Enhanced global attention upsample decoder based on enhanced spatial attention and feature aggregation module for semantic segmentation

This work designs a simple but effective decoder structure called Enhanced Global Attention Upsample Decoder (EGAUD) based on enhanced spatial attention (SA) and feature aggregation module for semantic segmentation, which is used to recover the prediction rich in detail and semantic information. First, Enhanced SA (ESA) and Feature Aggregation Module (FAM) are used to enhance the details and semantic information of shallow and deep features, respectively. ESA integrates the SA module and residual enhancement module to enrich spatial information while ensuring the effectiveness of backpropagation. FAM uses three parallel pooling operations to expand the receptive field of features, to optimize the semantic information. Then, the authors use Global Attention Upsample (GAU) to recover the feature resolution. From bottom to top, GAU utilises two pooling operations to fully integrate the features of adjacent layers. The results on two datasets, PASCAL VOC 2012 and PASCAL‐Person‐Part, show the effectiveness of the authors’ model.