Unsupervised Cigarette Packaging Defect Detection Method Based on Multi-Scale Feature Matching

During the cigarette packaging printing process, defects can occur due to environmental changes, printing machine malfunctions, or human errors. Currently, visual inspection by humans is commonly used in production, but this method is inefficient and highly subjective. Most existing defect detection methods based on computer vision require extensive manual annotation. However, due to the uncertainty in defect types, forms, and sizes, as well as the limited availability of defect samples, these methods struggle to meet the requirements for supervised detection. Therefore, we propose an unsupervised defect detection method for cigarette packaging based on multi-scale feature matching. This method only requires normal images during the training phase. By providing a pre-trained image classification network as the teacher network, knowledge across multiple scales is extracted through an attention module into a student network with the same structure but without pre-training, to learn the feature distribution of normal images. During the testing phase, when an image with a defect passes through the network, it will produce a multi-scale feature distribution different from that of normal images. The difference between the abnormal and normal feature distributions is then analyzed to detect and localize defects. This method achieved an average detection and localization accuracy of 98.1 % (PRO), outperforming existing unsupervised defect detection methods and providing robust support for the automated inspection of cigarette packaging.