Research On Methods For Target Detection And Recognition In Hyperspectral Images By Fusing Multi-dimensional Features

With the rapid development of military surveillance and reconnaissance,the demand for detection and recognition performance of imaging system is increasing.Traditional single or multi spectral imaging is difficult to achieve high-performance detection and recognition of targets.Hyperspectral imaging technology has become an important development direction of space optical load in the future because of its advantages of obtaining multiple continuous spectral images with high spectral resolution at the same time.However,according to the principle and the existing research of spectral imaging technology,hyperspectral image has the limitations of large amount of data,high redundancy of information and low spatial resolution,which leads to the problems of large amount of calculation,low efficiency of data interpretation,and overlapping of target and background spectra.Lead to the difficulty of target detection and recognition further increases.Can not meet system’s real-time interpretation for the target.Therefore,it is necessary to study efficient target detection and recognition strategies in hyperspectral images.In order to solve the above problems,this paper analyzes the factors that affect the detectability and recognizability of the target,and studies the methods of spectrum band optimization,target detection and recognition.The main research works are as follows:(1)The influencing factors of target detectability and recognizability are analyzed.From the perspective of the whole imaging link,the effects of spectral aliasing,radiation attenuation,image quality blur,noise interference equivalent generated in the imaging process of the target/background radiation,atmospheric transmission and detection system on the detectability and recognizability of the target are analyzed.According to the different effects of each link,some suggestions are given for the optimization of spectrum band,the design of target detection and recognition methods,which provide theoretical support for the follow-up research.(2)A method of spectrum band optimization for detection and recognition task is proposed.Aiming at improving the detectability of the target,and according to the different characteristics of background clutter and system noise in different spectrum band.Based on the signal separability maximization,the measurement model of background clutter and system noise is established.Proposed the optimization method of target detection spectrum band.In addition,according to the different characteristics of target recognizable performance in each spectrum band,established category separability measurement model and abundance inversion reliability measurement model.Proposed a method of optimization of target recognition spectrum band.The simulation experiment is carried out,and the conclusion of optimization of target detection and recognition spectrum band is given.(3)A method of detecting weak and small target in complex cloud background is proposed.In view of the complex and changeable cloud background and the multi-scale and multi working condition characteristics of the target.Taking the high detection probability and low false alarm probability of the algorithm as the design target.Based on the mathematical morphology theory and combined with the idea of multi-scale filtering fusion,an all-round multi-scale morphological adaptive filter is constructed,which suppresses the background effectively.At the same time,the research on adaptive threshold segmentation target detection method is carried out to eliminate the false alarm while extracting the target.Aiming at the problem that the pipeline filtering algorithm is difficult to deal with the background clutter points,multiple suspected target points and different speed target points in the window efficiently,a multi pipeline filtering method based on the target scale and motion estimation is proposed to confirm the target.The performance of the detection algorithm is verified by the simulation image.(4)A method of mixed pixel target recognition in complex cloud background is proposed.According to the principle of category separability maximization,the feature extraction method of target derivative spectral curve is proposed.Introduced the idea of spectrum curve coding,and the data set of target recognition is constructed.Based on the framework of linear mixing theory,and combined with the idea of resolution conversion and feature enhancement,the aerial motion point target unmixing model is established.The aerial motion point target atmospheric correction model is proposed based on the radiation transmission theory.Aiming at the problems of unbalanced decision tree structure design and classification hyperplane tilt suppression,studied a reasonable and efficient model construction method,and proposed a target classification method based on decision tree multi classification support vector machine.Finally,the performance of this method is verified by simulation experiments.