Robust Learning of Noisy Labels Based on Semi-Supervised Techniques
Xinru Zhou, Menghan Gong, Zheng Li, Yiming Liu, De Yu, Qian Zhang · 2025
Computer vision tasks often require extensive labeled datasets, which is expensive. The presence of noisy labels in these datasets can notably impair deep neural networks’ performance. In our earlier work, we introduced a noisy label learning method that combines a Mixup loss with a recalibration strategy, operating independently of any prior knowledge about the noise type. Since the nature of the noise—whether symmetric or asymmetric—can often be determined through manual analysis of a small subset of data in practice, this paper enhances the model robust by incorporating noise type information. We employ a pre-training loss function that combines cross-entropy loss with a penalty term to boost model performance. We evaluate our approach against previous methods on both synthetic and real-world noisy datasets, focusing on test accuracy.