Image Segmentation Based On Semi-Supervised Learning And Region Characteristics

Segmentation is one of the most difficult and important steps in digital image processing. Segmentation accuracy determines the eventual success or failure of computerized analysis procedure. Therefore, the research of image segmentation has been getting great attention. And a number of different algorithms have been proposed.Based on the application of medical image segmentation and SAR image segmentation, this paper researches the segmentation of abdominal organs from CT scans using 3D region growing algorithm, and the segmentation of SAR image based on two kinds of semi-supervised learning algorithm. The main contributions can be summarized as follows:(1) A new method of abdominal organ extraction based on twice 3D region growing algorithm is proposed. In order to inhibit the phenomenon of over-segmentation appeared in the traditional 3D region growing algorithm, the algorithm uses the Canny edges and the morphological edge of the image obtained from the first 3D region growing to restrain the second growth of the region. The experiments show the algorithm not only can effectively inhibit the phenomenon of over-segmentation, but also can significantly reduce the number of holes in the segmentation results. It can extract abdominal organs effectively.(2) A semi-supervised spectral clustering based on distance learning algorithm is proposed. The algorithm uses pairwise data to learn a distance metric, and uses the learned distance to build the similarity matrix. The constrained K-means clustering algorithm is used to cluster the feature vector obtained from spectral mapping. The algorithm makes full use of the prior information to enhance the stability of the algorithm. The experimental results on SAR images and texture show that the algorithm has a better performance than the traditional spectral clustering algorithm.(3) A SAR image segmentation method combining watershed and improved Laplacian Support Vector Machine (Laplacian SVM) algorithm is proposed. To solve the problem how to select the scaling parameter, the self-tuning graph is introduced into the Laplacian SVM algorithm. Additionally, because the graph and kernel function idea are used in the Laplacian SVM algorithm, there will be large amount of storage capacity and computing problems when its application to image segmentation problems. To speed up the algorithm, the proposed method firstly uses the watershed algorithm to do image segmentation, and then applies the improved Laplacian SVM algorithm to classify over-segmentation regions.