Traditional industrial inspection relies on the human eye, which not only results in low productivity but also creates problems for production lines due to factors such as relying too much on a person’s subjective perception. Given the widespread application of computer vision, an increasing number of algorithms have been introduced into the field of industrial product inspection, aiding in the efficient development of both light and heavy industries and alleviating worker stress. Based on the YOLO (You Only Look Once) architecture, object detectors continue to iterate versions, propelling significant advancements in the field of target detection, showcasing its powerful advantages. In this paper, we selected the SSGD dataset to detect defects in the surface process of smartphone touchscreens. Firstly, the dataset undergoes data augmentation, followed by the introduction of the YOLOv8 algorithm for detecting defects in smartphone touchscreens on the assembly line and categorizing these defects. Additionally, we propose the BRA-YOLO architecture, which integrates BRA (Bi-Level Routing Attention) into YOLOv8. It employs dynamic sparse attention, focusing solely on relevant label content, thus achieving high efficiency. Moreover, it is memory- and computation-friendly, as it disregards irrelevant information and prioritizes essential details. By emphasizing attention on small targets and selecting the appropriate p2 model for small targets, the experimental results achieve an mAP50 accuracy of 71.3%.