Wafer Surface Defect Detection Based On Improved YOLOv3 Network

In the semiconductor field, wafer defects often reduce the yield of the chip, thereby affecting the performance of the chip. Aiming at the problems of low accuracy and insufficient real-time performance of traditional wafer defect detection technology, a wafer surface defect detection method based on improved YOLOv3 network is proposed. First, the k-means algorithm is used to re-define the a priori box for clustering to increase the scale adaptability. Then, the Darknet-53 feature extraction network is improved, Mish activation function is added to improve the stability and accuracy of the network, and the SPP module is introduced to solve the problems of local information missing and uneven size. Finally, the SPP module is introduced to finally improve the multi-scale feature fusion network, adding a larger-scale prediction, and improving the detection accuracy of the network for small targets. Experiments show that the average detection accuracy of the algorithm model in this paper is about 13% higher than that of the YOLOv3 model, and the detection speed has also been greatly improved. Compared with some other existing target detection algorithms, this method has better detection effect and robustness, reduces the false detection rate, and can be used in related industrial fields.