Abnormal Image Detection Method Based On Deep Learning

Abnormal image detection is an important branch of computer vision.Its purpose is identifying data that does not fit the normal pattern from a large number of samples.It has certain research significance and application value in the surface defect detection and hyperspectral image processing,etc.Current deep learning-based algorithms for anomaly image detection is usually better than traditional methods in efficiency and stability.However,models trained by deep learning often have the disadvantages of low stability,easy model collapse,and low discrimination between normal points and outliers.Aiming at the above problems,this paper proposes an improved algorithm based on fAno Gan.The main research works and achievements are as follows:(1)Optimize the anomaly detection architectureBy studying the basic principles of various Gans and combining them with fAno Gan algorithm,anomaly detection is carried out in a unified configuration environment,and then the experimental results are analyzed.Finally,the BEGAN architecture is used to replace the original WGAN-GP,which improves the detection performance.(2)Improved BEGAN model.Although the performance of BEGAN is better,but it still has some flaws,therefore it has been improved to some extent.It includes the modification of the network layer and activation function,and the replacement of the original single time scale by the double time-scale update rule,and then the Pulling-away Term is used to add the PT term into the loss function to reduce the generation of similar images.The idea of self-attention is combined with the model to improve the whole of the image quality.(3)Select the appropriate calculation method of abnormal scoreIn order to solve the problem of low discrimination of the model,this paper studies a variety of methods to calculate anomaly scores,including using f-Ano Gan’s outlier definition,f-Ano Gan’s pixel space loss,f-Ano Gan’s discriminator loss,and SkipGanomaly’s outlier definition as the measurement standard,respectively.Experiments were carried out under the same conditions,and the results were visualized.Finally,the positive abnormal data were divided only by the image distance of f-Ano Gan.The final improved model f-Ano Gan-III is compared with f-Ano Gan-I,f-Ano Gan-II and the classic anomaly image detection algorithms Ganomaly and Skip-Ganomaly respectively on the pavement crack data set and concrete crack data set SDNET2018.AUC values of 0.90 and 0.77 were achieved,and its performance in major evaluation indicators such as EER was better than other models,which verified its excellent performance and strong generalization and anti-interference ability.