| Anomaly detection is a hot research topic in computer vision,bioinformatics,and wireless security.However,most traditional anomaly detection methods usually regard all normal samples as one normal pattern due to the high cost of annotation.Ignoring the multi-pattern of normal samples in the real world will decrease the detection accuracy of traditional anomaly detection methods.Because detection methods are more difficult to capture the patterns of samples.In this dissertation,we consider the multi-pattern of samples to propose a MultiPattern Adversarial learning Network model(MPGAN),which introduces interpolation techniques to discover the latent patterns of samples.We theoretically analyze the advantages of interpolation technique during the detection processes.In the experiment part,MPGAN method shows its’ effectiveness in mining the pattern information.Although MPGAN can capture multiple hidden patterns of samples,the basic reconstruction error function of MPGAN influences the detection accuracy.In this dissertation,we propose a novel anomaly detection method called Multi-Pattern Anomaly Detection model(MPAD),which cooperates with multi-pattern information and a pattern classifier together to design a new reconstruction error function.In the experimental part,the MPAD achieves the highest detection accuracies in three real-world image datasets.In a summary,we introduce the interpolation technique to propose the MPGAN method to capture the hidden multiple patterns of samples.Moreover,we propose the MPAD method by designing a better reconstruction error function to further improve the detection accuracy.The experimental results show that our proposed methods have the best detection accuracies comparing with the latest six anomaly detection methods in MNIST,Fashion MNIST,and CIFAR10 datasets. |
