AtICNet: semantic segmentation with atrous spatial pyramid pooling in image cascade network

This paper describes a new type of image segmentation method based on deep convolutional neural networks (DCNN) in the actual autonomous driving scene. The spatial pyramid pooling model is used to identify and segment the actual scene to complete the machine-aware task. In order to improve the information aggregation of the whole image, we use atrous convolution for multi-scale feature extraction based on the pyramid structure of image cascade network (ICNet), removing a residual module in the fifth stage of the network, in order to reduce the scale of the convolutional layer. The feature map is preprocessed by padding and atrous convolution before the four-level spatial pyramid model. Then, conventional convolutions are introduced to compose the atrous spatial pyramid pooling (ASPP) structure. Finally, the four feature maps in the pyramid are merged with the feature maps before input into the pyramid. This paper analyzes the spatial pyramid model, receptive field, and dilation convolution in detail and propose atrous image cascade network (AtICNet). Experiment results in the cityscape dataset have shown that AtICNet has some improvements over ICNet, by improving the accuracy of the segmentation.