Self-label correction for image classification with noisy labels

Label noise is inevitable in image classification. Existing methods usually lack the reliability of selecting clean data samples and rely on an auxiliary model to correct clean samples, which quality has a great impact on the classification results. In this paper, we propose the Dual-model and Self-Label Correction (DSLC) method to select clean samples and correct labels without auxiliary models. First, we use a dual-model structure combining contrastive learning to select clean samples. Then, we design a novel label correction method to modify the noisy labels. Finally, we propose a joint loss to improve the generalization ability of our models. In the experiment, we demonstrate the effectiveness of DSLC on various datasets, which achieves comparable performance to state-of-the-art methods.