Minimum Category Confusion for Semi-Supervised Industrial Quality Inspection*

Quality inspection is an important task in industrial scenarios and has attracted attention in recent years. However, the cost of industrial application deployment is high. On the one hand, labeling a large number of defective samples for each industrial application is costly. On the other hand, unlabeled quality inspection data are relatively easy to collect, but the lack of labeling information makes it a high cabinet. Hence, this article studies defective quality inspection based on the semi-supervised learning setting, and only a single labeled defect sample of each category and some unlabeled samples are provided. While semi-supervised learning methods have achieved excellent performance on the benchmark datasets, these methods in industrial scenarios still face challenges due to the large variance in defect appearance. Additionally, the current method either involves a complex process of noise labels or the design of a threshold scheme of the pseudo-label selection, resulting in a high complexity of model optimization. Specifically, we introduce the category confusion loss of unlabeled data samples to reduce the pairwise class confusion of unlabeled instances, thereby improving the discrimination of the classifier. Experiments on the industrial quality inspection dataset has shown the effectiveness of the method. Numerical experiment results indicate that the proposed method outperforms other semi-supervised methods and yields 16% relative improvement on the benchmark dataset NEU-CLS.