Image Segmentation

Semantic Segmentation

Pixel-level classification.

Architectures

FCN: fully convolutional, replace fully connected with 1x1 conv. U-Net: encoder-decoder with skip connections.

Mask R-CNN: extends Faster R-CNN with segmentation heads.

Loss Functions

Cross-entropy per pixel. Dice loss: F1 score-like. Combination often used.

Applications

Medical imaging. Autonomous driving. Satellite imagery.

Key Takeaways

1. FCN, U-Net are common architectures
2. Per-pixel classification
3. Skip connections preserve detail