At present, the glass screen defects recognition mos tly rely on manual detection, eye long time will lead to visual fatig ue, thus affecting the accuracy of the glass screen defect detection and recognition of the time used for each piece of glass screen. T his paper proposes a machine vision-based glass defect detection method, for the glass screen scratches defective features are often presented inconspicuously, resulting in faulty detection, omission detection of the situation, using the detection process of two angl es of light with the left side of the light and 45 ° oblique angle of t he light, to solve the weak scratches presented inconspicuous pro blem. Data preprocessing is performed using grayscaling and ima ge differencing to remove redundant information. The algorithm uses the Yolov8 neural network, which is made more suitable for industrial machine environments by thinning the network and usi ng slim-neck to reduce the amount of Yolov8 network computatio n. The attention mechanism Triplet Attention is added after the S PPF module of the backbone network to ensure good detection. T he improved network weight file size drops by 11.3%, mAP0.5 ris es by 1.4%, and the detection speed is 13.2ms, which effectively re places manual labor to achieve fast and effective defect detection.