Reference

B station video

Differences from previous models

1. The application scenario of the U-net series is a small target, so local features are sufficient.But when it is used to segment large targets, it is not very friendly.
2. In order to increase the receptive field, the traditional segmentation algorithm will choose pooling, but this will lose some information.
3. Based on the above problems, DeepLab proposes a delated convolution to increase the receptive field.

Advantages of atrous convolution

1. Increase the receptive field
2. Complete hole convolution by setting the dilation rate parameter, without adding additional calculations
3. The receptive field can be expanded by any multiple of the parameters without introducing additional parameters
4. The application is simple, just set a parameter in the convolution layer.
(learnable atrous convolution?)

SPP (Image Pyramid: Obtaining multi-scale (multi-receptive field) features of an image)

ASPP uses atrous convolution based on SPP

DeepLabv3 Model