An encoder-decoder-based image segmentation method for abrasive height detection of diamond wire
Abstract
The diamond wire (DW) is widely used in the field of slicing semiconductor materials such as monocrystalline silicon and polycrystalline silicon. The quality of the wafer is greatly affected by the protrusion height consistency of the abrasives consolidated on the surface of the DW. However, the online detection accuracy of the abrasive protrusion height is still unsatisfactory, which is mainly caused by the rough DW image segmentation results. In this paper, a diamond wire image segmentation model (DWISM) is proposed based on deep learning. To improve the segmentation accuracy of the DW image, a hard-points selection component is designed to adaptively select difficult-to-predict pixels in high-frequency regions such as the edge of the abrasives. The feature vector extraction and category re-prediction of these points are implemented in a specially designed refine head. Extensive comparative experiments are implemented to verify the performance of the proposed method. The results show that DWISM could achieve 90.8 mean Intersection over Union (mIoU) with 11.22 Frames Per Second (FPS) inference speed. The promising results demonstrate the great potential of DWISM for improving the protrusion height online detection accuracy of the abrasives consolidated on the surface of DW.
