Intelligent Glass-Defect Detection Using Convolutional Neural Networks

Inspection of industrial glass has to be very accurate as even small scratches, bubbles or marks could highly undermine its structural reliability but manual checking and traditional vision systems are incompatible with the process, must be protected from glare sensitivity (REF) and do not generalize well when new type of defects appear. For deep models, the accuracy is significantly better than, however most require expensive computation or well-controlled lighting. To bridge these gaps, a lightweight convolutional neural network method is proposed that features multi-scale feature fusion, gradient-field enhancement, and focal loss to highlight the subtle abnormalities. Experiments on different lighting and defect types achieve 99.1% accuracy, 98.9% mAP, 99.0% precision, and 98.7% recall at a speed of 70 fps on an edge device. The analysis indicates that the false rejects declined by 31% in contrast to rule-based baselines, which can be deployed practically for online inspection. Increased robustness under reflected noise and varied backgrounds strongly suggests potential to enhance manufacturing yield and reduce expensive post-processing failures.