Research On Image Anomaly Detection Based On Feature Similarity

Anomaly detection is a classic machine learning problem.Anomalies are also called outliers,and are usually defined as samples that are significantly different from other data.The goal of anomaly detection is to learn a discriminative model that can separate anomalies among test samples well.Anomalies often mean system risks,malicious operations,and faulty products,etc.In order to ensure system works properly,many applications need to eliminate anomalies,therefore the anomaly detection task has a wide range of practical requirements.In recent years,deep learning has developed rapidly.The methods based on deep convolutional neural networks for anomaly detection has taken place of traditional machine learning methods and has become the mainstream in this field.The work of this thesis mainly focuses on the anomaly detection problem of image data,and conducts an in-depth study of the problem based on the features extracted from images by deep network.Image anomaly detection problem can be divided into image-level one-class classification task and pixel-level anomaly location task.In the data scenario of anomaly detection,the model can only observe normal samples during training,and lacks information about anomalies.Therefore,how to extract discriminative features from images has become a critical problem in anomaly detection.In this regard,this thesis proposes an anomaly detection method based on metric learning and a pixel-level anomaly detection method based on feature similarity.The model first extracts the features from images,and then calculates the image similarity in the feature space to solve the anomaly detection problem.Specifically,the main work of this thesis are as follows:This thesis reviews the existing methods on anomaly detection,including traditional machine learning methods and deep learning-based methods which are popular in recent years.This thesis first discusses the traditional methods and points out their advantages and limitations.For deep learning methods,this thesis divides them into three categories: autoencoders and their variants,methods based on adversarial generation networks,and methods based on features.This thesis also introduces the research status of each category and analyzes the advantages and disadvantages of different methods.This thesis reviews the history of anomaly detection methods,evaluates and analyzes the advantages and limitations of existing methods,and describes the problems and challenges to be addressed by the anomaly detection work in this thesis.For one-class classification task in anomaly detection,this thesis proposes an new method based on metric learning.This method uses the rotation transformation of the test image to construct a self-labeled dataset.Based on this,the method uses the framework of deep metric learning to map the sample to a suitable feature space,so that the distance between similar images in this space is shorter than that between dissimilar images.The method uses the convolutional neural network as the backbone model,and obtains the normal score according to the maximum feature similarity between the test images and the training set.Experiments on image datasets show that the method proposed in this thesis works well on anomaly detection tasks.For pixel-level anomaly region localization task in anomaly detection,this thesis proposes a new method based on feature similarity.Because of anomaly detection target dataset with few samples and large image similarity,the method in this paper uses a pre-trained network model as the image feature extractor.The method first preprocesses the training data,constructs a feature retrieval library,and then also performs feature extraction on the test images,and calculates the similarity between the test image block and the training data as the normal score.This thesis proposes two methods for calculating similarity and applies them to image patches at different scales.After that,the results at different scales are summarized to obtain the final pixel-level anomaly score.Experiments on the anomaly detection dataset show that the method proposed in this thesis can effectively locate anomalous regions in images.