Wafer maps defect recognition based on transfer learning of handwritten pre-training network

Abstract

A new method of defect recognition based on handwriting classification pre-training transfer learning is proposed for efficient defect recognition of wafer maps. Using deep learning can better identify wafer map defects, but one problem that a large amount of training data is needed to build a deep network model that can effectively recognize complex images. Another problem is the uneven distribution of the number of defect types in the wafer map during the actual manufacturing process. To solve these two problems, this paper uses the deep convolutional neural network trained in the MNIST dataset for handwriting classification as a pre-trained network for transfer learning. Since the MNIST data set contains simple basic patterns such as lines and circles that are common in the defect mode of wafer maps, the training of a deep network requires fewer data. Moreover, transfer learning reduces the amount of data for deep network training by transferring the parameters of the pre-trained network to the pattern recognition and classification model of wafer maps. This method uses a tenfold cross-validation method to verify multiple sets of different size subsets of the WM811K data set. The average recognition accuracy of each group of 10 experiments is above 94.9%. It has good recognition effect.