Research On Semi-Supervised Image Anomaly Detection Method Based On Auto-Encoder Network

Image anomaly detection is an important issue in the field of computer vision.Many algorithms used in image analysis are susceptible to out-of-distribution samples,leading to incorrect and overconfident decision-making.Therefore,anomaly detection is also one of the biggest challenges for safe and robust deployment of machine learning algorithms in various fields.Supervised learning methods based on a large number of labels have problems such as difficulty in data labeling and poor generalization performance.For anomaly detection tasks,abnormal situations are diverse and uncertain,and it is often difficult to obtain a large amount of abnormal label data.The use of semi-supervised learning methods to solve the problem of image anomaly detection has a good research prospect.Based on the semi-supervised self-encoding network,this paper studies the problem of image anomaly detection,and explores the application and advantages of the semisupervised learning method based on the self-encoding network for the two major fields of medical image and industrial image.According to the image characteristics of the two fields,the corresponding image reconstruction strategies are designed.In addition,a medical abnormal image data set is constructed.The effectiveness of the proposed method on the standard data set is verified through experiments,and the effect of demarcating abnormal regions is achieved while detecting abnormal samples.This paper studies the semi-supervised learning anomaly detection algorithm based on self-coding network.The main work contents and innovations include:1.The basic flow and framework of image anomaly detection using semi-supervised learning method based on self-coding network are studied.For the medical image field,a foreground mask reconstruction strategy based on super-pixel blocks is designed to improve the restoration and reconstruction ability of self-coding network for normal medical images.2.Medical image anomaly detection data set is constructed.Medical image anomaly is divided into two categories: natural anomaly and synthetic anomaly.Brain MRI anomaly samples are constructed and combined with normal samples as the test set of medical image anomaly detection task.3.The characteristics of industrial images are studied,and an anomaly detection method based on semi-supervised learning is designed with self-coding network structure,which realizes the reconstruction strategy of generating input images directly through neural network without manual preprocessing.Experimental results show that the proposed method has good performance on standard data sets.