Research On Visual Inspection Of Wafer Surface Defects

Various defects are prone to occur in the production process of semiconductors.It is necessary to find defects early in the production process,find out the cause of defects,and discard the defective samples.It prevents the defective dies from continuing to be processed and affects production efficiency.In the method of wafer defect detection,automatic optical inspection(AOI)is a fast and low-cost visual inspection technology.Therefore,this paper focuses on the AOI system.This paper introduces the requirements and mechanical structure of the wafer defect detection AOI system,and develops the inspection process plan and communication scheme for the mechanical structure.In this paper,the gray-scale template matching algorithm is used to extract the die image on the wafer.For the mapped image after template matching,the candidate die position is obtained by threshold segmentation and gray maximum value,and the method of non-maximum value suppression is adopted.For the contamination defect,it is proposed to extract the contaminants at the edge feature with gradient images.If clear and blurred die images in the same camera image,the edge feature is correctly connected by the region growth method.For the scratch defect,the scratch skeleton is extracted by the Zhang-Suen thinning algorithm.According to the direction of the skeleton and the slope of the skeleton fitting straight line,the same scratch is connected and the different scratches are divided.For the ink dot defect,the defect is extracted according to the area of the connected component and the aspect ratio of the minimum circumscribed rectangle.For the poor cutting defect,the inclusive region of the minimum circumscribed rectangle of the connected area is used to merge the connected components,and the number of blocks with poor cutting separation is obtained according to the number of merged connected component blocks.For the corner breakage defect,the line of the edge of the die image is extracted by the Hough line,and the defect is determined according to the length of the line,and the position of the defect is determined by the distance between the end point of the line and the edge of the image.For the texture of the wafer image,the L0 gradient minimization algorithm smoothes the texture and preserves the defective parts and edges,and analyzes the smoothing effect.This paper designs the wafer defect inspection process and develops the appropriate software.The traditional machine vision algorithm detection can only be used for specific application scenarios,and it is difficult to detect complex crystal dies and defects.Based on the current good effect deep learning model YOLO v3,this paper detects the more complex scratches and contaminant defects on the die and the location and size information of the defect is extracted.This paper modifys the parameters of the network to make it more suitable for wafer defect detection.The experiment proves the feasibility of the above template matching algorithm and the algorithm of grain dot,contamination,cornering and poor defect detection by actual samples.Through the actual defect samples,it is proved that the YOLO v3 model can detect defects with complicated shapes,large numbers,and relatively small sizes on the dies.For the device communication needs,this paper designs the workflow and software based on the SECS/GEM communication protocol,and customizes the wafer defect data format.