D2CL: A Dual‐Dimensional Contrastive Learning Method for Enhancing the Performance of Weakly Supervised Semantic Segmentation

ABSTRACT Weakly supervised semantic segmentation plays a pivotal role in domains such as autonomous driving and medical image analysis. However, existing approaches often rely on limited semantic cues from single images or paired samples, leading to underutilized intraclass information and entangled interclass features—both of which significantly impair segmentation performance. To address these challenges, we propose a novel dual‐dimensional contrastive learning (D2CL) framework that explores fine‐grained feature attributes both across and within views, thereby promoting intraclass compactness and interclass discriminability in the feature space. Specifically, the interclass prototype contrastive learning module constructs a cross‐view dynamic prototype memory bank and imposes a contrastive loss to enhance category‐level distinctiveness. In parallel, the intraclass pixel contrastive learning module focuses on pixel‐wise variations within the same category from a single view, enabling the model to capture more refined semantic details and better handle intraclass diversity. Extensive experiments conducted on the PASCAL VOC 2012 and MS COCO 2014 datasets demonstrate that D2CL consistently boosts the performance of multiple baseline models. For instance, the mean intersection over union of the baseline model SEAM is improved from 64.5% to 67.7%, while another model AMN sees an increase from 69.6% to 71.8%, highlighting the general applicability and effectiveness of our method.