Research Of Image Segmentation Algorithm Based On Region Merging

As the rapid advancement of modern imaging equipment and technologies,including various optical imaging systems and microwave imaging systems,the automatic interpretation and recognition of digital image data become a support for the application in advanced imaging systems in practical environments.Image segmentation technology is a process of partitioning an image into several non-overlapping regions with similar properties,which is one of the important basic research issues in computer vision and image processing field.The most popular techniques can be divided into four categories: clustering based segmentation algorithms,region merging based segmentation algorithms,graph theory based segmentation algorithms and classification based segmentation algorithms.The main idea of clustering based segmentation algorithms classify the dataset into a multiple number subsets according to some defined distance measure or other means of measure;In case of region merging based segmentation approach,the whole image is divided into a set of disjoint region,and then using a merge strategy to merge the adjacent regions based on similarity measures;Graph theory based segmentation algorithm uses undirected graph to represent image data,and convert the segmentation issue to an undirected graph partition problem;The classification based segmentation algorithm uses a trained classifier to label image pixels.This paper addresses the two problems of similarity measurement criteria and region merging strategy between adjacent regions in the region merging based segmentation approach,and segmentation algorithms for images with color and depth information(called RGB-D image)and synthetic aperture radar(SAR)images are proposed respectively.The main contents of this dissertation are as follows:(1)Aiming at the image segmentation of RGB-D images of indoor scenes,an unsupervised image segmentation method using color-direction based region merging is proposed.Firstly,in order to achieve the regional similarity measure combining color and plane direction,the method utilizes the data-driven which is calculated by using a multi-directional bisemi-circle Gaussian function to construct the adaptive weight;Then,a region merging threshold function is designed to guide the region merging process based on the difference in plane direction between regions.The experiments show that the method improves the edge quality and the region quality.(2)To address the problem of adjacent areas with nearly similar textures and directions cannot get the ideal segment results.We propose A novel image segmentation method for RGB-D images from the characteristics of color and geometric information in indoor scene.It begins by applying a data-driven to combine color and depth information,and using the plane normal difference at the edge positions between the mainly planes as a penalty term.Next,a region merging cost function is proposed by utilizing the plane co-planarity on the region merging threshold,which generates a set of regions.Finally,it applies a statistical region merging method on resulting regions to obtain the final segmentation.The experimental results show that this method has obvious advantages in edge quality and region segmentation quality than the state of the art unsupervised techniques.(3)To overcome the problem of heavily irregular for the shapes of boundaries caused by strongly speckle noise and texture information,we propose a hierarchical region merging method with geometrical edge penalty to segment a SAR image into un-overlapped scene areas.It applies watershed transform on the thresholded edge strength map(ESM)extracted by using multiscale Bhattacharyya distant of the SAR image to obtain an initial segmentation results.In order to guarantee the final segmented regions are delimited by smooth boundaries,a novel region merging criterion is proposed in terms of statistical measure based on the nonparametric Kuiper's distance between two adjacent regions,and geometric penalty term using geometrical information of common boundaries between the adjacent regions.The hierarchical segmentation results are yielded through changing of a parameter in the geometric penalty term to gradually increase the strength of geometric penalty term.Experimental results on real SAR images show that the proposed method is more effective on the segmentation of SAR images in complicate scenes than several recent state-of-the-art methods.