AM-PSPNet: Pyramid Scene Parsing Network Based on Attentional Mechanism for Image Semantic Segmentation

In this paper, AM-PSPNet is proposed for image semantic segmentation. AM-PSPNet embeds the efficient channel attention (ECA) module in the feature extraction stage of the convolutional network and makes the network pay more attention to the channels with obvious classification characteristics through end-to-end learning. To recognize the edges of objects and small objects more effectively, AM-PSPNet proposes a deep guidance fusion (DGF) module to generate global contextual attention maps to guide the expression of shallow information. The average crossover ratio of the proposed algorithm on the Pascal VOC 2012 dataset and Cityscapes dataset reaches 78.8% and 69.1%, respectively. Compared with the other four network models, the accuracy and average crossover ratio of AM-PSPNet are improved.