Robust Learning Against Label Noise Based on Activation Trend Tracking

Defect detection of industrial products plays an important role in industrial production. Deep learning methods have been introduced to replace manual labor and save costs. Training deep neural networks (DNNs) requires large datasets with precisely annotated labels. However, it is challenging to obtain such clean datasets, especially in industrial scenes. For instance, in tire industry, classification of tire defects is ambiguous and thus labeling a big and clean dataset for defect detection is almost impossible. Therefore, training robust models with datasets containing noisy labels has received increasing attention. Different from existing methods, we conduct research on the learning process of such datasets from a new perspective, by analyzing the channel-wise activation intensity of training samples. Deep models exhibit a two-phase learning process when training with noisy datasets, which is quite different from training with clean labels. This motivates us to propose a novel method called activation trend tracking-based robust learning (ATTRL). During the training process, activation intensity is monitored and used to adjust the loss function to avoid overfitting on noisy labels. To illustrate the effectiveness of the proposed method, empirical analysis and robustness verification are presented through extensive experiments. Experimental results on synthetic and tire datasets demonstrate that our approach is tolerant to high proportions of noisy labels and is superior to other existing methods, exhibiting a good prospect for industrial application.