Fig. 5From: Ancient mural segmentation based on a deep separable convolution networkPSP-M workflow. The input image is subjected to deep separable convolution to form a feature map, which further undergoes maximum pooling and processing with the pyramid pooling module to obtain the semantic information of feature maps. Then, upsampling is performed for the image containing the semantic information using bilinear interpolation. Finally, feature maps of different hierarchies are spliced up to obtain the predicted imageBack to article page