Research On Unsupervised Anomaly Detection Algorithm Based On Image Reconstruction

Analysing the abnormal samples has attracted great reasearch attention in many fields,e.g.,industrial defect detection,medical image diagnosis and remote sensing hyperspectral image analysis.However,the captured abnormal samples are various and they are of a small quantity,which make it extremely difficult to construct a supervised anomaly detection model in the real-world application.To overcome these difficulties,the unsupervised anomaly detection methods have been widely used and become one of hot researches in the field of computer vision.Unsupervised anomaly detection aims to distinguish abnormal samples that are globally or locally different from the training set in the test stage.It can effectively alleviate the lack of abnormal samples and has high application value.Recently,the reconstruction-based anomaly detection methods have made great progress.However,the reconstruction-based anomaly detection methods suffer from the inference of shallow features and the overexpression of neural networks.Thus the reconstruction-based methods cannot meet the requirements of real-world application scenarios.The thesis focuses on the reconstruction-based anomaly detection method,and conducts research on the aforementioned problems:(1)To alleviate the inference of shallow features,this thesis proposes an unsupervised anomaly detection method based on representation transform perception.Firstly,the thesis performs a parameterized affine transformation on the input image,and gets the input-transformed image pair.Then,two additional decoders are introduced to predict the parameters of image transformation.Finally,the image transformation learning encourages the anomaly detection model to learn the transformation-detectable features of the normal image in the content level and structure level.Through the transform prediction learning,the image feature contains high-level semantic information and rich intrinsic attribute information,which effectively alleviates the inference of shallow features.Furthermore,a transformation consistency constraint is proposed to maintain the attribute consistency between the input image and the generated image.The consistency constraint further improves the reconstruction quality of the normal image.The comparative experiments on the defect dataset MVTec-AD and the semantic dataset CIFAR10 validate the effectiveness and robustness of the proposed method.(2)To alleviate the overexpression of neural networks,this thesis proposes an unsupervised anomaly detection method based on representation compression.Firstly,the method discretizes the latent features through a finite quantizer.The quantification process removes the information redundancy of the latent features,and learns the compact representations of the normal samples.Through the quantification compression,the network’s overexpression is constrained,and the expression difference between normal samples and abnormal samples is enhanced.As a result,the abnormal images cannot be corrected reconstructed,which increases the reconstruction error of abnormal images.Secondly,the skip-connection approach is adopted to increase the reconstruction quality of normal images by reusing the low-level features and the highlevel features.This method simultaneously meets the two requirements of abnormal detection expression differentiation and reconstruction clarification,which effectively alleviates the overexpression of the network.A large number of experiments conducted on two public datasets MVTec AD and CIFAR10 show that this method can effectively alleviate the overexpression of neural networks and improve the performance of unsupervised anomaly detection.