1. Introduction

• The motiviation of this paper is to solve the problem of max pooling and sub-sampling reduce feature map resolution.
• The encoder of SegNet is indentical to CNN of VGG16. In decoder, the proposed model used max-pooling indices from encoder.

3. Architecture

• The encoder of the model has 13 convolutional layer and each encoder layer has a corresponding decoder layer. Therefore there are 13 decoder layer. The output of final layer will be get through multi-class soft-max classifier and it will calculate probability of each class for a independent pixel
  
• In encoder, it is easy to loss boundary information of image during max pooling. Therefore, this paper proposed the method to store max-pooling indices for each encoder feature map.
• Decoder upsamples its input feature map using max pooling indices.(Figure 3) The final upsamples feature representation is fed to soft-max classifier. As a result, the output of the soft-max classifier will be (number of classess) dimentional output which represent probability for each class.

Reference
https://arxiv.org/pdf/1511.00561.pdf