Weakly Supervised Semantic Segmentation Network Based on Feature Decoupling and Pixel Aggregation

Image-level annotated weakly supervised semantic segmentation has garnered considerable attention because of its efficient annotation process and remarkable scalability. Most studies typically generate pseudo labels by leveraging class activation maps produced by classification networks, which are then used to train segmentation models. However, conventional class activation maps tend to focus on only parts of the foreground objects and are prone to encroaching into background regions, thus limiting segmentation performance. To address these issues, this paper proposes a weakly supervised semantic segmentation Network based on feature Decoupling and pixel Aggregation (DANet). The network consists of a decoupling branch and an aggregation branch. The decoupling branch, by disentangling foreground-background features, compels the model to concentrate on stable target representations while ignoring background noise. This effectively eliminates false associations between foreground and background, thereby addressing the co-occurrence problem. Meanwhile, the aggregation branch delves into global contextual information through a pixel aggregation module, generating more comprehensive class activation maps, which in turn yield improved segmentation performance. Benefiting from the synergy of the two branches, the network demonstrates advanced performance on the PASCAL VOC 2012 and MS COCO 2014 datasets.