A band-shaped Mura detection method based on unsupervised deep learning

摘要

Mura is a phenomenon in which the displays have various uneven display defects. The band-shaped Mura has the characteristics of irregular shape and different sizes. And the new shapes and sizes of Mura may appear at any time during the inspection process. Therefore, traditional image algorithms are difficult to detect the band-shaped Mura anomaly. In response to the above problems, this paper proposes the Res-unetGAN network, which is an unsupervised anomaly detection method based on generative adversarial network. We design resnet50 as the encoding network of the generator to obtain the latent feature vectors. To improve the quality of reconstructed samples, we combine the skip-connection structure into the generator to guide the decoder. The discriminator is a convolutional neural network based on the Depthwise Separable Convolution. The purpose is to distinguish between normal samples and reconstructed samples, and form a game process with the generator. The network only needs normal screen samples during the training process. In the test, since the Mura sample has not been trained, the reconstruction error score of the Mura sample will be higher. After repeated experiments on the band-shaped Mura data set, the highest auc of 0.995 was obtained, which is better than several models for comparison.