Improving Weakly Supervised Semantic Segmentation with Contrastive Learning and Graph-Based Contextual Reasoning

Weakly Supervised Semantic Segmentation (WSSS) aims to achieve pixel-level segmentation using limited annotation information but often encounters foreground-background confusion, particularly in complex scenes. To address this issue, we propose two key innovations: (1) a method for generating foreground instance masks and calculating contrastive loss to maximize the distinction between foreground and background features, and (2) the integration of a coordinate space graph reasoning module in the convolutional neural network to enhance global contextual representation. These methods significantly improve the quality of pseudo-labels and overall segmentation performance.