Weakly supervised semantic segmentation based on positive and negative hybrid learning

Most of the current weakly supervised semantic segmentation (WSSS) methods follow the pipeline of first generating pseudo-masks via class activation maps (CAMs) and then using the pseudo-masks to train a fully supervised semantic segmentation model. However, these methods essentially make use of weak labels (lacking pixel-level annotations) to generate pseudo-masks, and as a cost, even with high-quality pseudo-masks, noisy labels are still inevitably present, which will have an impact on the subsequent segmentation performance. Therefore, in this paper, we propose a method based on positive and negative hybrid learning to improve WSSS in terms of mitigating the effect of noise. Instead of using the generated pseudo-masks directly for the training of the segmentation network, our method first designs the K-L divergence criterion to distinguish the pseudo-masks into clean labels and noisy labels, and then applies positive and negative learning to the clean labels and noisy labels to train the semantic segmentation network so as to effectively improve the performance of the segmentation network. The experimental results on the general dataset PASCAL VOC 2012 val set and test set show that the mean Intersection over Union (mIoU) values of the proposed method reach 70.6% and 71.7%, respectively, which outperforms current weakly supervised semantic segmentation methods and proves the effectiveness of the proposed method.