CAGK-YOLO: An Efficient Detection Model for Micro-Scale and Irregular Surface Defects Based on Deep Learning

Surface defect detection is an indispensable step in the production and manufacturing process. However, existing methods are susceptible to imaging noise and often struggle to achieve satisfactory results in recognizing micro-scale and irregular defects. This paper proposes an efficient detection model CAGK-YOLO for real-time mechanical components micro-scale and irregular surface defects detection. Firstly, the Coordinate Attention (CA) module is added to the original YOLOv5 to augment the feature extraction capability for defect images. Then, we utilize the K-means algorithm to align the initial anchor frames of the vanilla YOLOv5 model more closely with the distribution of original defects, thereby enhancing the model's ability to detect irregular defects. Lastly, we replace the original three convolutions module in the backbone of YOLOv5 with the Ghost module to accelerate the defect detection process. We also integrate an up-sampling mechanism to further refine the accuracy in detecting micro-scale defects. The experimental results show that the CAGK-YOLO achieves superior performance with comparison models on the benchmark NEU-DET, NEU-Micro, and Samsung oil stain datasets. Moreover, ablation experiments are delivered to verify the validity of CAGK-YOLO's substructures.