U-Net §

Architecture §

U-Net, introduced in 2015, boasts a symmetrical structure comprising two paths:

• Contracting Path: Captures contextual information.
• Expanding Path: Enables precise localization, especially at object or feature boundaries.

Training and Performance §

Trained on a limited image dataset, U-Net surpassed sliding window CNNs, excelling in tasks like medical image segmentation.

Diverse Applications §

Beyond typical CNN usage, U-Net shines in generative and diffusion models (e.g., SD Imagine, DALL-E), initially applied to medical image segmentation. Later, it adopted tasks such as:

• Image segmentation
• Generating segment masks
• Upscaling resolution
• Diffusion models(gaussian noise to newly generated images)
• Cascading diffusion for higher resolution( 3 unit in a row for higer resolution )

See all of above cases we are taking image as input to give image as output

Main Logic §

The connecting path copies features from the opposite encoder part, adding them to the decoder. This consolidates both semantic and spatial information for precise pixel-level segmentation.

• as decoder has semantic infomation (like this total area has bike )
• as endoer has more spacial features (like these are the pixels the bike is)
• by taking both into consideration we get pixel perfect segmentation

Paper Link §

U-Net: Convolutional Networks for Biomedical Image Segmentation