Detection and Recognition of Shallow and Fine Scratches on Mobile Phone Screens via CGM-YOLOv8

With the development of science and technology, cell phones have become an indispensable part of people's daily lives, and as a direct link between human–computer interactions, the quality of the user's experience directly affects the user's experience, which is a key factor in the sales of manufacturers. To ensure the production of high-quality cell phone screens, in the production process, cell phone screen defect detection, in which the screen surface of shallow fine scratch defect detection is difficult, is particularly important. The existing algorithms do not have high accuracy in detecting shallow scratches on screen surfaces, and the rates of misdetection and omission are high, which cannot meet the production demand. To improve the accuracy of shallow screen scratch detection, a shallow screen scratch detection network based on YOLOv8, CGM-YOLOv8, is proposed. A C2f module is added at the front end of the backbone network, which reduces the number of computations and can capture the features of the inconspicuous target more. The three traditional C2f modules at the back end of the backbone network are subsequently improved with Global Group Coordinate Attention (GGCA), which is designed to highlight inconspicuous features by capturing feature information in both the height and width dimensions to make the network more focused on the feature representations of shallow and fine scratches. Multiscale Dilated Fusion Attention (MDFA) is used instead of the traditional C2f module in the head network part, and the MDFA module fuses the attention mechanism through multiscale null convolution, which can significantly expand the sensory field of the network and enhance the comprehensiveness of extracting the feature information of the shallow and fine scratches. These methods can effectively improve the efficiency of shallow and fine scratch detection on the screen and improve the precision by 4.9%, the recall by 6.27%, and the mAP@0.5 by 4.8% compared with traditional YOLOv8.