Print Defect Classification And Detection Technology Based On Deep Learning

Print is the main carrier for the public to get information about the items,and we always hope to get clearer and more error-free printed materials of all kinds,but in the actual manufacturing process,it is easy to produce defective products,and the defects caused by ink leakage,ink jet,ink color,etc.are often small and difficult to detect,so it becomes a challenge to improve the accuracy of quality detection in the production environment while ensuring real-time.Currently,industrial defect target detection methods still have bottlenecks such as complex models or algorithms,low real-time performance,and poor generalization performance,and recent research hotspots are mainly focused on model accuracy and lightweighting.For the problems of low accuracy of small target defect detection,large differences in defect categories and high cost of labeling,and difficulty in obtaining abnormal data for industrial printed products,the specific research of this paper is as follows.In this paper,a single-stage defect detection algorithm based on multilayer feature fusion of residual blocks is designed.The method based on the idea of migration learning and is improved based on the YOLOv5.The structure of the backbone network is adjusted considerably to improve the characterization ability of the deep-level large model for local and global features.Both the edge rejection method and the calculation of edge loss are adjusted to improve the effect of edge rejection and edge accuracy.To address the problem of small defective data sets and insufficient data volume in industrial scenarios leading to the model’s failure to converge,a semi-supervised learning approach based on generating pseudo-labels with strong data enhancement strategy is proposed to establish a semi-supervised YOLO framework,so that the training of the model no longer requires a large amount of data.In order to meet the demand of real-time industrial production,this paper designs a self-learning method of pseudo-labeling and logits knowledge distillation fusion.Under the teacher-student network structure,feature migration from the teacher network to the student network is achieved by using the predicted values of the teacher model as soft labels to achieve the purpose of lightweight models.The printing defect detection method proposed in this paper can solve the problem of printing defect target detection accuracy to some extent,and improve both for the data and the model itself to achieve an m AP of 0.82,which improves 26.1% in comparison with the same structural model,and reduce the number of parameters to only 2% of the large model through knowledge distillation,and finally reach a realtime detection speed of 50 fps.