| Image mosaicing technique is to process several input images which have overlapswith each other, then registrate and transform them into one refrence image, and finallyresample and stitch them to form a wider perspective image, which still include theinformation of each image. However, when the images are captured from theconsiderable views, there exists parallax among these images. Especially, if there existsocclusion and disocclusion among them, the ghost will appear in the mosaicing imagewith the traditional image mosaicing methods.In the traditional image mosaicing technology, in order to reduce the ghost, aparallax-free camera configuration is necessary, in this camera configuration to reduceparallax the camera can not be moved, or dense sampling is used. However, thisconfiguration is often unrealistic, and can not be applied to all occasions. For example,hand-held cameras are always being moved. Although the mosaicing quality of theparallax-free images achieved by the traditional method is very good, when the parallaxin images cannot be ignored, the mosaicing quality will be reduced dramatically.In this paper, a noval image mosaicing technique is presented based on depthimage. The depth image is obtained from the stereo information generated frommulti-view images.The main contents of this thesis are as follows:(1) We have investigated the image segmentation based on graph cuts, mainlyfocused on energy minimization theory and-expansion algorithm. Then, we appliedthe graph cuts method to stereo matching, and gave out the definition of energyfunction.(2) In order to reduce the search time and improve the registration precision of thestereo matching, epipolar rectifcation method is applied to transform the matchingpoints of two images into the horizontal straight line. In order to adapt to the groundtruth images, a quasi-Euclidean uncalibrated epipolar rectification method is used,which directly uses matching points to realize image rectification.(3) A novel image mosaicing method based on depth image is put forward in this paper. First, the depth images of the left and right views are obtained based on stereomatching method, then a wider perspective image is achieved based on the principle offoreground overlay background.Simulation and ground truth images experimental results prove that this method isfeasible, and it can eliminate ghost and deformation caused by traditional mosaicingmethod. |
