Application Of Improved Anisotropic Diffusion Filtering In High Resolution Remote Rensing Image Segmentation

With the development of satellite technology,the spatial resolution of satellites is getting higher and higher.However,the automatic processing methods of remote sensing images are more and more unable to meet the needs.Because of the more abundant information of the high resolution image,the background noise in the image has become more complex,and the traditional and widely used medium and low resolution remote sensing image processing methods have been unable to meet the needs very well.Among them,fast and accurate identification of high resolution remote sensing images is one of the most urgent problems to solve.The segmentation technique is the first step of high resolution remote sensing image recognition.Firstly,the remote sensing images with complex background noise and texture are segmented,and then the segmented regions are analyzed and processed to realize the recognition of high resolution remote sensing images indirectly.After studying a lot of image segmentation algorithms,we use watershed transform algorithm as the segmentation method.However,because of the overcomplexity of the background noise and the texture information of the ground objects,the segmentation results of the traditional watershed transform algorithm are often overdivided,which requires the improvement of the traditional watershed transform algorithm.In order to solve the problem of over segmentation of images,anisotropic diffusion filtering algorithm is introduced and improved.The improved algorithm not only removes the background noise,but also preserves the edge information and highlights the details of the edge position.First,the gradient mode in anisotropic diffusion filtering is improved.The non local mean algorithm(NLM algorithm)is used to replace the original gradient mode after the smooth gradient model,which enhances the noise resistance of the filter,especially the strong noise and the edge information.Second,we improved the method of obtaining diffusion coefficient in the original algorithm.An adaptive method is proposed to achieve the automatic acquisition of diffusion coefficient,which reduces the influence of human factors.Third,the multiscale morphological gradient was used.The traditional single scale morphological gradient can not extract gradient information accurately.In this paper,a multi-scale morphological gradient is adopted.This not only avoids the mutual influence between the edges of the large scale morphological gradient,but also shows that the maximum value of the gradient is not consistent with the real edge information.It also avoids the good denoising effect of the small scale morphological gradient,thus enhancing the image’s boundary information and the anti noise of the gradient image.Fourth,the extended minimum transformation is used to eliminate the local "Valley" and "burr" in the gradient image,and to retain the significant "mountain peaks" in the gradient image,and to further avoid the occurrence of over segmentation.Finally,using the method proposed in this paper and the mainstre am commercial segmentation software eCogniton on the market,the aeria ed.The results of six indexes are evaluated:over segmentation index(OR),under segmentation index(UR),differential distance(Dij),potent ial segmentation error(PSE)and polygon quantity ratio(NSR).The res ults show that the segmentation accuracy of the proposed method is hi gher than eCogniton.