Image Segmentation by Position-Edge-Aware Network

Existing segmentation methods still suffer from problems with the loss of location information and edge details of segmented objects, resulting in the inability to efficiently model long-range data dependencies, further affecting the segmentation performance of the model. To address this issue, an edge- and position-aware adaptive attention block (EAPA) is designed. It can aggregate features along two spatial directions and then calculate offsets from the input features to obtain a boundary-enhanced feature map. The attention block captures cross-channel information while incorporating boundary details and position sensitivity, enabling the extraction of fine-grained details. Furthermore, to further improve the model's segmentation performance and the feature extraction capability of edge detail information. This paper presents an element-by-element weighted (EW) feature fusion module. This module combines high-level and low-level features to generate a fused feature map that contains rich semantic information and edge details. Experimental results demonstrate that our proposed network architecture with these two modules performs better in the Cityscapes dataset, achieving superior segmentation performance.