Sitemap
Jo Wang

Jo Wang

Member-only story

U-net

3 min readOct 21, 2024

U-Net is a convolutional neural network architecture for segmentation (specialized for tasks that require pixel-wise predictions)

The U-Net architecture is notable for its U-shaped structure, which facilitates precise localization of features while maintaining a high level of contextual information.

Focus on Localization: Designed to capture spatial hierarchies and details necessary for accurate segmentation, making it effective for tasks where precise pixel-level output is needed.

Press enter or click to view image in full size
https://media.geeksforgeeks.org/wp-content/uploads/20220614121231/Group14.jpg

Characteristics of U-Net

  • Symmetric Architecture / Encoder-decoder Structure/ U-shape: The symmetric encoder-decoder structure in U-Net refers to its two main components: the contracting path (encoder) to capture context and the expansive path (decoder) to enable precise localization. This architecture is designed to effectively capture both the context and fine details necessary for image segmentation. The U-shape allows for a balanced combination of context and localization.
  • Skip Connections: Connects layers from the encoder to the decoder, allowing for the retention of spatial information that might be lost during downsampling.
  • Efficient Training: U-Net can be trained on relatively few images due to its architecture, which utilizes data augmentation techniques effectively.

--

--