Bisenet features

BiSeNet characteristic ：

1. Dual branch network

A network emphasizes details , Its characteristic graph is at least as small as 1/8 The size of the , In other words, try to keep more details of the feature map ;

The second is lightweight , He only uses three convolution layers , And that's what happened

The third is the branch of high-level semantic features （context path）, Emphasize the deeper depth , The deepest can reach 1：32 Under sampling ,

The fourth is the lightweight model , Use inception To achieve , While maintaining a relatively deep depth . Still keep a relatively small amount of calculation , Through global average pooling , Realize the capture of context information , The main idea is simple and lightweight backbone network

The fifth is feature fusion , The characteristic graph space of the two branches is different , There will be great differences in characteristics , Fusion will lead to many noise points , So it uses batch normalization , Reduce differences between feature representations , Weighted fusion using attention mechanism , To emphasize high-level semantic features

2. Feature fusion module

First, stack on a channel , That is, in series , Then a common convolution is used for processing after the channel is stacked , Then a mechanism similar to attention is used , Finally, there is a short circuit

3. Loss function

Use the main loss function to monitor the whole network , Use the auxiliary loss function to monitor the deep network on the right , The main purpose is to speed up the training progress

Use BiSeNet Realize your own semantic segmentation