As a regression-based target recognition algorithm, YOLOv3 can achieve rapid and accurate recognition of multiple targets. YOLOv3 trains on the global region of the input image, which speeds up training and allows for better discrimination between targets and background. First, the Darknet-53 backbone network is used to extract the surface defect features of aluminum profiles, and then the target frame is used to directly predict the target category and location. The surface defects of aluminum profiles are irregular in shape, random in location and vary in size. It is difficult to accurately identify small defects by directly applying the YOLOv3 model for identification.
This article improves the YOLOv3 model based on an in-depth analysis of its characteristics. Expand the original 3-scale recognition structure to 4 scales to improve the ability to identify small defects; construct an initial target frame suitable for aluminum profile surface defects through re-clustering analysis, and improve the model parameters of the YOLO algorithm; use multi-scale training methods to train The process is optimized to enhance the model's adaptability and recognition accuracy to defects of different scales, and to solve problems such as difficulty in identifying surface defects and low accuracy of aluminum profiles.