IAC-ReCAM: Two-dimensional attention modulation and category label guidance for weakly supervised semantic segmentation.

Weakly supervised semantic segmentation (WSSS) approaches aim at pixel-level semantic category prediction using only image-level labels. The existing classifier-based method ReCAM has achieved good results, however, the classifiers tend to only focus on the most discriminative regions, resulting in an uneven distribution of fea-tures in the resulting class activation maps (CAMs). Besides, the classifiers are susceptible to image background interference and generate false activation mapping. To solve the above problems, we propose an improved method IAC-ReCAM, which introduces an activation network that integrated attention modulation and category label guidance based on the ReCAM method. We utilize the attention modulation module to reassign the feature distribution of the CAMs from the perspective of channels and spaces in turn. Meanwhile, we use the class label guidance module to suppress the generation of false activation mapping. Furthermore, we verified the effective-ness of the IAC-ReCAM method improvement work on both PASCAL VOC 2012 and MS COCO 2014 datasets, our method outperforms a large number of existing mainstream methods. Among them, compared with the ReCAM method, the mIoU of the pseudo-labels on the two datasets is improved by 2.9% and 1%, respectively.& COPY; 2023 Published by Elsevier B.V.