Feature decoupling and recombination for weakly-supervised semantic segmentation

The feature coupling caused by co-occurring classes is a major challenge, particularly impacting Weakly Supervised Semantic Segmentation. Existing decoupling methods can only increase the feature differences between co-occurring classes within already coupled features without actually separating the coupled features. To resolve this fundamental limitation, this work introduces an explicit Feature Decoupling Module (FDM) in this paper, which strives to separate the features of co-occurring classes based on semantic aggregation and contrastive learning. Building on the FDM, we further propose a Feature Recombination Module (FRM) that can recombine features of different classes to form new complete feature maps. These new feature maps disrupt the initial coupling relationships between co-occurring classes and effectively enhance inter-class differences during the class recombination process. Experimental validation across standard benchmarks demonstrates our method's effectiveness, achieving improved mIoU on PASCAL VOC 2012 dataset.