General Segmentation Context: SegNet And PSPNet¶

This appendix gives compact background on two general semantic-segmentation architectures that are useful for reading medical segmentation papers, but are not core medical architectures in this book. They are not implemented locally, not listed in data/architectures.yml, and not part of the main architecture catalog.

Use this page as context after reading FCN, and before comparing encoder-decoder and context-module ideas in U-Net and DeepLabv3+.

Why These Stay In The Appendix¶

SegNet and PSPNet are important general semantic-segmentation references. They help explain design ideas that appear in later segmentation work, but this repository keeps the central architecture list focused on medical image segmentation and a small number of general roots.

Model	Context role	Local status
SegNet	General encoder-decoder segmentation with pooling-index-based upsampling.	Appendix only, not implemented locally
PSPNet	General scene parsing model that uses pyramid pooling for global context.	Appendix only, not implemented locally

SegNet¶

SegNet is an encoder-decoder segmentation architecture. The encoder applies convolution and pooling to build compact feature maps. The decoder upsamples those feature maps back toward image resolution and produces dense segmentation predictions.

The distinctive SegNet idea is how decoder upsampling uses the max-pooling indices saved by the corresponding encoder stage. During max pooling, the encoder records which spatial location won inside each pooling window. During upsampling, the decoder uses those recorded indices to place activations back into sparse higher-resolution feature maps before later convolution refines them.

graph LR Input["Input image"] Encoder["Encoder<br/>convolution + max pooling"] Indices["Saved pooling indices"] Decoder["Decoder<br/>index-guided upsampling"] Head["Pixel classifier"] Output["Segmentation map"] Input --> Encoder --> Decoder --> Head --> Output Encoder -. pooling switches .-> Indices -. guide upsampling .-> Decoder

Compared with FCN, SegNet emphasizes a more explicit decoder path instead of only upsampling coarse logits. Compared with U-Net, SegNet reuses pooling indices rather than concatenating full encoder feature maps at matching resolutions. That makes it a useful contrast for understanding how different decoder designs restore spatial detail.

PSPNet¶

PSPNet, or Pyramid Scene Parsing Network, focuses on global context. A backbone extracts a feature map, then a pyramid pooling module summarizes that feature map at several spatial bin sizes. Those pooled summaries are resized back, combined with the original features, and used to predict dense segmentation labels.

The key idea is that local features can be ambiguous without broader scene context. Pyramid pooling gives the model several context views, from coarse global summaries to finer regional summaries, before the final prediction head.

Compared with FCN, PSPNet adds a structured context aggregation module on top of dense prediction. Compared with DeepLabv3+, PSPNet is another way to build multi-scale context: PSPNet pools features at multiple grid scales, while DeepLab-style modules commonly use atrous convolutions and ASPP branches.

Relationship To The Main Reading Path¶

Idea	Main catalog anchor	How SegNet or PSPNet helps
Dense pixel prediction	FCN	Both models still produce per-pixel segmentation outputs.
Decoder-based localization	U-Net	SegNet shows a decoder that restores resolution using pooling indices instead of full skip-feature concatenation.
Multi-scale context	DeepLabv3+	PSPNet shows pyramid pooling as a different context module from ASPP.

References¶

SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation, arXiv: 1511.00561.
PSPNet: Pyramid Scene Parsing Network, arXiv: 1612.01105.