Research On Object-oriented Segmentation Of High Resolution Remote Sensing Image

With the development of remote sensing technology, the high resolution remotesensing image gradually go to the general public, which traditional image processingmethod based on pixel is not suitable for dealing with. So an object-oriented methodhas emerged. In object-oriented method, object is image’s indivisible minimalprocessing unit instead of pixel which can make full use of sundry characteristics ofthe object, improve the processing accuracy. As the key step in remote sensing imageprocessing, the quality of remote sensing image segmentation shows direct influenceon the analysis, understanding and application of image. In order to improve thesegmentation accuracy and play positive role on the following steps, object-orientedis applied to the segmentation. Originating from the object-oriented perspective, thispaper studies the remote sensing image segmentation techniques, which contain thefollowing work and innovation.According to the characteristics of high resolution remote sensing image, thispaper enhances remote sensing before segmenting. Firstly, the improved multi-scaleRetinex algorithm is used to enhance image’s brightness, and keeping image’ssaturation and hue in the same which can not only improve the contrast, but alsomaintain color information of images. Then this paper uses color histogramequalization method to further enhance image. Experiments demonstrate that thispaper enhances images on local and global with obvious effects of imageenhancement and great color keeping which is well prepared for next stage ofsegmenting.Because mean shift algorithm may produce less over-segmentation, which meetsthe large and detailed requirements, this paper uses mean shift algorithm to constructimage’s object by pre-segmenting. Through the process of iteratively calculate meanshift, mean shift algorithm congregate similar pixels into specific meaning of objects.According to the problem of high time complexity in iterative process, this paperadopts some strategies and multi-threading technology to improve iterative process,which can improve the efficiency of constructing object.With the object-oriented idea, similar objects which are treated as image’sindivisible minimal processing unit will be combined. Firstly, this paper calculates thegray average of pixels in the object, and combine objects whose scale are smaller than the threshold based on the minimum difference criterion of average. Then accordingto statistics every object’s color histogram, this paper takes Euclidean distance ofcolor histogram as criterion for further combining similar objects. Finally, White isused to mark the extracted contours and the final results of image segmentation can besaved after the entire object is represented by a pixel gray value in objects.The experiments demonstrate that our methods have good segmentation resultsand the segmentation results can accord with human visual effect.