Architecture
SegNet is deep convolutional neural network which consists of corresponding encoder and decoder for pixel-wise semantic segmentation. The encoder encodes or downsamples the input images into low resolution feature maps and the decoder network upsamples these low resolution encoder feature maps into the original resolution feature maps. It is similar to other architectures such as UNet and DeconvNet, however it differs from them in some ways. Segnet was designed to be an efficient architecture for pixel-wise segmentation and was motivated for the use in road scene understanding.
Encoder - Decoder Network
The encoder architecture is identical to the convolutional layers in VGG16 architecture. It discards the fully dense connected layers and keeps the 13 convolutional layers. Each encoder block consists of convolutional layers which produces a set of features maps, a batch normalization layer and a ReLU activation layer. It is then followed by a max pooling layer. The decoder also has the same number of convolutional layers. Each decoder architecture is similar to that of the encoder and only varies in terms of the upsampling layer.
In SegNet, downsampling is achieved by using max pooling layers after each encoder block. It has a pooling window size of 2 with a stride of 2. The model explores the idea of storing the max-pooling indices in each pooling window for each feature map. The upsampling layer in each corresponding decoder block upsamples the features maps by using the max-pooling indices.Output
The output of the final decoder of SegNet is fed into a trainable softmax classifier to predict the probabilities for the K number of classes. The output is a K channel image where each channel represents a class.
How’s it different
As stated earlier, SegNet has similarity with DeconvNet and U-Net in terms architecture. As compared to SegNet, DeconvNet requires more computational resources and is harder to train and in terms of U-Net, it does not reuse pooling indices but instead transfers the entire feature map to the corresponding decoders and concatenates them to upsampled decoder feature maps.
Implementation
I have implemented the SegNet architecture in Tensorflow 2 and it can be found here
Run the notebook 
Output Images
| Input | Predicted |
|---|---|
 |  |