Advancing weakly supervised semantic segmentation using class activation map augmented with graph attention module

In the field of computer vision, Weakly Supervised Semantic Segmentation (WSSS) at the image level poses a significant research challenge. Existing WSSS methods primarily rely on class activation maps (CAM). Due to the disparity between fully-supervised and weakly-supervised approaches, CAM often yields imprecise and coarse semantic information in generating target masks. To address the issues of semantic coarseness and detail loss in image-level WSSS, this paper introduces a novel approach that utilizes pixel-level relationships to optimize and refine the features extracted by the network. Maintaining the foundation of image-level class label weak supervision for semantic segmentation, this study forgoes extensive modifications to the model structure in favor of incorporating a Graph Attention Module (GAM). This module, by constructing a weighted undirected graph with pixels as vertices and pixel affinity as edges, simulates the interactions among neighboring pixels in graph representation learning. Such an arrangement allows for the effective propagation and integration of relational information between pixels, thereby enhancing the overall representational quality of the image. By employing graph constraints to facilitate semantic dissemination, our method optimizes the network's localization map, resulting in more precise CAMs and more accurate pseudo-labels for subsequent semantic segmentation network training. To evaluate our approach, we conducted quantity and quality experiments on the PASCAL VOC 2012 and MS COCO 2014 datasets. Compared to existing advanced weakly-supervised semantic segmentation methods, our approach shows notable improvements in performance.