Semantic boundary enhancement and position attention network with long-range dependency for semantic segmentation

Some factors hinder advancements in semantic segmentation techniques, including intraclass inconsistency and interclass distinction. In addition, existing methods with self-attention mechanisms disregard semantic boundaries in highest-level feature maps, which hinders further performance improvement. To improve semantic segmentation performance, we propose a semantic boundary enhancement and position network (SBEPNet) that enhances feature maps with semantic boundaries and captures useful contextual information along these boundaries. Specifically, edge enhancement is employed to extract semantic boundaries, which are used to enhance the high-level feature maps. The enhanced feature maps are input into the boundary enhancement attention module to guide the learning of the discriminative long-range dependencies along object boundaries. The resultant feature maps are further refined by fusing the feature maps from the position attention module and the original feature maps. The experimental results verify the effectiveness of SBEPNet, which demonstrates the high potential for improving the generalizability of semantic segmentation.