Boundary Constraint and Dynamic Graph Feature Decoupling for Weakly Supervised Semantic Segmentation

The weakly supervised semantic segmentation tasks have progressed through pseudo-labels derived from class activation maps. Most methods perform poorly in complex scenes, capturing local foreground features and relying on simple context. To address these challenges, we propose the boundary constraint and dynamic graph feature decoupling(BC-GFD) approach, which is designed to enhance foreground features representations through three key components: the Moment-Constrained Fusion strategy (MCFS), the Dynamic Graph Relational Mapping(DGRM) strategy, and the semantic difference measurement(SDM) strategy. Specifically, MCFS is designed to emphasize and constrain non-discriminative regions, while DGRM manages long-range dependencies between global contextual semantics to extract global fine-grained features. Meanwhile, SDM focuses on quantifying foreground-background disparity, enhancing the ability to decouple local fine-grained features. BC-GFD effectively focuses on constraining boundary features and integrating global-local complementary foreground features to enhance adaptability to complex contexts. Extensive experiments on the PASCAL VOC 2012 and MS COCO 2014 demonstrate the effectiveness of our method and achieve the state-of-the-art performance.