부산대학교 도서관

In the traditional industrial production process, the sorting of product parts mainly depends on human eye detection, which has high production cost and low production efficiency. At present, automatic identification of parts has become a major trend of the reform and development of the manufacturing industry. However, current researches of metal parts identification mainly focuses on the identification of common parts, such as bolts, nuts, gaskets, etc.. The number of parts is small, and the identification accuracy still needs to be improved in these researches. Based on the improved two-stage object detection algorithm Faster RCNN and the one-stage object detection algorithms YOLOv3 and YOLOv5s, this paper realized the recognition of 40 types of metal parts. The recognition accuracy of all improved models was more than 97%, of which the lowest was YOLOv3 (97.3%), followed by Faster RCNN, with the recognition accuracy of 98.1%. The recognition accuracy of YOLOv5s model was the highest, reaching 99.4%, and the detection speed reached 179 frames per second. The results showed that the improved YOLOv5s algorithm achieved the highest recognition accuracy and recognition speed in the identification of metal parts, and could meet the needs of industrial scenes. The research in this paper could play an important role in promoting the application of these algorithms in the industrial field.