Stereo Matching Of High-Resolution Images With Similar Texture

Stereo matching is one of the most important research topics in computer vision, which at-tempts to find identical points from two 2D images with overlapping areas and record the corre-sponding information in a disparity map. Combined with camera parameters, we can recover a 3D model from the disparity map of the reference image. How to accurately calculate the disparity map of the reference image according to a stereo image pair is the basic task of stereo matching and the critical step of reconstructing a 3D scene based on stereo image pairs.With the development of computer vision technology, binocular stereo matching has been widely applied in the industry, such as robot ranging, obstacle avoidance and 3D reconstruction for cultural relics excavation scenes in archaeological sites, and our research is inspired by the latter one. In practice we found that stereo image pairs taken from excavation site environments have their own particularity, that is, much similar texture areas close to the soil color are involved in those images. In these areas, candidates within the search range still have some distinguishable features, but we need cost functions with more powerful capacity of discrimination to identify them correctly. Beside that, in order to obtain disparity maps of higher resolution, we have to adopt corresponding images of higher resolution in stereo matching, but their resolution can reach up to millions or ten millions of pixels, and far beyond the level of traditional stereo images having hundreds of thousands of pixels to megapixels. Increase of the amount of input data bring higher requirements to the computational speed of stereo matching methods, and many researchers have realized this issue, so developing stereo methods with higer computational efficiency has been an important research direction of stereo matching.We summarize above application issues to a kind of more general academic problem-stereo matching based on high-resolution similar texture image pairs, and research it from three aspects:(1) Research on the cost function with a higher capacity of discrimination. The traditional cost function has two steps, initial cost calculation and cost aggregation, of which cost aggregation plays a decisive role in the discrimination ability of cost function. Combining with image segmentation technique, this thesis presents a new cost aggregation strategy and test it in two different types of cost function operators, Census transform and adaptive support weight approaches. Experimental results show that the proposed cost aggregation strategy significantly improves the discrimination capability of cost matching functions and reaches the lowest average error rate among the same kind cost functions.(2) Research on global optimization algorithms with better smoothing capability and adapt-ability. The theory basis of global optimization algorithms is MRF-MAP framework, and it finds the disparity at each position by solving an energy function, in which belief propagation (BP) is the most widely used optimization method. We improve the classic hierarchical BP algorithm by join-ing the B-spline fitting based de-noising and smoothing module and get much smoother and more accurate disparity maps; we also propose a BP algorithm using asynchronous message passing, and the experimental results show that the proposed algorithm not only has the comparable convergence speed with the classic BP algorithm, but also has better de-noising ability and consumes less storage space. In addition, based on the BP using asynchronous message passing, we design a complete stereo matching algorithm by integrating several existing techniques and conduct experiments, and the average error rate on standard data sets rank six at that time.(3) Research on extensible hybrid stereo matching schemes. After analyzing the research progress on acceleration about current stereo matching, we design a hybrid stereo matching com-putational framework. The framework adopts some most widely used stereo matching techniques currently, such as multi-core CPU and GPU acceleration techniques, the disparity estimation tech-nique based on plane fitting, and combines them appropriately. We test this computational frame-work at the local and global stereo matching algorithms, and the results show that the framework can further improve the computation speed for local stereo matching algorithms, and can effectively reduce the peak memory demand for global stereo matching algorithms. Besides that, by making full use of kinds of computing resources in a heterogeneous system, this framework also has a high scalability.Research on the stereo matching technique based on high-resolution similar texture stereo image pairs covers hotspot issues of current stereo matching technology, and we propose effective solutions in this thesis. Experimental results show that proposed solutions are effective, and good results have been achieved in tests based on standard data sets.