The Anomaly Analysis Of Visible And Hyperspectral Images

With the development of various imaging devices and application requirements,image data experiences the development from the panchromatic image,color map,multispectral image to hyperspectral image.These rich images provide the powerful data support for the processing of different practical problems.Among them,owing to its unique characteristics the abnormal data analysis is a kind of important research problem.Abnormal analysis has different forms in a number of different image spectrums.Specifically,it mainly refers to abnormal behavior detection based on the video content,which corresponds to the anomaly target detection in hyperpsectral image.Anomaly detection in crowd scene is a hot topic in the current intelligent surveillance field.Analyzing the behavior of the crowd and responding to the anomaly behavior timely,is able to strengthen public administration and public security and further promote the smooth progress of the Safe City.Besides,Hyperspectral anomaly detection is to locate the pixels as anomalies whose spectral characteristics significantly deviate from the background spectrum by fully utilizing the spectral characteristics without any spectral priors about the target and background.This property of the unsupervised technique makes hyperspectral anomaly detection have a wide of practical applications.Therefore,it is of great research value to study the effective anomaly analysis of visible light and hyperspectral images by computer vision and image processing techniques.This thesis aims to develop crowd anomaly detection and hyperspectral anomaly detection approaches,and it involves:1)The crowd scene is complex with high density and diverse activities,which makes anomaly detection greatly difficult.The tracking based methods aiming to track every individual in the crowd scene is still a challenging task.Tracking technology relied on algorithms are difficult to accurately track each individual,which will cause great detection error.And the methods rely heavily on the availability of a large number of labeled training data,whose requirement is difficult to be strictly satisfied in practice.Taking fully into account the obstacles faced by the two kinds of anomaly detection approaches,a novel online learning method is proposed in this work.Inspired by the interact influence of individual behavior and the crowd group movement trend,the adaptive selection of experts and the dynamical update of the model are adopted to avoid an exhausting training stage and a long-term tracking as well.At the same time,it reaps the benefit of low computation complexity and high accuracy of expert joint decision.2)Hyperspectral image covers a wide range of geography scope containing a variety of materials and the corresponding image scene is complex.Conventional detectors need some rigorous assumptions on the spectrum distribution of background,which is not fully reasonable for the real collected hyperspectral data and cannot describe the complex scene accurately.Naive assumptions in this case are unable to model the data and may limit their generative ability for a new hyperspectral data.In order to overcome these limitations existing in traditional methods,this paper proposes a novel scheme based on manifold learning and graph theory.Without any assumptions on the distribution of background statistics,our method is more adaptive to different kinds of real-world hyperspectral images.It can select pixels via constructing a vertex-and edge-weighted graph,which can strengthen the distinctiveness of anomalies.