Foster noisy label learning by exploiting noise-induced distortion in foreground localization.

Abstract

Large-scale, well-annotated datasets are crucial for training deep neural networks. However, the prevalence of noisy-labeled samples can cause irreversible impairment to the generalization of models. Existing approaches have attempted to mitigate the impact of noisy labels by exploiting the different loss or confidence distributions between clean and noisy data to detect and correct noisy labels. This paper investigates the noise-induced distorting effect on foreground localization by tracking the model's spatial attention distribution on visual activation maps. We observe that in clean samples, highly responsive regions usually focus on label-relevant foreground regions, whereas in noisy data, the model accidentally attends to uninformative background regions or cluttered object edges due to interference from label noise. Inspired by the observations, we propose a novel two-stage foreground localization-augmented noisy label learning framework named FLSC to concurrently boost the accuracy of sample selection and label correction for robust training. Specifically, FLSC first quantifies noise-induced distortion in foreground localization to foster conventional loss-based selection criteria by calculating the information reduction when deriving the foreground images from the original images based on the attention distribution. Next, we propose a noise-adaptive adversarial erasing strategy, which suppresses background activation by imposing adaptive erasure regularization, to eliminate overfitting to noisy samples while enhancing the learning of robust representations. To the best of our knowledge, it is the first attempt to exploit localization quality evaluation based on feature activation to address the label noise problem. Extensive experiments on synthetic and real-world datasets validate the superior performance of FLSC compared to state-of-the-art methods.