Nondestructive detection of surface defects of curved mosaic ceramics based on deep learning

Curved mosaic ceramics are the primary materials in creating three-dimensional mosaic ceramic artworks. However, they must be thoroughly inspected for defects to ensure the high quality of the final products. The unique small size, high reflectivity, and curved structure of these ceramics pose significant challenges for surface defects detection. Therefore, this paper proposes a deep-learning defect detection method based on an improved version of YOLOv7. Initially, highlights in the images are removed using minimum image fusion, followed by enhancement with the removed highlights. The model incorporates the Spatial and Channel-reconstruction Convolution (SCConv) module, the Centralized Feature Pyramid Network (CFPNet) module, and the Multi- Scale Dilated Transformer attention mechanism (Dilateformer) to create the YOLOv7-SCD model. Subsequently, defects are detected and identified. Experimental results indicate that YOLOv7-SCD increases detection accuracy and reduces missed detections compared to YOLOv7. The accuracy of this method in detecting surface defects in curved mosaic ceramics reaches 99 %, with a detection time of 18.3 ms per ceramic piece. Thus, it provides accurate and rapid detection of surface defects, meeting industrial real-time detection requirements.