Research On Key Technologies Of Natural Scene Image Segmentation

Image segmentation is one of the most challenging technology in the fields of imageprocessing and computer vision, it has been studied for many decades. Early imagesegmentation techniques are designed for grayscale images and structured scene images, theseimages with simple form and stable feature, are easy to be segmented. However, in recentyears, with the popularity the Internet and digital products, segmenting the natural sceneimage and extracting the target is a hot and difficult point in field of image segmentation.Through analyzing the color characteristics of natural scene images, this paper doprofound research on the existing segmentation methods, and put forward our own algorithms,the main contents as follows:First, we analyze the color space and calibrate the object of the image in the HSI colorspace. And then we calculate the threshold using Otsu algorithm to and binaries the image.Finally, we use morphological operations to denoise the image and fill the hollow. Weproposed a thresholding segmentation method based on color calibration and Otsu and get agood segmentation result.Second, since Otsu algorithm does not take the image spatial neighbor information intoconsideration, we combine the Markov random field with Otsu algorithm to integrate graylevel information and spatial correlation information for the pixels. In this paper, Otsuthresholding algorithm based on Markov Random Field is proposed. In this algorithm, theneighborhood rejectability function is imported to Otsu algorithm and an threshold selectionfunction is improved. The experiment results verify that applying our algorithm to road imagesegmentation can achieve good effects.Finally, through deeply studying mean shift and the minimum spanning tree, we presenta combined mean-shift with the minimum spanning tree image segmentation algorithm(MS-MST), in order to improve the operating efficiency of the classic mean shift. Thealgorithm first select a smaller spatial bandwidth, apply mean shift to over-segment image atthe faster speed. Then, we regard the over-segmentation region as the basic unit of subsequentprocedure that structuring a weighted region adjacency graph, and use minimum spanningtree algorithm to merge over-segment image. The experiment results verify that this algorithm,on the premise of ensuring the quality of image segmentation, substantially increases thespeed of the classic mean shift segmentation algorithm.