A Mura Detection Model Based on Unsupervised Adversarial Learning

Mura is a phenomenon in which the displays have various uneven display defects and has the characteristics of irregular shape and different sizes. For Mura detection, traditional detection methods have the following two problems: One is the problem of data imbalance. The second is that new shapes an...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2021, Vol.9, p.49920-49928
Main Authors: Song, Shubin, Yang, Kecheng, Wang, Anni, Zhang, Shengsen, Xia, Min
Format: Article
Language:English
Subjects:
Anomalies
Anomaly detection
Artificial neural networks
Convolution
Decoding
Feature extraction
Gallium nitride
Generative adversarial networks
Generators
Image reconstruction
Inspection
Mura
Normal distribution
Reconstruction
reconstruction error
Res-unetGAN
Training
unsupervised anomaly detection
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Mura is a phenomenon in which the displays have various uneven display defects and has the characteristics of irregular shape and different sizes. For Mura detection, traditional detection methods have the following two problems: One is the problem of data imbalance. The second is that new shapes and sizes of Mura may appear at any time during the inspection process. 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. The generative network is an autoencoder structure composed of ResNet50 and UNet to learn the normal distribution of normal samples. The discriminator is a convolutional neural network based on deep separable convolution and forms a game process with the generator. The network only needs normal samples during the training process, and the network is optimized by the error loss between the original samples and the reconstructed samples. In the test, a reconstruction error score will be designed according to the reconstruction quality, and the defect in the sample will be judged by the reconstruction error score, so as to achieve the goal of anomaly detection. After repeated experiments on the Mura data set, the detection accuracy of Mura defect is better than that of several models compared. The proposed model has a unique application prospect in other industrial anomaly detection since it only requires normal samples for training.
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2021.3069466