Study On The Hyperspectral Image Segmentation

In this thesis,the basic concept of hyperspectral image is illustrated,the characteristics of hyperspectral image ,which are stronger spectrum dependence of hyperspcetral and hybrid image-spectrum ,are studied and vertified;then the difficulties of hyperpectral image processing is analyzed; at last current technique of dimensionality reduction technique are summarized. Through summarizing current techniques of hyperspectral image segmentation ,a new algorithm of image segmentation based on hyperspectral curve characteristic is proposed: firstly, describes the spectrum curve by a certain amount of spectrum curve characteristics,and then achieves the purpose of pixel identification by the similarity of spectrum curve characteristics,at last the image segmentation is achieved. The main method of pixel detection describes as below: firstly extracting a series of characteristics of spectrum curve, and then executing the segmentation of using high-dimensional vector directly and transforming high-dimensional vector into scalar based on these characteristics.The main characteristics is extracted as follows: the number of wave crest which is above mean, the number of wave valley which is above mean, the number of wave crest which is below mean, the number of wave valley which is below mean, the number of uptrend bands, the number of downtrend bands, the band number of crest, and the band number of valley.The characteristics of spectrum curve can not only describe the spectral curve effectively and reflect the differences between pixels, but also can significantly reducethe computation and improve the efficiency of follow-up to deal with.The direct segmentation algorithm based on ISODATA achieve good result. Comparing to the segmentation result of original image, the segmental result based on extracted character image identified the superiority on segmenting texture information.The indirect segmentation algorithm which are the segmentation based on neighborhood characteristics angle and the segmentation based on character weighted have been experimented. Although the experiment didn't obtain good effect, it proved that the segmentation effect based on original characters is better than the segmentation effect based on transformed characters.