>
            <atom:updated>2021-01-21T17:01:49.889Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/711/0*0WDZ57NuiwqZxRS1.png" /></figure><p><strong>Switzerland, 30.10.2020</strong> — Parashift, provider of a first of its kind modern, machine learning-powered document extraction cloud platform, has just announced its formal entry to the US market.</p><p>The worldwide market for Document Capture generates roughly $6 billion in revenue per year and is projected to grow around 10% annually over the next seven years. Together, Europe and the US accounted for more than 50% market share in 2019. The US market, specifically, known for its maturity in digital transformation and technological innovation, is set to outperform the projected global growth significantly.</p><p>Parashift, whose name implies a revolutionary “paradigm shift” in document processing, develops cloud solutions that lead the market technologically by their innate ability to interpret a wide range of document types over the same endpoint, thanks in part to “swarm” AI learning capabilities. This enables organizations to capture the 80% share of documents previously thought impossible to process at low cost and with seamless onboarding. Parashift estimates this type of technology and processing paradigm has a global market potential of around $15 billion.</p><p>For Parashift’s long-term success, access to global talent and proximity to customers and partners is of fundamental importance, achievable only via having a footprint in the most relevant markets. Formalizing and increasing its foothold in the US is therefore a strategic decision that aims to assist organizations around the country in realizing their vision of business transformation, ultimately resulting in greater value generation for them and their customers.</p><p>“As one of the leading global technology ecosystems, the US West coast provides a host of advantages. We are very happy that Edward Isarevich is joining the company and supporting our process of expansion in the US market. This initiative will further enable us to create a more cohesive partner network and strengthen our capabilities globally,” said Alain Veuve, Founder and CEO of Parashift.</p><p>“Parashift solves a mission-critical aspect of today’s everyday business and digital transformation process in a meaningful way, helping businesses cope with some of the most pressing challenges they face. I’m enthusiastic about joining forces with the incredibly capable and proven team at Parashift to help accelerate its impact in a geographically crucial global market,” said Edward Isarevich, US Representative at Parashift.</p><p><strong>About Parashift</strong></p><p>Parashift AG, founded in 2018, is a deep-tech company focusing on the autonomization of document extraction through machine learning. Parashift is headquartered in Switzerland and aggregates expertise in the field of machine learning engineering and platform development. Forbes named Parashift 2019 as one of 30 promising AI startups in Europe.</p><p>Parashift AG<br>548 Market St, PMB 59616<br>San Francisco, California<br>United States</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22fec2993f2b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/parashift-announces-us-market-expansion-22fec2993f2b">Parashift announces US market expansion</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[alabus and Parashift Enable AI-Based Data Extraction for the Insurance Market]]></title>
            <link>https://medium.com/trea-ventures/alabus-and-parashift-enable-ai-based-data-extraction-for-the-insurance-market-fdbb529f056a?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/fdbb529f056a</guid>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Thu, 21 Jan 2021 17:00:28 GMT</pubDate>
            <atom:updated>2021-01-21T17:00:28.819Z</atom:updated>
            <content:encoded><![CDATA[<h3>alabus and Parashift enable AI-based data extraction for the insurance market</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*fI0mxCsF9O2nZysb.png" /></figure><p><strong>Sissach, Basel-Landschaft, Switzerland, 03.11.2020</strong> — alabus ag is a leading Swiss manufacturer of standard software for the insurance market. A large number of well-known customers trust in the expertise of alabus, especially when it comes to optimizing their document centered business processes.</p><p>The alabus team, which consists of IT and process management experts, combines many years of experience in the areas of process optimization and automation as well as information technologies with profound industry knowledge. Thus, alabus clients receive holistic consulting and close support and guidance in order to successfully navigate the digital transformation.</p><p>A comprehensive number of processes modeled with alabus software include documents that need to be processed as quickly and cost-effectively as possible in order to reduce throughput times and costs and improve customer-oriented services. Thus, the integration of Parashift enables to design process automation in a consistent way and to tap more potential from existing infrastructure.</p><p>«alabus is a customer-oriented and very professional software company in the field of process management, which manages to create significant added value for customers in the long term. We are very pleased to be working with alabus to further digitalize the insurance market,» said Alain Veuve, CEO of Parashift.</p><p>«The Parashift Platform, through its integration with alabus solutions, provides us with a simple and consistent solution to automatically capture business documents and avoid media breaks. We are convinced that we will be able to achieve a great added value for our customers,» says Uwe Störrlein, Co-CEO of alabus.</p><p>The strategic partnership aims to create decisive competitive advantages for the individual needs of the insurance market by offering a complementary service.</p><p><strong>About alabus</strong></p><p>alabus is a software provider and specialist for process management in the Swiss insurance market. Thanks to profound experience and highly specialized expertise, alabus is a strong partner that has been successfully accompanying customers for over 20 years on their way to a holistic, digital future.</p><p><a href="http://www.alabus.com/">www.alabus.com/</a></p><p><strong>About Parashift</strong></p><p>Parashift AG, founded in 2018, is a deep-tech company focusing on the autonomization of document extraction through machine learning. Parashift is based near Basel (Switzerland) and employs around 30 people in the field of machine learning engineering and development. Forbes named Parashift 2019 as one of 30 promising AI startups in Europe.</p><p><a href="https://parashift.io/">https://parashift.io/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fdbb529f056a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/alabus-and-parashift-enable-ai-based-data-extraction-for-the-insurance-market-fdbb529f056a">alabus and Parashift Enable AI-Based Data Extraction for the Insurance Market</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[More Automation in Logistics With Intelligent Document Capture]]></title>
            <link>https://medium.com/trea-ventures/more-automation-in-logistics-with-intelligent-document-capture-77135062e616?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/77135062e616</guid>
            <category><![CDATA[rap]]></category>
            <category><![CDATA[document-capture]]></category>
            <category><![CDATA[document-extraction]]></category>
            <category><![CDATA[logistics]]></category>
            <category><![CDATA[ocr-software]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Tue, 28 Jul 2020 15:33:11 GMT</pubDate>
            <atom:updated>2020-07-28T15:34:31.264Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jzXEFTM-MNLv6OtduDMMGQ.png" /></figure><p><em>This article was originally posted on </em><a href="https://parashift.io/"><em>https://parashift.io/</em></a><em>.</em></p><p>The way in which we handle the purchase and sale of services and products has changed considerably in recent years. Especially the restrictions and extraordinary circumstances caused by COVID-19 have clearly shown in many places what a complex, scalable and resilient logistics infrastructure has been developed parallel to that of stationary trade to satisfy our needs as consumers.</p><p>The purchasing behaviour and the expectations we have of a retailer or supplier in this context have of course changed just as significantly. When we buy something, at least most of us want to have it at home as soon as possible. We often only want to wait in an exceptional case. You can imagine that these expectations pose one or two challenges for logistics. For many consumers, however, the processes that enable next-day or even same-day delivery are not really visible. And if they do, then only to a limited extent.</p><p>Online retailers offering same-day delivery must ensure that their logistics and warehousing operations have implemented goal-oriented processes and tools. Processes such as inventory, purchasing, order administration, picking, delivery and others should be automated whenever possible. Otherwise, fulfillment becomes critical.</p><h4>How logistics has evolved</h4><p>Logistics deals with the planning, control, optimization and implementation of, among other things, flows of goods and information. There’s also quite a lot of research that is being done in professional sciences. The aim is always to ensure that</p><ul><li>the right good,</li><li>in the right amount,</li><li>in the right condition,</li><li>in the right place,</li><li>at the right time,</li><li>for the right customer and</li><li>is available at the right cost.</li></ul><p>In order to get goods from A to B, we have seen various innovations in transport. From simple carts and non-motorized boats, steam trains to cargo ships, trucks, aircrafts and drones. Thanks to these transport vehicles, we are now able to deliver goods more or less cost-efficiently to practically every corner of the world. As far as the flow of information is concerned, we have obviously also made good progress. I mean you’re reading this blog article thanks to one of the most fascinating innovations in this context — the Internet. With the Internet, the cost of distributing information has moved close to zero, and the Internet’s interfaces with the complex logistics infrastructures around the world are causing e-commerce to boom. Compared to the Internet, however, logistics has much more friction and unnecessary inefficiencies, which means that logistics companies have room for optimization. A modern, holistic process design plays a central role in this journey for optimization.</p><p>We need to find innovative ways to reduce costs and increase the throughput of warehouses. Traditional warehouses are often well organized. However, they often reach their limits and are therefore no longer really suited to cope with today’s volumes and requirements. What we need are partially or fully autonomous warehouses that can be productive 24/7 and without classic shift operation. This gives us a completely different quality of flexibility and opportunities to optimize output.</p><h4>Types of logistics</h4><p>Sub-areas of logistics, which offer all potential for increasing efficiency, include in particular</p><ul><li>Procurement logistics</li><li>Warehouse logistics</li><li>Production logistics</li><li>Transport logistics</li><li>Spare parts logistics</li><li>Maintenance logistics</li><li>Distribution logistics</li><li>Waste disposal logistics</li><li>Information logistics</li></ul><h4>Inventory management; a hard nut to crack</h4><p>To keep customers satisfied and improve customer retention, retailers must have enough products in stock or at least be able to stock up as quickly as possible. That’s also why in recent years, global inventories have increased significantly, despite the additional pressure on costs, picking and packing to meet deadlines, and on shippers to deliver packages in ever shorter time frames. However, the upside that can be realized in this context, at least for the big players, compensates in many cases for the increased expenses. This is actually not really surprising when we consider that the acquisition of new customers is a large chunk of any budget.</p><h4>A new era: E-Commerce</h4><p>As mentioned at the beginning, the current extraordinary situation around the globe has led to major structural changes. Retail markets are in decline in many places, while e-commerce is flourishing and sales at individual retailers are more than twice as high as usual. This is most likely due to the marked change in behaviour during the period of social distancing. It remains to be seen how this will behave after a «normalization».</p><p>But even beyond these special times, e-commerce has gained considerable traction and the increased push towards e-commerce in recent months has only made it more obvious than ever that the continuous rise in demand is leading to difficulties in meeting it, proving that companies need to implement technologies to automate their logistics processes. This insight becomes even clearer when we consider that e-commerce sales currently account for only <a href="https://www.emarketer.com/content/us-ecommerce-2019">about 12% of global retail sales</a>.</p><h4>Automation in logistics</h4><p>Logistics automation refers to the use of control systems, machines or software to improve operational efficiency. The term is open per se, broadly comprehensive and refers to all functions that are performed at the technological level.</p><p>Processes along the supply chain, including procurement, distribution, customer service and reverse logistics, are among the many processes that can or must be automated in a warehouse or distribution center. The problems described above should make it clear that e-commerce and automation must be optimally coordinated. After all, one cannot really develop, cannot really mature without the other. Automation is therefore absolutely essential in order to successfully scale e-commerce further.</p><p>For CEOs, automation in logistics has become a high priority due to the increasing shortage of labour, the growing demand from online retailers and technological progress. This is because warehouse and office workers can work more closely and profitably with different technologies, reducing their workload and creating added value for the company. Machine learning-based systems are the central and increasingly important sparring partners in this context. Their ability to cope extremely well with complexity and to make it more accessible to us is one of the various unique selling points that make them so meaningful. In many areas of logistics, interesting and significant process innovations can be expected from the increased use of such technologies.</p><h4>Intelligent Document Capture to promote automation</h4><p>Today, machine learning technologies are used very pragmatically, for example in document extraction. Because in logistics there are many documents to pass, check and approve. These include invoices, orders, bills of lading, certificates and customs documents. Typically, long tables with information about all shipped goods such as item description, quantity, unit price, tax rate, etc. have to be entered into a leading system. Doing all this manually would take hours or even days. By automating this document data entry work with machine learning-based OCR software, it is instead possible to extract the same amount of data in just a few seconds. This helps to significantly improve document handling efficiency, make downstream process automation more fluid and robust, and simplify customs compliance.</p><p>Through this automation-driven optimization of logistics, goods movements can be tracked more transparently at a higher level and inventory management can be fully integrated by the companies’ ERP systems. This results in greater organizational control over all company-relevant resources. In addition, companies have near real-time access to data streams, which enables targeted and timely interventions or is also a significant advantage purely for reporting, as the management can make better decisions based on relevant data more quickly. The ability to extract more errors from processes, easily scale implementations and build up more speed in workflows are additional constructive effects that can bring decisive advantages to a logistics company.</p><p>So, Parashift provides you with a key element for the transformation to a top modern and efficient logistics company with the versatile Parashift Platform, which enables you and your team to generate a large impact with relatively little effort, leading to more efficiency and agility in your logistics processes.</p><p>That’s why you should now make use of this chance and <a href="https://blog.parashift.io/cs/c/?cta_guid=4cb23862-a7ff-4b42-b4f4-27b5375a5279&amp;placement_guid=f449b583-44bf-4bd4-9438-6881487e5f4b&amp;portal_id=5804034&amp;canon=https%3A%2F%2Fblog.parashift.io%2Fblog%2Fintelligent-document-capture-in-logistics&amp;redirect_url=APefjpFF1hKJY8R8IP0OHAFfWHvnTylmwSiJPTkIQSg4QNeuDv1sVFRVLbl5_8qUcYhcFh-_SH7fSI0dBhAy7snJO9YHP2fVUuXJK-QJOgxKJrh05zxhJlygRXlc3MwghIXvbdqp2DsKkCQH2w1hvZraBEbdYIWdNoELyueZOMYaocB1oHKhQKI&amp;click=1071f309-7418-481f-8b29-324630ed16b2&amp;hsutk=187f2d589bd9cb8846e58325e2b02131&amp;signature=AAH58kEdGK9Epm7OClJU4Af0LkjRagbl_w&amp;utm_referrer=https%3A%2F%2Fblog.parashift.io%2Fde%2Fblog%2Fintelligent-document-capture-in-logistics&amp;pageId=32492249587">register now for a free 14-day test account</a> and <a href="https://meetings.parashift.io/meetings/thilo-rossa">contact one of our specialists </a>for a first assessment of possible solutions. We look forward to getting to know you!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=77135062e616" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/more-automation-in-logistics-with-intelligent-document-capture-77135062e616">More Automation in Logistics With Intelligent Document Capture</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Evolution of Document Capture]]></title>
            <link>https://medium.com/trea-ventures/the-evolution-of-document-capture-3acb86c9a140?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/3acb86c9a140</guid>
            <category><![CDATA[document-capture]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[document-extraction]]></category>
            <category><![CDATA[ocr]]></category>
            <category><![CDATA[document-management]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Tue, 19 May 2020 09:48:16 GMT</pubDate>
            <atom:updated>2020-05-19T09:48:16.923Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This article was originally posted on </em><a href="https://parashift.io/"><em>https://parashift.io/</em></a><em>.</em></p><p>Countless documents such as orders, delivery notes and invoices are issued on a daily basis. These documents must then be captured in corporate IT systems in some way. Nowadays, technological developments have led to efficient methods of passing these documents on to downstream processes in a machine-readable and structured form. At <a href="http://parashift.io/">Parashift</a>, we mainly use machine learning for this purpose. Because this type of technology allows us to reduce the necessary human intervention to a minimum. This offers numerous advantages. Not only does it save time and money, but it can also avoid rather monotonous and tedious tasks. As you can certainly imagine, not so long ago, document capture looked quite different. From the first inventions to the latest developments, the following article is intended to give you an impression of how modern document processing has developed over the years.</p><h3>Beginnings and first attempts</h3><p>Scientists have always dreamed of using machines to replicate human capabilities. This includes reading and categorizing what is read. The origins of optical character recognition (OCR) date back to 1870, when <a href="https://en.wikipedia.org/wiki/Timeline_of_optical_character_recognition#cite_note-:0-1">Charles R. Carey,</a> an American inventor, invented the retina scanner — an image transfer system that uses a mosaic of photocells. Two decades later, the development of a successive image scanner by the German inventor Paul Nipkow gave the decisive impetus for today’s televisions or reading machines and thus for character recognition.</p><p>However, the first OCR devices were primarily intended for blind and visually impaired people, such as the <a href="https://en.wikipedia.org/wiki/Optophone">optophone developed in 1912 by the Irish inventor Dr. Edmund Fournier d’Albe</a>. The optophone was a scanner that, when moved across a printed page, produced sounds that correspond to certain letters or characters so that they can be interpreted by a blind person.</p><p>It was not until the 1950s and the invention of the Optacon that the potential applications that OCR can have in the business world were recognized. From that time on, the focus of development was also on the possible applications of OCR for businesses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/599/0*lVr9MrrYB90Spt0o.png" /><figcaption>Optophone by Dr Edmund Fournier d’Albe</figcaption></figure><h3>The first OCR devices</h3><p>In 1950, the technological revolution was advancing at a rapid pace, resulting in increasing data volumes and thus making electronic data processing increasingly important. At that time, data was still entered by punch cards and initially processed by hand. However, the processing of increasing amounts of data required a cost-effective alternative to manual processing. At the same time, machine reading technology became sufficiently mature for commercial applications. As a result, OCR machines went into mass production in the mid-1950s.</p><p>The <a href="https://en.wikipedia.org/wiki/Timeline_of_optical_character_recognition">first real OCR machine was installed in 1954</a> at the American magazine Reader’s Digest. This device was used to convert typewritten sales reports into punch cards that could be read and searched by computers. It was not yet possible to extract only the most relevant data, so the entire documents were read in. Moreover, the devices were still quite bulky and expensive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/399/0*YCl5dktzXcWcN4Lw.png" /><figcaption>An example of a punch card</figcaption></figure><h3>Commercialization and first automation</h3><p>The commercial OCR systems that appeared between 1960 and 1965 can be described as the first generation of OCR. This generation made it possible to automate the first few steps in the process of document processing. However, the first devices still had a very limited number of symbols and letters that could be read. And since the symbols were specially designed for machine reading, the first symbols did not look very natural either.</p><p>With the further development of the devices, the first multifont devices appeared, which could read up to ten different fonts. The IBM 1418 Optical Character Reader is a representative model for this generation. These devices worked with so-called template matching, which made the above-mentioned automation possible. The OCR devices could now be trained to read only predefined sections of documents. This meant that a large amount of work that previously had to be done by humans cloud now be taken over by the OCR devices.</p><p>In the mid-1960s and early 1970s, the second generation of OCR machines appeared. These systems were capable of recognizing normal, machine-printed characters and also had the ability to recognize handwritten characters. However, the character set for handwritten characters was limited to numbers and a few letters and symbols.</p><p>The famous and also first system of this kind was the IBM 1287, which was presented at the 1965 World Fair in New York. Further technological advances were observed at Toshiba, who at the same time launched the first automatic letter sorting machine for postal codes. Also in the same period, Hitachi succeeded for the first time in producing a high performance OCR machine at low cost.</p><p>Another milestone was the progress made in standardization. For example, the first optically machine-readable font <a href="https://en.wikipedia.org/wiki/OCR-A">“OCR-A”</a> was developed in 1968 by Adrian Frutiger from Switzerland. This font was highly stylized and designed to facilitate optical recognition, but still remained legible for humans. A few years later, Frutiger designed a new typeface, <a href="https://en.wikipedia.org/wiki/OCR-B">“OCR-B”</a>, which replaced its predecessor. The reason for the new typeface was its low popularity in Europe and the technological advances made in optical character recognition in the meantime. Compared to OCR-A, this typeface therefore resembles a normal, sans-serif print font much more. It was then declared a worldwide standard in 1973 by the <a href="https://en.wikipedia.org/wiki/International_Organization_for_Standardization">International Organization for Standardization</a>, or ISO in short. Some attempts were made to merge the two fonts into one standard, but instead devices were developed that could read both standards.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/599/0*8NcImR83NalzcAf5.png" /></figure><h3>Technological progress and the Kurzweil Reading Machine</h3><p>For the 3rd generation of OCR systems, which came onto the market in the mid-1970s, the greatest challenge was the recognition of documents with poor quality and large printed and handwritten character sets.</p><p>Significant advances in hardware technology have resulted in lower costs and improved performance of the devices. Although more and more sophisticated OCR machines came onto the market, simple OCR machines were still widely used. In the time before personal computers (PCs) and laser printers began to dominate the field of text production, typing was a special niche for OCR. The uniform printing distance and the small number of fonts made simple OCR devices very useful. Rough drafts could be created on ordinary typewriters and fed into the computer via an OCR device for final processing. In this way, word processing programs, which were a costly resource at the time, could be used by several people, thus avoiding high acquisition costs.</p><p>At that time, researcher <a href="https://en.wikipedia.org/wiki/Ray_Kurzweil">Raymond Kurzweil</a> succeeded in launching the first commercial reading machine capable of translating printed matter into spoken words: The <a href="https://en.wikipedia.org/wiki/Reading_machine">Kurzweil Reading Machine</a>. A milestone in history. For the first time there was a commercial solution that allowed blind people to read.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/399/0*6IUl0B-j8y-PpaCS.png" /><figcaption>Student of the Perkins School for the Blind uses the Kurzweil Reading Machine</figcaption></figure><h3>OCR software and machine learning-based OCR</h3><p>Although OCR devices became commercially available in the 1950s, only a few thousand systems had been sold worldwide by 1986. The main reason for this was the cost of the devices. However, as hardware became cheaper and OCR systems became available as software packages, sales figures also increased significantly.</p><p>Advances in computer technology have made it possible to integrate the recognition part of OCR into software packages that work on PCs. However, OCR software also had disadvantages, particularly in terms of speed and the type of character sets read. Today, several thousand OCR systems are sold every week. In addition, the cost of Omnifont OCR devices has been falling year after year, which has led to OCR devices finding their way into the mainstream.</p><p>A few years ago, the first attempts to combine OCR with machine learning were made. In 2013 the famous <a href="https://en.wikipedia.org/wiki/MNIST_database">MNIST (Modified National Institute of Standards and Technology database) </a>was established. This database consists of handwritten letters and numbers and is used for machine learning training. With this, document capture has made the step into the latest digital revolution and the focus has been on complete automation since then. And researchers are registering quite promising successes here, although it must also be clearly emphasized that we are only at the beginning and that there are still numerous technical challenges to be mastered.</p><h3>Template-based OCR vs. machine learning-based OCR</h3><p>Although template-based OCR is much older than machine learning-based OCR, the results of template recognition are usually quite good and consistent, which means that much of the manual work can be automated. However, automation also requires a lot of prior configuration for each field to be extracted. Defining in advance the exact location on the document where the information is to be found and how it is to be extracted is crucial. In contrast to the complete automation that is possible with machine learning-based OCR, this can therefore be a major and very costly disadvantage.</p><p>Another major disadvantage of template-based OCR, as already mentioned, is that document structures that deviate from the template do not work. In other words, as soon as the document changes, even minor layout changes, the machine must be reconfigured, which is extremely time-consuming. In addition, the setup of the templates is not quite simple either and usually requires external experts. Therefore, the successful management of templates can quickly become very cost-intensive and a major challenge.</p><p>Machine learning-based OCR systems offer a solution to these problems. The machine learning algorithms are trained with the help of training sets, i.e. data sets for training purposes. A validation set is then used to optimize the resulting model structures and, if necessary, adjust some design parameters. Finally, a test set is used to measure the performance of the algorithms and to check their suitability for generalization. In doing so, for example, the accuracy, sensitivity or speed of the generated system is checked. If the results are not satisfactory, the algorithms must be further optimized with new training sets.</p><p>Due to the way the algorithms of machine learning-based OCR solutions work, there are significant advantages over template-based OCR solutions. They differ in the way that they offer more flexibility and do not require templates. This means that your company no longer needs to carry out activities such as maintenance, consulting and coordination with regard to templates, which means significantly lower costs for you. By continuously training the algorithms on all customer documents, performance is also continuously improved, which in turn leads to less human intervention and lower costs over time. And you don’t have separate large-scale projects for extraction rate improvements, which quickly generate exponentially higher costs with increasing extraction quality.</p><p>Machine learning-based OCR solutions therefore offer numerous advantages over template-based OCR solutions. Nevertheless, it should be noted once again that the template approach can be very accurate. But as soon as different layouts and document changes come into play, machine learning-based OCR solutions are the clear winners.</p><h3>Test now!</h3><p>Convince yourself of the potential of machine learning-driven OCR and upload various test documents (ATTENTION: currently, we only process documents from purchasing: offer, order, order confirmation, delivery note, pro-forma invoice, invoice, receipt, credit note and dunning letter).</p><p>PS. By the way, on working days between 08:00 and 17:00 CET, we also validate the extraction results within 3 hours. So with Parashift you get outstanding data quality and do not need to tie up resources unnecessarily for tasks that are not directly part of your core business.</p><blockquote><a href="https://id.parashift.io/users/sign_up?product=dc&amp;p2p_enabled=true">Test here!</a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3acb86c9a140" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/the-evolution-of-document-capture-3acb86c9a140">The Evolution of Document Capture</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 100 Swiss Startup Award Public Voting: We Need Your Vote]]></title>
            <link>https://medium.com/trea-ventures/top-100-swiss-startup-award-public-voting-we-need-your-vote-50b6320d7086?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/50b6320d7086</guid>
            <category><![CDATA[rap]]></category>
            <category><![CDATA[document-extraction]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Mon, 18 May 2020 12:53:52 GMT</pubDate>
            <atom:updated>2020-05-18T12:53:52.183Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uRYNZI9waKpup3upfXkzHw.png" /></figure><p>Every year, the TOP 100 Swiss Startup Award presents the 100 most innovative and promising Swiss start-ups selected by a 100-member jury of leading investors and experts. The TOP 100 Swiss Startup Award Public Voting runs parallel to the ranking of the expert jury and highlights an outstanding Swiss startup, which will also be celebrated on the stage of the Award Night on September 9, 2020. This year, we want to be part of it. And to do so, we now need your vote.</p><h4>How to vote for <a href="http://parashift.io/">Parashift</a> in less than 1 minute</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/599/0*sOfgStr9alKY3Gkb.jpeg" /></figure><p>1. Visit <a href="https://www.startup.ch/index.cfm?CFID=366000509&amp;CFTOKEN=c6a3f06f99dd1973-205C1643-CD12-EB19-7798A3B48DC11B3E&amp;page=129367&amp;profil_id=16355&amp;backpage=129571">this link</a>.</p><p>2. Click on the golden button “vote now”.</p><p>3. Log in with your LinkedIn profile. Done! 🤗 We thank you very much for the support!!</p><p><em>All votes must be cast by midnight on 12 July 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50b6320d7086" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/top-100-swiss-startup-award-public-voting-we-need-your-vote-50b6320d7086">Top 100 Swiss Startup Award Public Voting: We Need Your Vote</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beetroot and Parashift Jointly Offer AI-Based Solutions for Digital Document Processes out of the…]]></title>
            <link>https://medium.com/trea-ventures/beetroot-and-parashift-jointly-offer-ai-based-solutions-for-digital-document-processes-out-of-the-bc8ae6c171d4?source=rss-37f341b69068------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc8ae6c171d4</guid>
            <category><![CDATA[document-extraction]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[rap]]></category>
            <category><![CDATA[document-management]]></category>
            <dc:creator><![CDATA[Mattia Rüfenacht]]></dc:creator>
            <pubDate>Mon, 18 May 2020 06:03:15 GMT</pubDate>
            <atom:updated>2020-05-18T06:03:15.764Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beetroot and Parashift Jointly Offer AI-Based Solutions for Digital Document Processes out of the Cloud</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*787-cVyJXCjmZPilj0_M3A.jpeg" /></figure><p><em>This article was originally published on </em><a href="https://parashift.io/"><em>https://parashift.io/</em></a><em>.</em></p><p>The way in which companies work together and communicate is changing fundamentally. Beetroot accompanies companies on their way to digitalization: with individual solutions for internal communication and the digital working environment.</p><p>Beetroot focuses on strategy development and consulting in the area of IT renewal. The specialists develop digitization strategies, create IT assessments and evaluate new applications for their customers. Beetroot supports the implementation of these strategies through coaching and IT project management and assists companies in change management initiatives.</p><p>The second area is organization and cooperation in companies. Here, the “Digital Workplace” is at the centre. Everything revolves around the question of how employees can better collaborate with applications such as intra- and extranets and other collaboration platforms. Beetroot is making an important contribution, especially in the current situation where many companies are forced to move to the “Digital Workplace” more quickly.</p><p>It goes without saying that this type of digitization always involves documents and their data extraction:</p><p>“Parashift develops technologies that make new digital processes possible in the first place, because the costs for extracting different types of documents were simply too high until now. This has great financial and qualitative advantages, especially when processing digital archives,” says Daniel Schnyder, Senior Consultant and Managing Partner at Beetroot.</p><p>“Over the past few years, Beetroot has built up impressive expertise in terms of experience and skills in the field of digital transformation and is one of the leading consulting firms in Switzerland. We look forward to working with them more intensively in the future,” said Alain Veuve, founder and CEO of Parashift.</p><p>Beetroot and Parashift are jointly focusing on medium-sized and larger companies in Switzerland. Both believe that the digital transformation will accelerate significantly in the future and that the joint offering will provide a competitive advantage for their customers.</p><p><strong>About Beetroot:</strong></p><p>Beetroot is a specialist in digital work organization and strategic IT consulting and is involved with modern forms of work and communication technologies. With extensive experience and a wealth of expertise, Beetroot helps companies prepare for the digital future — with a clear focus on the practical benefits for the company and its teams.</p><p><a href="https://beetroot.ag/">https://beetroot.ag/</a></p><p><strong>About Parashift:</strong></p><p>Parashift AG, founded in 2016, is a deep-tech company focusing on the autonomization of document extraction through machine learning. Parashift is based near Basel (Switzerland) and employs around 20 people in the field of machine learning engineering and development. Forbes named Parashift 2019 as one of 30 promising AI startups in Europe.</p><p><a href="https://parashift.io/">https://parashift.io/</a></p><p>If you have follow-up questions, do not hesitate to contact us:</p><p>Daniel Schnyder | <a href="/service/mailto:daniel.schnyder@beetroot.ag">daniel.schnyder@beetroot.ag</a> | +41 41 700 30 41</p><p>Mattia Rüfenacht | <a href="/service/mailto:mattia.ruefenacht@parashift.io">mattia.ruefenacht@parashift.io</a> | +41 61 508 77 77</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc8ae6c171d4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trea-ventures/beetroot-and-parashift-jointly-offer-ai-based-solutions-for-digital-document-processes-out-of-the-bc8ae6c171d4">Beetroot and Parashift Jointly Offer AI-Based Solutions for Digital Document Processes out of the…</a> was originally published in <a href="https://medium.com/trea-ventures">TREA Ventures</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>