Machine Vision-Based Defect Classification Algorithm for Rolled Packages

Under the growth of industrial digitization, networking and intelligence, more and more industrial enterprises are undergoing digital transformations. In the tobacco industry, potential issues may arise during the packaging phase, leading to packaging defects. In the traditional process, the camera shoots at a fixed point and compares the designated focus area image with the standard image to determine whether it is defective or not. This thesis investigates a defect classification method dedicated to the field of tobacco production, using machine vision algorithms to detect the images of cigarette packaging. The objective is to determine the presence and the type of defects, which is conducive to realize a automated and intelligent production mode. The method combines traditional computer vision techniques with deep learning models, and contains two phases: first, traditional CV methods are directly used for some defects with obvious features and some positional offset defects, which improves the efficiency and accuracy, and facilitates the subsequent classification; second, deep learning methods are used for the remaining defects. For the multi-label classification, a defect classification network is constructed and trained based on ResNet-34. After experiments and research on all the defects (12 types), this method can meet the needs of defect recognition.