Feature / FCN (Fully Convolutional PSPNet (Pyramid Scene
SegNet UNet
Model Network) Parsing)
ResNet (commonly ResNet-
Backbone VGG / other CNNs VGG16 Custom encoder-decoder
50 or -101)
Architecture Encoder-Decoder with pooling Symmetric Encoder-Decoder with Encoder + pyramid pooling
Encoder + upsampling layers
Type indices skip connections module
Transposed convolutions Upsampling with pooling Bilinear upsampling +
Upsampling Transposed convolutions
(deconvolution) indices fusion
Skip
No No Yes No
Connections
Excellent on small datasets, fine Captures global + local
Strengths Simple and end-to-end trainable Efficient decoder, less memory
details context
Mobile/low-memory Medical imaging, small dataset Scene parsing, large-scale
Best For Baseline segmentation
environments cases segmentation
SegNet
UNET
PSPNet