A Stereo Matching Algorithm Based On Feedback

Stereo vision is one of the main research questions in the field ofcomputer vision. In other words, the associated depth information will beobtained by searching for the corresponding relation of the pixels betweentwo images in the same scene. Stereo matching is the most important anddifficult part of the stereo vision. The depth map generated by stereomatching not only can be provided to the depth image basedrendering(DIBR), but also saves the bandwidth of the multi-view video.Thus, the stereo matching has a significant impact on making thetransmission of the image or the video more efficient. However, due to thestereo matching itself is a ill-posed problem, the matching algorithm doesnot have a unified standard, which becomes a difficult problem of research.We have made some improvement by analyzing the main matchingalgorithms existed, and a novel algorithm of stereo matching based onfeedback is proposed, in which the feedback information is taken into account.(1) A method which adds feedback information into data costaccording to belief propagation (BP) is presented to optimize the disparityin chapter1. The main procedures of this approach is: Firstly, the disparitymap is created through the images of left and right views by stereomatching algorithm. Then we take advantage of the left view image andrelated disparity values to render the right view image which is used tocompare with the real right view image. Finally, the comparative results areprovided for the energy function of BP algorithm in order to regenerate themore accurate disparity map.(2) Another method which is used to obtain more precise disparitymap by adding feedback information into energy function of the graph cutalgorithm is also proposed in chapter2. This approach proceeds as follow:The depth map is first generated via the images of left and right viewsaccording to the graph cut algorithm. Next, the left viewpoint image andassociated depth information are used for generating the virtual image based on depth image rendering. After that, the rendered image is restoredwhich is further compared with the original rendered image with holes.Finally, feed the comparative results back to the energy function andregenerate the depth map which can make the rendered virtual view imagemore accurate. The experimental results show that the proposed methodcan obtain satisfactory rendering images.