Detection model of electroplated diamond wire surface abrasive distribution characteristics based on YOLO-EDW

Abstract

Electroplated diamond wire (EDW), a primary cutting tool for hard and brittle materials such as monocrystalline silicon, polycrystalline silicon, and sapphire, demonstrates processing efficiency and quality that are directly influenced by its surface abrasive distribution density. While computer vision-based methods have been proposed for efficient abrasive detection, current approaches exhibit limited accuracy in handling complex scenarios involving small abrasive sizes, dense distributions, and occlusions. This study introduces YOLO-EDW, an enhanced YOLOv8-based object detection model that utilizes EDW scanning electron microscope (SEM) images. The proposed model significantly enhances detection accuracy for small, densely distributed, and occluded abrasives through the integration of four key components: the Separation Enhanced Attention Module (SEAM), Spatial Depth Convolution Module (SPD-Conv), Optimized Object Detection Layer, and Wise-IoU Loss Function. Experimental results on the EDW-SEM dataset demonstrate that YOLO-EDW achieves improvements of 6.61 % and 7.16 % in mAP50 and mAP50:95 metrics, respectively, compared to the original YOLOv8. Compared with eight mainstream models such as FCOS-nas, YOLOF, and Faster-RCNN, YOLO-EDW shows excellent performance in dealing with abrasive densities and occlusions. This study achieves precise identification of abrasive distribution features, enabling efficient quantification of effective abrasives along the EDW's semicircular circumference and establishing a robust foundation for accurate abrasive density calculations.