Hierarchical Activation Dual-Backbone Network for Weakly Supervised Semantic Segmentation

Weakly supervised semantic segmentation (WSSS) aims to achieve segmentation using weak labels, thereby reducing annotation costs. Current mainstream WSSS methods utilize class activation map (CAM) to generate pseudo-labels. However, CAM primarily focuses on discriminative regions beneficial for classification, which adversely affects the final segmentation performance. In this article, inspired by the learning mechanism in cognitive science theories, we propose a novel hierarchical activation dual-backbone network (HADBN) to obtain high-quality CAM. HADBN employs serial dual backbones to decompose the CAM activation process into a free-training main-body learning part and an iterative detail supply part. Through this decomposition, HADBN highlights detailed regions that conventional CAM often overlooks, and achieves significant performance improvements through the precise application of algorithms. Furthermore, HADBN only takes supervised training with lightweight module on the main-body part of CAM to avoid introducing excessive computational cost. In extension experiments on the VOC and COCO datasets, HADBN significantly improves the quality of CAM and achieves final segmentation results with mean intersection over union (mIoU) scores of 73.6 and 45.3, respectively, which are competitive with recent state-of-the-art methods. Additionally, the framework of HADBN demonstrates a certain level of versatility, making it well-suited for adaptation to other methods and improving their performance.