High-Resolution Guided Up-Sampling Edge-Enhancing Semantic Segmentation Network for Semiconductor Wafer Surface Defect Detection

This is the authors' abstract. We don't add key points for this paper.

Semiconductor wafers are the cornerstone of the intelligent electronic devices, and the precise fabrication of them is regarded as the crown of the manufacturing industry. With the total size of the semiconductor wafers getting smaller and their packaging process getting more complex, it becomes more and more difficult to inspect the surface defects of semiconductor wafers. Herein, a high-resolution guided up-sampling (HRGU)-based semantic segmentation network is demonstrated for high-accuracy wafer surface defect inspection. First, a novel HRGU block is developed to recuperate the information loss caused by the down-sampling of the encoder, by guiding the up-sampling process in the decoder. Second, a boundary refine module (BRM) taking the image gradient map into account is designed to further improve the segmentation performance at edges. Finally, a boundary focal loss (BFLoss) is utilized to constrain the training process. Experiments are carried out on several typical open industrial defect datasets and our own wafer surface defect datasets. The proposed network exhibits a more superior detection performance compared to the other classical traditional counterparts, achieving a high segmentation metric mIoU of 80.71%, 87.05%, 91.23%, and 94.18% on the KolektorSDD dataset, Magnetic Tile dataset, and two our own wafer surface defect datasets gathered from industrial production lines.