Image Matching For Depth Recovery And Shape Recognition

Image matching is an attracted area in the field of computer vision and image understanding, which has been widely used in multi-sensor image fusion, environmental monitoring, target recognition, and 3D reconstruction. Although image matching obeys a universal framework, each class of problem is designed by its specific application. This paper focuses on the technique of stereo matching for depth recovery and model-based matching for shape recognition.In stereo matching for depth recovery, both local and global stereo matching algorithms are mainly studied after reviewing the issues of basic theory, difficulties, occlusion detection, evaluation and methods for stereo matching.In the research of local stereo matching, two new methods are proposed for riding out the demerit of area-based methods used widely. Firstly, an improved approach that only uses gray information to decide the window size is proposed. Since the selection of window size doesn't need both local gray and disparity variation, it can easily make use of integral image technique to speed up computation and the calculation is independent of window size. Secondly, an algorithm based on disparity point for dealing with large occlusion is developed, which extends the first method. It can resolve the problem of slanted surface and occlusion, since some parameters are used to describe the slope of line and occlusions distinctly in the parameterization of a disparity point.In the research of global stereo matching, two novel approaches are proposed in order to solve the problem that segment-based method is readily affected by the errors of initial color segmentation. Firstly, the pixel-set based stereo matching algorithms by using tree re-weighted belief propagation is offered. The initial color segments is segmented again according to the histogram of initial ground control disparity points. Therefore it can suffer less from the error that disparity boundaries appear inside the regions with similar color. Secondly, the symmetric segment-based algorithm is presented, which detects occlusion in segment level by extending uniqueness constraint from pixel domain to segment domain. The experiment results on the standard data sets show that our overall performance ranks the third out of about 24 algorithms.In the research of model-based matching for shape recognition, two algorithms for application are proposed. Firstly, the approach to recognize planar polygon by local invariant features based on cross-ratio is proposed. It has little computational complexity and stable performance, especially for the local incomplete planar polygon, even losing a few vertices. Secondly, the method based on gradient template image to recognize number imaging in complicated circumstance is offered. It can remove the effects by the illumination change, dust, oil paint breaking off, machine shaking, objects motion, low quality number and occlusion in the product line environment of a steel mill. The final recognition rate of above 80% is achieved in an experimental product line.