Semi-supervised Semantic Segmentation Method For High-Resolution Remote Sensing Image Based On Feature Field

With the rapid development of sensor technology and aerospace platform technology,the capability of remote sensing technique in earth observation has also been greatly improved,followed by a large number of remote sensing data with more and more information.High-resolution remote sensing technique observes the earth in a more refined way,and the obtained remote sensing data also contains more accurate and rich features,which can be expressed as texture color,geometric structure and spatial relationship.These increasingly refined features of the terrain make it difficult to transform the information in such rich remote sensing images into geographic information that can be edited and processed for analysis.In the face of the increasing amount of high-resolution remote sensing image data,how to effectively perform target recognition,data mining and information extraction is particularly important.A semi-supervised semantic segmentation method for residential building object in high-resolution remote sensing images based on Scale-invariant Feature Transform theory(SIFT)feature field is proposed in this paper,which can make the object foreground area and the background area marked automatically in the high-resolution remote sensing image,and then the marked foreground and background area can be used to perform the evolution cut of the target object by the graph cut algorithm.The main contents of this paper include:(1)A introduction of the necessity of two-dimensional image object classification and segmentation,organization structure,technical flow,development status and application prospects of this paper;(2)Realizing the cutting of the target object sample area in the high-resolution remote sensing image,and performing feature point matching through the feature point information of the sample area and the original remote sensing image,mapping the feature points of the sample area to the original image,and mapping the original remote sensing image and the sample area One-to-many matching;(3)Using the modulo of the feature vectors and matching results of the SIFT feature points,do the corresponding processing of neighborhood grayscale color histograms for matched feature point pairs,and the precision(P),accuracy(A),recall rate(R)and F value(PARF)are used to filter the feature points by iteratively update the quantitative results of the histogram comparison,then select the histogram result value when the F value is the maximum as the threshold value to perform preliminary filter on the matched feature point pairs;(4)The statistical filtering in the 3D point cloud filtering method is used in the feature point filter process of this paper,and obtain the feature points of the final target object region by remove the discrete points in the feature points after the preliminary filter;(5)The feature points of the target object area are marked as the foreground points,and the remaining feature points are marked as background points,and the target object is subjected to evolutionary segmentation by using the graph cut algorithm.(6)Quantitative evaluation of target objects segmentation results in high-resolution remote sensing images using PARF index values.In this paper,the semi-supervised semantic segmentation of target objects in highresolution remote sensing images is realized by SIFT feature field and graph cut algorithm.The high-resolution remote sensing images of two urban residential quarters in Wuhan area are segmented and extracted.The experimental results show that the proposed method in thins paper has almost no gaps in the target segmentation method,and the evaluation values of the PARF index are above 0.7,and the recall rate(R)are above 0.9,which indicate that the method is effective and the target objects in the image can be segmented well according to the texture gradient,spatial neighborhood and scale and direction invariance of high-resolution remote sensing images.