Surface defect detection of ceramic disc based on improved YOLOv5s

Abstract

Addressing the challenges in detecting surface defects on ceramic disks, such as difficulty in detecting small defects, variations in defect sizes, and inaccurate defect localization, we propose an enhanced YOLOv5s algorithm. Firstly, we improve the anchor frame structure of the YOLOv5s model to enhance its generalization ability, enabling robust defect detection for objects of varying sizes. Secondly, we introduce the ECA attention mechanism to improve the model's accuracy in detecting small targets. Under identical experimental conditions, our enhanced YOLOv5s algorithm demonstrates significant improvements, with precision, F1 scores, and mAP values increasing by 3.1 %, 3 %, and 4.5 % respectively. Moreover, the accuracy in detecting crack, damage, slag, and spot defects increases by 0.2 %, 4.7 %, 5.4 %, and 1.9 % respectively. Notably, the detection speed improves from 232 frames/s to 256 frames/s. Comparative analysis with other algorithms reveals superior performance over YOLOv3 and YOLOv4 models, showcasing enhanced capability in identifying small target defects and achieving real-time detection.