ASPP is a module that can be placed anywhere by us, such as the following picture:
Source: ASPP (Atrous spatial pyramid pooling) originally came from the DeeplLab v2 network, and was later improved and strengthened in the DeepLab v3 network. This article uses the ASPP structure in DeepLab v3 for analysis and source code implementation.
The purpose of the module is to increase the receptive field of the network without changing the shape or downsampling, and enhance the ability of the network to obtain multi-scale context.
Module structure:
- A total of four branches are connected in parallel, and the first branch is a 1×1 ordinary convolutional layer.
- The second and third branches use a 3×3 expansion convolution, and the expansion coefficient rate used by different branches