A Logit Adjusting Transformer for Class Imbalance in Surface Defect Recognition

Surface defect recognition plays a crucial role in ensuring product quality in manufacturing systems. Deep convolutional neural networks (CNNs) achieve excellent performance on surface defect recognition, which requires a large amount of data and roughly uniform distribution of class labels for training. However, in the real world, surface defect recognition of products typically exhibits defect irregularities and imbalanced class distribution, some of which have very few samples. This is a great challenge to the generalization of the model on such classes. Therefore, this paper proposes a logit adjusting transformer-based method (LAT) to address the class imbalance problem in surface defect recognition. LAT encourages a larger margin between logits of the minority class and the majority class to achieve accurate recognition of the minority class. The LAT achieves better results compared with competing methods on a dataset of printed circuit boards (PCB) surface defects with five classes collected from a real-life manufacturing factory. The accuracy reaches 93.06% on the dataset with only ten samples in the minority class, which is improved by 7.52% to the best competing method.