Classification of Wafer Backside Images Via Fasterrcnn-Based Neural Network

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

Automatic classification of defect patterns in wafer backside image stands as a challenging problem in semiconductor manufacturing industry, especially different wafer backside defects start to impact the front side pattern defined and cause yield loss because of the design rule shrinkage. Deep neural network has recently shown decent progress in image classification, including wafer bin maps, but has rarely been applied in wafer backside image identification as far as we are aware. A Faster-RCNN based neural network model has been proposed for defect pattern detection and classification in wafer backside images. As a prepositive strategy, customized data augmentation has been proposed to diminish the expense of explicit manual supervision. Meanwhile, Class Activation Map (CAM) of each image has been generated to guide the training process. The results of this practical study showed the validity and practical viability of the model even at little expense of manual-labeled images.