A Light CNN Model for Defect Detection of LCD

Abstract

The quality control of LCD manufacturing process is very important to minimize cost and maximize product quality. This paper designs a light convolution neural network (CNN) model with fewer parameters that detects surface defects and identifies their types in thin film transistor liquid crystal display (TFT-LCD). We choose different sizes of small convolution kernels to extract the shallow features of the image, and introduce the sparse convolution structure to extract the multi-scale deep features. In addition, because there are few defects in actual industrial production lines, the defect sample numbers may be not sufficient for obtaining much more local features directly based on small sample learning. We use the deep convolutional generative adversarial network (DCGAN) to generate data. The network parameters of the designed model are updated by the re-learning the generated samples based on the original sample and DCGAN. The original small samples are directly sent to the deep feature extraction layer to further strengthen training, so that the model has the ability of continue learning. We make a series of experiments on the real images using the trained model combined with a sliding window technique to detect and classify the detects in the original images. The experimental results show that the model can effectively improve the detection rate and reduce the missed detection rate.

This research is supported by the key project of National Nature Science Foundation (U1604262).