We propose a two-stage label noise learning framework. Classification by deep learning with datasets with noisy labels. Filtering noise samples based on the loss of the warm-up stage is a common method, but it is impossible to judge the optimal warm-up length and thus suffers from memory effects. This paper is inspired by the use of contrastive learning self-supervision as pre-training to replace the warm-up part, and pseudo-labels for subsequent loss calculations, which can avoid the influence of noisy labels on pre-training. The second part is a semi-supervised algorithm that applies data augmentation. The adopted boosting strategy is to use weak boosting for any loss analysis task and strong boosting for gradient descent. Learn a more accurate representation for each image. Through these two steps, the performance of the network can be improved. Experiments show that our framework is more robust to data classification problems with noisy labels, and also works well in datasets with high noise rates. At the same time, since it is a two-step framework, it is also easier to split and apply in combination with other noisy label learning algorithms.
|