A representation points based method for object detection on the surface of electroplated diamond wire saw

Abstract

Electroplated diamond wire saw (EDWS) is the main tool for slicing semiconductor materials such as monocrystal silicon, polycrystal silicon, and sapphire. The distribution density of abrasive and agglomeration on the surface of EDWS is an important index to evaluate its manufacturing quality. Limited by the diversity of object morphology in EDWS images, designing a unified feature extractor by traditional methods is pretty hardship. In this paper, a method named electroplated diamond wire saw detection network (EDWSDN) is proposed based on deep learning, which can adaptively extract the shape and posture of objects by using representation points. In addition, an adaptive positive points selection (APPS) module is designed to realize the quality evaluation and redistribution of positive samples. Extensive comparative experiments are implemented to verify the performance of the proposed method. Put the confidence threshold into 0.5, the proposed method can achieve 89.5 mean of Average Precision (mAP) while ensuring an inference speed of 21.5 Frame Per Second (FPS). The promising results demonstrate that the proposed method has good enough performance to be applied to the online detection of EDWS manufacturing quality.