Cross-Attentive Multimodal Enhancement for Weakly Supervised Referring Image Semantic Segmentation

Weakly Supervised Semantic Segmentation (WSSS) reduces annotation costs but often suffers from incomplete target activation. To address this, we propose Cross-Attentive Multimodal Enhancement for Weakly-Supervised Referring Image Semantic Segmentation (CAME). CAME introduces CrossModalityAttnFusion (CMAF) to guide language features using visual cues, thereby improving semantic alignment. A Positional Encoding 2D (PED) module enhances structural perception through coordinate encoding. Structure-Aware Consistency Loss (SACL) enforces boundary alignment via edge constraints. Additionally, Entropy Constraint Loss and Focal Loss V2 suppress low-confidence noise and emphasize difficult regions. This joint optimization improves pseudo mask quality, resulting in sharper boundaries and enhanced regional consistency. Experiments show that CAME achieves mIoU scores of 54.96%, 57.08%, and 53.01% on the validation and test sets of RefCOCO, 48.25%, 46.37%, and 45.09% on RefCOCO[Formula: see text], and 54.89% on the validation set of RefCOCOg, significantly outperforming existing weakly supervised methods. These results confirm the effectiveness and strong generalization ability of the proposed approach.