Research On Spatio-temporal Anomaly Detection Algorithms Of Image Data

The algorithms of image anomaly detection usually need to consider the temporal and spatial characteristics of the image data.In the existing research,anomaly detection based on the temporal characteristics of image data mostly require pixel-level annotations of abnormal objects and cannot detect unlabeled anomalies,the algorithms without pixel-level annotations have complex model structures and abnormal judgment methods.The anomaly detection algorithms based on the spatial characteristics of image data,including the methods which based on the reconstruction and cluster.The anomaly detection based on the reconstruction have better results,but the model structure are more complex.In order to solve the above problems,our research based on the temporal and spatial characteristics of image data respectively.The main research contents are as follows.(1)The thesis proposes an anomaly detection algorithm which based on the image time series prediction without pixel-level annotations.In order to avoid the pixel-level annotations of abnormal objects and not increase the complexity of the model structure,also use the simply and effectively method to detect the abnormal objects.The thesis studies the video frame anomaly detection method which based on the prediction model,and designs the shallow one-dimension convolution and shallow two-dimension convolution time series prediction models respectively.These model make full use of the video data properties to predict different video frames and the model structure are simple.This way of detect the anomalies by predicting the difference between the predict frame and the actual frame does not require pixel-level annotations,and the abnormal judgment method is simple.The thesis studies the fusion method which based on the prediction results of various time series prediction models,so as to make full use of data characteristics and improve the accuracy of anomaly detection.(2)In this thesis,an anomaly detection method by improved GANomaly which based on the spatial characteristics of image data is designed.The anomaly detection effect based on GANomaly is better.GANomaly model can ignore the temporal characteristics of video and detect anomaly,but the structure of the model is complex.Then,in order to maintain the effect of anomaly detection and simplify the network structure,the thesis studies the anomaly detection model based on improved ganomaly,simplifies the network structure based on the feature representation ability,and designs a new loss function for the simplified structure.(3)Applications analysis and test of anomaly detection algorithms.For different types of video anomalies,the applicability of the model which based on the temporal and spatial characteristics of image data is studied.The thesis synthesizes the anomaly detection results of temporal models and spatial model to improve the anomaly detection capabilities of various anomaly scenes.