부산대학교 도서관

Compared with large and medium targets, the resolution and characteristic information of small targets are limited, resulting in low recognition accuracy of such small target detection. This paper proposes a tube product defect detection method based on block processing to identify and detect missing prints, indentations and scratch defects in tube product images collected by industrial line scan cameras. Consider the working principle of line scan camera, establish a machine vision imaging model, analyze the relationship between the acquisition accuracy, travel frequency and motion speed of the measured object of the camera, and use the camera to collect pictures of tube products to form a data set. Since the scale of the research object to the image is small and cannot be directly used for detection, the data set pictures are first chunked, then the improved YOLOv4-Tiny network model with the addition of CBAM attention mechanism is entered at the same time to detect the recognition of three types of defects, and finally the detection results of each image block are comprehensively evaluation. The experimental results show that the proposed surface flaw detection method of tube products based on block processing has an mAP of 97.46% and a detection speed of 31.59 Fps for three types of defects, which can meet the requirement of small object detection.