Robust Weakly Supervised Semantic Segmentation with Noisy Pseudo-Labels via CNN-Based Learning

Weakly supervised semantic segmentation aims to train segmentation models with image-level labels instead of expensive pixel-level annotations. However, generating pseudolabels from class activation maps (CAMs) often introduces noise, hindering the performance of the segmentation network. This paper presents a novel CNN-based framework that addresses the challenge of noise in pseudo-labels for robust weakly supervised semantic segmentation. Our approach introduces a noise-aware learning module that explicitly models the uncertainty in the pseudo-labels during training. We incorporate a confidence estimation branch that predicts the reliability of each pseudolabel, allowing the network to focus on more reliable regions during training. Extensive experiments on the PASCAL VOC 2012 dataset demonstrate the effectiveness of our approach in mitigating the impact of noisy pseudo-labels and achieving significant improvements in segmentation accuracy compared to existing weakly supervised methods. Our method achieves state-of-the-art performance, showcasing the importance of noise handling in weakly supervised semantic segmentation.