Research On Hyperspectral Fast Anomaly Detection Based On Progressive Line Processing

With the rapid development of imaging spectroscopy,hyperspectral imagery(HSI)with the property of image and spectrum has been gained wide attention in the field of agricultural monitoring,ore mining,military reconnaissance and so on.Compared to the traditional Two-dimensional image,HSI contains both the spatial information and spectral information of the objects so that it has the potential to discover the subtle differences of ground materials that cannot be visually inspected in a multispectral image.It lays a good foundation for the target detection.Hyperspectral target detection technology as a hotspot in the field of hyperspectral image processing can be roughly divided into two categories: supervised target detection and unsupervised target detection(anomaly detection)based on the availability of priori target information.In the application,the acquisition of prior information is very limited so that the anomaly detection which the priori spectral knowledge is unnecessary is of more value for academic research.However,with the spatial and spectral resolution of HSI continues to increase,coupled with the growing quantity of information is the huge pressure of data storage,transmission and further processing.Furthermore,it is particularly significant,as some moving objects are highly desirable to be detected on a timely basis.Therefore,it is urgent to improve the timeliness of anomaly detection while ensuring the accuracy of the detection.This paper mainly studies the real-time anomaly detection algorithm based on the imaging mechanism of push broom scanner.The main research contents in this paper are as follows:First,LRXD(Local Reed-Xiaoli detector)needs to repeatedly calculate local mean and covariance matrix while detecting the abnormal points,this results in lower detection efficiency.To solve the problem,a fast anomaly detection algorithm based on the semi-window is proposed.The concept of Kalman filtering and Woodbury matrix is imported to recursively update the complex matrices so as to improve the efficiency of the algorithm.In this case,LRXD algorithm can be implemented quickly with the sliding of the semi-window.Secondly,aforementioned BLRXD algorithm is not true real-time processing algorithm but actually fast algorithm which can process data without much time delay.For the problem,a real-time anomaly detection method is proposed based on the framework of progressive line processing.In order to keep the causality of real time,the local sliding causal window is introduced to detect anomaly targets.In terms of the high computational complexity caused by the inversion of matrix,the Woodbury's lemma is employed to update the status information of current data through iterating data status information at the inversion of large matrix,thereby improving the processing efficiency.Finally,real-time RXD algorithm resolves the problem of causality and timeliness,but it cannot avoid undesirable detection performances.Accordingly,an approach to the progressive line processing of KRXD(PLP-KRXD)that can perform KRXD line by line is presented.Parallel causal sliding windows are defined to ensure the causality of PLP-KRXD.Then,with the employment of the Woodbury matrix identity and the matrix inversion lemma,PLP-KRXD has the capacity to recursively update the kernel matrices,thereby avoiding a great many repetitive calculations of complex matrices,and greatly reducing the algorithm's complexity.