Optical Remote Sensing Image Semantic Segmentation And Target Extraction Algorithm

With the development of aerospace technology,remote sensing image has become an important source of national defense security,geographic survey and aerospace technology in every country.Image semantic segmentation is an important direction of optical remote sensing image research.In recent years,more and more scholars have carried out a lot of research on the semantic segmentation of optical remote sensing image,and achieved remarkable results.This paper takes optical remote sensing images as the research object,realizes the semantic segmentation of images through the traditional superpixel clustering method,and realizes the semantic target extraction after detection and clipping for specific target objects.Finally,corresponding improvements are proposed for the superpixel segmentation,semantic segmentation and semantic target extraction respectively.The specific work is as follows:1)In this paper,a superpixel generation method of remote sensing image based on edge tracking is proposed.Aiming at the problems of large amount of calculation and long operation time in remote sensing images with large amount of data processed by SLIC(Simple Linear Iterative Clustering),an edge-tracking method for generating superpixel of ET-SLIC(Edge Tracking-SLIC)was proposed.By finding the boundary points of each superpixel block from four different directions,a closed interval was formed to quickly generate superpixel.Experimental results show that the proposed algorithm can reduce the computational cost significantly while ensuring the quality of superpixels,and the running time of remote sensing images with large amount of data is shortened by half compared with SLIC.2)This paper proposes a semantic segmentation algorithm for remote sensing images based on superpixel spectral clustering.Remote sensing images are processed into superpixel blocks,and we use the spectral clustering algorithm to realize the image segmentation.Due to the low segmentation accuracy of spectral clustering algorithm,we propose a spectral clustering optimization algorithm.By increasing the spectral clustering K value to ensure that the initial clustering accuracy,and then the texture information is used to replace the spatial position information in the similarity matrix to get the improved similarity matrix.Finally this paper put forward the global optimal merge algorithm to realize the remote sensing image segmentation.Coupled with the SVM(Support Vector Machine)classifier tag of each area,to achieve the purpose of semantic segmentation.The semantic segmentation of remote sensing image is realized by the method of segmentation first and then annotation,and the results were compared with eCognition software.The results showed that the F1 score and Kappa coefficient of this algorithm reached 87.46% and 0.8314 respectively,both of which were greatly improved compared with eCognition software.3)In this paper,an algorithm based on the ARPN(Attentional Region Proposal Network)and GrabCut is proposed for semantic target extraction of remote sensing images.In the semantic segmentation of remote sensing images,people usually focus on the semantic segmentation of specific objects rather than the semantic segmentation results of the whole image.Therefore,ARPN algorithm is adopted in this paper to achieve the detection of specific targets,so as to obtain the category information and position information of the target object.Then,the position information is taken as the input of GrabCut algorithm for clipping.Finally,the specific semantic target extraction results of remote sensing images are obtained by adding category information to the output.An improved GrabCut algorithm combining ET-SLIC,significance detection and morphological method was proposed due to the high computational complexity and low segmentation accuracy of GrabCut algorithm.The experiment has proved that the algorithm of F1 score reached 88.36%,Kappa coefficient reached 0.8034,showing good extraction performance.