A Light-Weighted CNN Model for Wafer Structural Defect Detection

Abstract

Silicon wafer is the raw material of semiconductor chip. It is important and challenging to research a fast and accurate method of identifying and classifying wafer structural defects. To this end, we present a novel detection method in terms of the convolution neural networks (CNN), which achieve more than 99% detection accuracy. Due to the wafer images are not available by open datasets, a set of imaging acquisition system is designed to capture wafer images. Digital image preprocessing technology is utilized to split a wafer image into thousands of silicon grain images. The proposed model, called WDD-Net, uses depthwise separable convolutions and global average pooling to reduce parameters and calculations, adopts multiple 1*1 standard convolutions to increase the network depth. Specifically, two types of CNN models, VGG-16 and MobileNet-v2, are adopted for comparative analysis. Using the aforementioned three models, the comparative experiments are implemented on data sets that consisting of more than ten thousand grain images. The experimental results show that compared with VGG-16 and MobileNet-v2, the detection speed of the WDD-Net is 105.6FPS, which is 5 times faster. The model size of the WDD-Net is 307KB, which is much smaller than the other two. Furthermore, the WDD-Net directly completes the data collection and defect detection process through the local computing equipment, which is suitable for edge computing.

Keywords

• Image classification
• neural networks
• semiconductor manufacturing
• machine learning