An Improved Detection Network for Small Glass Defects Based on Enhanced Multi-scale Feature Extraction

With the increasing use of glass across various industries, intelligent glass defect detection has become a critical issue to ensure high-quality standards while minimizing labor costs. However, glass defects are featured by tiny size, resulting in poor detection accuracy. To this end, we adopt the advanced YOLOv8 framework for this task with specific improvements for Small Glass Defects, named SGD-YOLO. Specifically, we adopt the Enhanced Multi-scale Deformable Convolution (EMSDC) C2f to fully extract local fine-grained information and global contextual information of small objects by adaptive tuning the multi-scale convolution kernel. Our experimental results on the public Smartphone Screen Glass Dataset confirm the validity of our method, outperforming available methods in terms of accuracy and efficiency performance, especially for small glass defects.