Research On Image Salient Region Extraction Algorithm Based On Unsupervised Matting Model

The extraction and recognition of salient region in nature image are a hot area of computer vision and image understanding. Image salient region detection which has the widespread application value, could be applied in the fields as target marking, content-based multimedia information retrieval. There are two kinds of algorithm for region detection and edge detection:bottom-up and top-down image based segmentation methods. Bottom-up based methods could detect the edge which comes from the dramatic change of the colors or gray. But this method may bring the problems of over-segmentation and fault-segmentation of the target area as the mechanism of visual attention has not been concerned. Top-down based method could detect the edge of the target area better but need much more guidance of prior information. And the results are far from satisfactory when the method is used in computing the complex natural scenes. How to get the information of position and edge of salient region in nature image is the topic of this dissertation.A new unsupervised matting model is designed for solving the above mentioned issues in this dissertation. With good robustness, it could detect the edge of the salient region in nature image on the basis of the different kinds of foreground image in different kinds of images. The following topics are fully studied in this dissertation.(1) Extracted feature point of SIFT from both of salient foreground region and non-salient background region, and provided the data for training the SVM classifier.(2) This dissertation utilized the matting algorithm to detect the edge of the salient region in nature image. Studied the characteristics of Trimap in the algorithm of learning based digital matting, and acquired the theory basis of Unsup-Trimap.(3) An Unsup-Trimap algorithm is designed based on the theory above. The design of this algorithm includes both of coarse localization and accurate positioning, and rationalization of Unsup-Trimap in the application of salient region extraction has been proved. The algorithm of detecting the edge of the salient region in nature image is designed based on the research above.(4) A new system of salient region extraction and recognition has been designed. This innovative system has better performance because of the new algorithm.