Research On Image Segmentation Based On Spatial And Similarity Feature

Image segmentation is used to divide the image into a plurality of regions having the same or similar characteristics but does not overlap each other, and its purpose is to extract the regions of interest. Image segmentation, which is the key step from image understanding to image analysis, is significant in image analysis.This technology was widely used in medical image processing, pattern recognition and computer vision technology and so on. However, with the continuous upgrading of electronic imaging equipment, the pixel size of the image becomes larger and larger, and the type is also increasingly tend to be more and more diverse, because of its unknown complexity and inherent particularity, the quality and speed of image segmentation is facing new challenges, causing researchers’ a lot of attentions. People has done a lot of work in the field of image segmentation and made many achievements, but in order to meet the higher demands of modern society for the image, we are still in the long way of doing researches to explore new related theories and methods in the image segmentation.This paper mainly studies the SAR image segmentation and natural image segmentation,the main contents include the following three aspects:1. A SAR image segmentation method based on super-pixel spatial correlation characteristics has been proposed. The algorithm uses neighboring and similar super-pixels of each super pixel to generate spatial position weighting factors and similar characteristics weighting factor, and apply them to the traditional FCM clustering algorithm so as to improve defects that segmentation effect is not ideal that resulted from did not consideration spatial location information in traditional FCM clustering algorithm. Ideal segmentation results had been obtained since the algorithm is applied to SAR image segmentation and natural image segmentation.2. A Spectral clustering image segmentation method based on global similarity has been proposed. The algorithm is committed to building a similarity matrix that express the image information sufficiently. In similarity matrix, the similarity was constructed between not only pixels and pixels, but also super pixels and super pixels, and the similarity between super pixel and pixels has been introduced, and uses a Gaussian model to express the similarity between them. After processing, the similarity matrix was applied to the existing spectral clustering model, and has improved the treatment effect of the image.3. A spectral clustering segmentation algorithm based on super-pixel encoding has been proposed. The algorithm construct a multilayer undirected weighted graph firstly, and then makes the establishment of the similarity matrix not only takes into account the relationship between the intra and inter layer, but also reduce the errors of those regions that may contain many objects. Use of super-pixel image encoding to express a similar relationship between the super-pixels and super-pixels. Such similarity matrix is significant when the split edge is not robust enough so as to obtain a ideal segmentation.