Semi-supervised Image Anomaly Detection Based On Generative Adversarial Network

Anomaly detection is the identification of events or observations that do not match other items in a dataset.As a research field with wide application value,the effect of image anomaly detection has been greatly improved with the development of deep learning in recent years.Most of the traditional image anomaly detection algorithms based on deep learning are supervised algorithms,which need labeled data sets containing positive and negative samples to train the model,although it has achieved good results on standard datasets,however,compared with the supervised anomaly detection algorithm,the semi-supervised or unsupervised anomaly detection algorithm has more research significance and application value because of the imbalance of data and the complexity of anomaly.Because of its unique generator-discriminator structure,the Generative Adversarial Network(GAN)provides a new idea for anomaly detection algorithm based on Deep Learning: In Training,GAN only needs to learn the data distribution of normal samples.In the test,the difference between the reconstructed image and the original image is used to judge whether the input is an abnormal sample or not.Based on this idea,this thesis carries out the following research on anomaly detection:1)Two stage anomaly detection model based on GAN: in order to make the discriminator learn the subtle disturbance of the generator in reconstructing the abnormal image more accurately,a low-quality generator is added to the generation network to convert the discriminator task in GAN.In addition,a pseudo-anomaly module for data enhancement is added to improve the training effect of the discriminator.The principle of the improved model is consistent with that of the original GAN,so the improving method can be integrated into any existing one-class generative adversarial networks.2)In order to solve the problems of complex images,unclear details and large background interference in real application scenes,an improved GANomaly model based on skip connection is proposed.Skip connection is added to GANomaly model to learn image details and generate more accurate image space.In addition,the backpropagated gradient loss function is introduced to improve the detection accuracy.In order to verify the performance of the two anomaly detection models and combine the theoretical research with the practice,this thesis uses the coal mine data set collected in the real coal mine production environment to carry on the experiment.The two stage anomaly detection model based on GAN and the improved GANomaly model based on skip connection have achieved the accuracy of 71.3% and 77.4%respectively.The average accuracy of the improved GANomaly model based on skip layer connection is 70.1% in all anomaly classes of coalmine dataset,which is better than other anomaly detection models based on GAN.