Hyperspectral Image Description Based On Local Feature In Spatial-Spectral Domain And Its Application In Target Detection

Image description in hyperspectral images is an important research direction in the fields of information science,pattern recognition and machine vision.With the development of imaging equipment technology,high resolution hyperspectral images can provide more detailed features,which is conducive to accurate object detection.Combining with the characteristics of high resolution hyperspectral images,this paper studies the key technologies of image description based on local description in spatial-spectral domain and its application in target detection.The main work includes the following four aspects:A spatial-spectral SURF feature detection and description algorithm for hyperspectral images based on geometric algebra(GA-SSSURF)is proposed.Feature point detection is the precondition of accurate classification and object detection in hyperspectral images.Starting from the characteristics of hyperspectral data,we use geometric algebra as a mathematical analysis tool to establish the spatial-spectral domain representation model of hyperspectral images.On this model,we propose a Hessian matrix generation method for the geometric algebraic space of hyperspectral images.On this basis,a spatial-spectral SURF feature detection and description algorithm based on geometric algebra is proposed,which enhances the uniqueness and affine invariance of local features and improves the accuracy of local description of hyperspectral images.A dual locality-constrained linear coding algorithm for 2D images(DLLC)and a spatial-spectral locality constrained linear coding algorithm for hyperspectral images(SSLLC)are proposed.Image local feature constrained coding is the link between local features and visual vocabulary vectors.It is the core step of image classification and object detection based on local description.In this paper,a dual locality-constrained linear coding algorithm for 2D images is introduced by utilizing the correlation of local feature points,which reduces the ambiguity of mapping between feature points and dictionary words,and enhances the expressive ability of 2D images.On this basis,a spatial-spectral locality constrained linear coding algorithm for hyperspectral images is proposed.While using visual words to constrain the local features of hyperspectral images,the algorithm introduces the local correlation information of adjacent bands of hyperspectral images and maps them as feature points to dictionary words.The algorithm combines the spatial-spectral information of the neighborhood of interest points,reduces the problem of synonymity and polysemy in the process of feature coding,and enhances the ability of describing hyperspectral images.A pyramid matching model in spatial-spectral domain and a pyramid matching model in gradient direction are proposed for hyperspectral images.Optimizing feature combination can improve the expressive power of feature vectors to images,which is an important subject in the field of object detection in hyperspectral images.At present,the classical spatial pyramid matching model(SPM)is commonly used to optimize the combination of local features in spatial-spectral domain.SPM can segment hyperspectral images in spatial domain,combine features in subdivision space,and encode spatial information into feature vectors.However,this simple segmentation method ignores the important information of local features such as spectral domain distribution and spectral domain correlation,which leads to the low robustness of traditional feature combination model.For general hyperspectral images and hyperspectral images of objects with similar texture or shape,we use spatial-spectral correlation and gradient information of hyperspectral images to design two different local feature combination algorithms,namely,spatial-spectral pyramid matching model(SSPM)and gradient direction pyramid matching model(GPM),to improve the accuracy of image representation.An object detection algorithm based on multi-scale region matching for high resolution hyperspectral images is proposed.Traditional object detection algorithms for hyperspectral images are mainly for low-resolution satellite remote sensing hyperspectral images,which are difficult to process high-resolution hyperspectral images effectively.Based on the region-related information of high-resolution hyperspectral images,the sliding window is used to traverse the hyperspectral images using the spatial-spectral local features of the hyperspectral images proposed in the previous chapters,and the spatial-spectral stereo matching model is used to search the suspected object region in the sliding window.In the suspected object area,a smaller sliding window is used to detect the object,and the object is accurately located by fusing the similarity between the multi-scale window area and the target,to improve the performance of object detection and enhance the accuracy and robustness of object detection.