An online automatic carbide insert high-resolution surface defect detection system based on template-guided model

Abstract

Carbide insert is a fundamental tool in manufacturing, and it has been widely applied to cut raw materials or machine workpieces. In the production process of carbide insert, surface defect detection system plays a crucial role. Current general-purpose defect detection methods remain challenging due to the low efficiency of high-resolution images and high diversity of carbide inserts, which affect the practical application in manufacturing. In this paper, we propose a specifically designed carbide insert defect detection algorithm based on template-guided framework called TG-Net to address these issues. In contrast to previous general-purpose approaches that merely encode the defect image, we innovatively utilize a template image to guide the entire prediction. First, a siamese lightweight network is employed to extract multi-level features of the reference and defect image-pair. Then, the context and template guided attention module is adopted to fuse adjacent feature maps guided by difference maps at all levels, which promotes effective information to propagate from high-level feature maps to low-level ones. Benefiting from learning the difference information between image-pair, our algorithm can rapidly generalize to new types of carbide inserts without training again. On our carbide insert dataset, the proposed method yields the best prediction accuracy of 38.80% with the least parameters and reaches a real-time inference speed of 5.03 frames per second (FPS) on an image of 5120 × 5120, indicating that our approach achieves a trade-off between accuracy and efficiency when handling high-resolution images. Furthermore, a hardware carbide insert detection system is proposed, integrating the TGNet algorithm and deployed in the practice of production, demonstrating the effectiveness of our system.