High-Resolution Remote Sensing Image Sparse Decomposition And Its Linear Texture Information Extraction

Target recognition and information extraction are the key technology of high-resolution remote sensing applications.With the increase of spatial resolution,the spectral characteristics of surface features in high-resolution remote sensing image are not only become more obvious,but the spatial structure information such as geometry structure and texture are also very prominent,which offers the possibility of high-resolution remote sensing image to achieve more detailed and more accurate information recognition.However,the pixel-based image analysis method does not meet the requirements of information extraction of high-resolution remote sensing image,as a bottleneck to large-scale applications of high-resolution remote sensing image.In order to achieve the object of more precision and more efficiency identification using abundant characteristic information of high-resolution remote sensing image,information extraction methods of high resolution remote sensing images has become an inevitable requirement from the change of the pixel-based methods to the feature-based methods.Then,how to realize the change from pixel-based methods to feature-based methods has become the frontier scientific problem of remote sensing science.Biological visual identity objects are born with the ability of high efficiency and real-time performance;the results of sparse decomposition model is consistent to the law of biological visual perception,which provides a new research idea to information extraction in high resolution remote sensing image.Funded by Chinese National Programs for High Technology Research and Development(Grant No.2008AA12Z106)titled "Research on the novel technique of image segmentation based on high-resolution remotely sensed imagery" and the National Natural Science Foundation of China(Grant No.40801166)titled "Research on the algorithm of multi-scale segmentation based on frequency feature from high-resolution remote sensing image",this thesis discussed the information extraction method used in high-resolution remote sensing image.With sparse decomposition theory and the Quickbird image of study area,this thesis explored a model of image decomposition of high-resolution remote sensing image based on sparse decomposition.Information load of decomposition image was also analyzed.In addition,an academic idea of information extraction based on feature-based methods was studied.The main aspects include the following.(1)On the basis of over-complete dictionary,the complex structure of image and the optimal performance of sparse constraints on image features are better characterized.This thesis put forward an image decomposition model,which could effectively separate high-resolution image into the high-frequency component and the low-frequency component.Through sparse decomposition theory,I found over-complete dictionary had better approximation performance of image structural components than a single orthogonal basis,and the optimal solution of sparsely constraints was consistent to the human visual cortex to stimulate the expression of the complex objects,which directly inspired to design an image decomposition model,showing the ability to effectively decompose high-resolution remote sensing image into high-frequency component and low frequency component,and would help features identification and information extraction from high-resolution remote sensing image.(2)Information load of reconstructed images,combining high frequency component with the low frequency component,provided a guidance on recognition objects and extraction information.Feature of decomposition images in study area was analyzed by the peak signal to noise ratio and root mean square error;the results showed that reconstruction images of the high-frequency component and the low-frequency component contained enough information to object recognition.This thesis also studied features of luminance response from the high-frequency component and the low-frequency component.Results exhibited that reconstruction image of the high-frequency component would help to distinguish and extract structural information from high-resolution remote sensing image,and reconstruction image of the low-frequency component would help to distinguish and extract color information.These findings would provide a guidance to achieve target identification and information extraction based on the decomposition images.In order to verify this conclusion,based on the idea of feature-based on recognition and extraction,a technological exploration was carried out using reconstructed image of the high-frequency component.(3)Through characteristics of reconstruction image of the high-frequency component and the idea of the best sparse approximation,a denoising method of a linear feature enhancement for high frequency component was proposed.Though reconstruction image of the high-frequency component could effectively characterize structure information of original image,but the noise of original image still remained in reconstruction image of the high-frequency component due to a variety of factors in imaging process.Noise suppression was another problem to feature extraction and target recognition based on reconstruction image of the high-frequency component.The proposed denoising method was Curvelet,moreover,multi-level threshold settings were also applied according to the noise distribution at different scales of high-resolution remote sensing image.These methods effectively eliminated the noise of the high frequency component,and made the linear texture feature much more outstanding which is conducive to its high-precision extraction.Finally,multi-level thresholds were set using multi-scale characteristics of objects in study area image.Based on these multi-level thresholds,linear texture information was extracted from reconstruction image of the high-frequency component after denosing.Then the extraction of information was superposed to reconstruction image of the low-frequency component,the geological significance of the object was thus obtained,which verified the validity of research ideas in this thesis.The accuracy rate of accuracy evaluation was higher than 87%,which could meet the needs of more accurate object recognition and information extraction in the field of high-resolution remote sensing image.As a summary,this thesis studied image decomposition method of high-resolution remote sensing image and method of linear texture extraction.This thesis put forward an image decomposition model for high-resolution remote sensing image,and proved a method consisting of information extraction from remote sensing image,which included information decomposition,linear feature enhancement and threshold settings on each scale and was successful in linear texture extraction.The study was new in information decomposition of high-resolution remote sensing image and linear objects extraction.Further studies as follows:other objects recognition and extraction from high-resolution remote sensing image based on sparse decomposition;a series of problems of the transition from object graph to remote sensing mapping;object recognition and information extraction using low-frequency component.And applications of this method to the geological structural information extraction and other similar studies were also required.