Enhanced Sagger Crack Detection Integrating Deep Learning and Machine Vision

In recent years, target inspection has found extensive utilization within the industry, making it crucial to detect defects in industrial products to ensure quality. To address the challenges posed by large brightness differences, attached dirt, and complex backgrounds in saggers, we propose a sagger defect recognition method that integrates deep learning target detection and machine vision feature extraction. This method commences by employing the photometric stereo method to construct a curvature map of the sagger surface, reducing the interference from brightness differences and dirt. Next, an improved YOLOv5s target detection model uses the surface curvature map as an image source for crack detection. The model incorporates the Faster Block module in the backbone network and an efficient coordinate attention mechanism, embedding position information into channel attention to enhance the model’s understanding of crack defects. Finally, the method extracts crack geometry features from the target region, using feature scoring to confirm whether a crack defect is present. Compared with existing methods, this approach provides a new solution for detecting sagger cracks in complex backgrounds. Field applications and test results demonstrate that this method effectively improves the accuracy of sagger crack defect recognition.