Research On New Techniques Of Target Recognition In Hyperspectral Images

Hyperspectral image is a new type of remote sensing data, which has become recent research focus in this field. In applications of national defense, hyperspectral image provides sufficient spectral details for classification and detection of military targets. Hence, it's necessary to research effective algorithms of accurate target recognition using hyperspectral information.This thesis mainly researches hypersepctral target recognition algorithms, which aims to realize effective detection using signal processing and pattern recognition techniques, fully mining spectral and spatial information within the image, and cooperating hyperspectral image and optical image. Firstly, this thesis begins with spectral detection algorithms to research how to detect targets using spectral information. This thesis performs detailed study on current spectral detectors from the perspect of pattern recognition, which puts main attention to decision boundaries. Geometric explanations are then given, based on which a modified spectral detector is proposed. The modified spectral detector can greatly restrain background interference by reasonable decision boundaries. After that, spatial detection algorithms are researched, and two plans are designed based on binary image and gray image, which utilize invariant moments and similarity measures for matching respectively. Finally, based on spectral and spatial detection results, research are done to find how to cooperate spectral and spatial information for"refined target recognition", which is a new concept. Based on distribution of spectral and spatial detection results, a"spectral-spatial cooperative feature"is constructed, and a hyperspectral refined target recognition system is built. By this system, each pixel can be judged as"actual target","surface-covered target","superior target"or"background".Main contribution of this thesis lies in two innovations. First, a modified spectral detector is proposed, which can effectively restrain false alarms. Second, a hyperspectral refined target recognition system is built, which introduces the concept of"refined target recognition"and defines"spectral spatial cooperative feature". This system can effectively cooperate spectral and spatial information from different sources, and realize accurate target recognition.