DTP-Match: A Semi-Supervised Model with Method Dynamic Threshold and Pair Loss for Canned Food Defect Detection

Foreign matters may be mixed into canned food in the production process. The traditional defect detection method of canned food mainly relies on manual visual inspection, which is generally low in accuracy and efficiency. The defect detection method based on deep learning can realize the automatic detection and identification of defects, and achieve high detection efficiency. High performance deep learning models usually require a large number of high-quality labeled data to train, but building such data sets requires a lot of time and energy. Semi-supervised learning can effectively use unlabeled data and improve the performance of the model with only a small amount of labeled data. This study improves the existing semi-supervised classification algorithm for canned food data, and proposes a detection method of foreign matter defects in canned food based on dynamic threshold and pair loss. In view of the unbalanced distribution of defect categories in canned food data and the differences in learning difficulty of different categories, the dynamic threshold strategy is introduced to adjust the thresholds of different categories to adapt to the learning difficulty of different categories and adjust the learning focus of the model in real time. In view of the problem that the unlabeled data in the training process of the traditional semi-supervised model is not fully utilized, resulting in the waste of information resources, a set of pair loss items is designed to make full use of the similarity information between unlabeled data. The experimental results show that the method proposed in this paper has achieved good results in the detection of defects in canned food, achieving an accuracy of 91.79% on 30% labeled data, which is superior to other semi-supervised methods such as FixMatch and FlexMatch. It provides a new approach for defect detection in canned food.