Study On Defect Detection And Classification Of Wafer SEM Image Based On CNN

With the rapid development of semiconductor industry,the critical size of transistors on chip has become smaller and smaller,bringing great difficulties to integrated circuit manufacturing.In order to cope with the increasing numerous defects of different types on the wafer surface,it is necessary to study more accurate and faster defect detection and classification algorithms.Based on images of nanoscale resolution imaged by scanning electron microscope(SEM)from wafer surfaces,the problem of wafer defect classification has become a problem of SEM image classification,and the problem of wafer defect detection has become an object detection problem.This thesis first introduces the basic structure and principles of convolutional neural networks(CNN),then the "ZFNet" image classification algorithm,and finally the improvement and development of object detection algorithm from Region-CNN to Fast-RCNN and then to Faster-RCNN.Aiming at the problem of defect classification of wafer SEM image,we applied a CNN named"ZFNet" to classify the defect regions of SEM image.The SEM image data contains 9 types of defects and non-defect,meaning 10 types altogether,and the test F-score reaches 97%.The"ZFNet" defect classification algorithm we proposed is accurate with strong data adaptability.Aiming at the defect detection problem of wafer SEM image,this paper implements a"patch-based CNN" defect detection algorithm based on "ZFNet" defect classifier,which can detect the location and type of defects from SEM images at the same time.In order to improve the accuracy and speed of detection algorithm,in view of the particularity of defect detection problem,another defect detection algorithm is implemented by modifying the region proposal network(RPN)structure in "Faster RCNN".The SEM image data contains 9 types of defects,and the F-score of detection test is 92%,meanwhile the consuming time is only 5%of patch-based detection algorithm on same platform.The "Faster RCNN" defect detection algorithm we proposed is accurate and fast with strong data adaptability.