Unsupervised Visual Anomaly Detection And Recognition Based On Deep Learning

Developing computer vision models to detect and localize the unexpected or anomalous structures within images is very important and challenging,especially when there are no anomalous image samples or priors available and the anomalous patterns tend to appear in the limited local areas of images.Because it requires that the vision models discern the minor visual variations without any supervision.This detection task is significant and meaningful.On the one hand,since the research paradigm of our visual anomaly detection is beyond the assumption that building machine learning models in a static and closed environment,the relevant research will drive the development of artificial intelligence in both theory and application.On the other hand,visual anomaly detection has a broad application prospect.For example,in the field of intelligent manufacturing,it can be applied to defects detection;in the field of medical image analysis,it can be used to detect the lesions in medical images;in the field of intelligent security,it can be used to detect the abnormal events in videos.This thesis mainly studies the unsupervised visual anomaly detection techniques.Based on the relevant theories and technologies of deep learning,the thesis focuses on the problem of unsupervised anomaly detection for images.The research contents and main contributions are as follows:1.In view of the shortcomings of the current reconstruction-based image anomaly detection methods,a novel feature-based reconstruction mechanism is proposed to address the visual anomaly detection problem.By leveraging the discriminative and sparse deep convolutional image features,the feature-based reconstruction mechanism can be effectively used to detect and identify potential abnormal patterns in images.In order to enhance the representation ability of the convolution features,a pre-trained deep convolution network and a multi-scale image region feature generator are proposed to construct the multi-level and multi-scale representation of images,so as to improve the performance of visual anomaly detection.Besides,in order to effectively and efficiently reconstruct the multi-level convolutional features,a deep yet efficient convolutional autoencoder is proposed which then can be leveraged to detect anomalous patterns within images via fast feature reconstruction.The experiments demonstrate that the featurebased reconstruction method are able to detect and recognize the anomalies in local image region quickly and effectively.In addition,for image anomaly detection,different levels of convolution features may convey some specific useful information.2.To address the deficiency of the current feature-correspondence-based anomaly detection methods,a novel learning-based feature correspondence mechanism is proposed.The proposed correspondence mechanism is very different from the commonly adopted search-based correspondence mechanism,such as the nearest neighbor method.It is end-to-end and learnable,and can directly estimate the best correspondence among image features on-demand,so as to calculate the feature correspondence in real-time.To effectively implement the proposed feature correspondence mechanism,an asymmetric deep dual network framework is proposed,which consists of a visual feature extraction network and a feature estimation network.The dual network estimates the potential anomalies in the image by directly modeling and evaluating the feature correspondence between its two network branches.Besides,in order to improve the robustness of the correspondence mechanism,a novel feature enhancement strategy based on self-supervision learning and a multi-context residual learning network module is specifically proposed.The experimental results demonstrate that the proposed featurecorrespondence-based anomaly detection model can effectively discover and detect anomalies in images,especially for industrial anomalies.Compared with other image anomaly detection approaches,it achieves the best average detection performance both on the benchmark dataset and a real industrial anomaly dataset.3.A novel detection mechanism based on feature inpainting is proposed for the first time in the field of visual anomaly detection.Specifically,a deep feature inpainting approach that incorporates the priors of anomalies is developed.The approach is implemented with a cascaded model which consists of an anomaly mask estimation network,a deep feature inpainting network,and an anomaly scoring and detection module.Firstly,the anomaly mask estimation network captures the priors of possible anomalies within images by a rough end-to-end mask prediction of the potential abnormal regions in images.Then,the prior information is taken by the deep feature inpainting network to improve the feature inpainting performance.Finally,by evaluating the disparities of the features before and after the inpainting operation,the anomaly scoring and detection module can discover and detect the possible anomalies within images effectively.In addition,in order to further enhance the performance of the feature inpainting network,a new network learning module coined as the global semantic enhanced dilated dense block is specifically designed.Since the module can capture multiple different context information of the image,it further helps the feature inpainting network repairing the possible abnormal feature patterns,and generate reasonable normal feature patterns.The experiments demonstrate that the method can detect and recognize anomalies in local image regions effectively.In addition,it is found that the anomaly mask estimation network is also a satisfying anomaly detector though only trained with the synthesized anomaly samples.The findings will inspire new ideas for the research of image anomaly detection.