基于一维匹配的宽基线图像多维匹配方法研究

To meet the practical needs of image matching, this project focuses to study the three main issues with wide-baseline image matching. First, the project will study an efficient algorithm for the one-dimensional (1D) mapping. One issue with the 1D mapping is that it can only make use of a small amount of data in mapping. As a result, the derived result from mapping can be unreliable sometime. To make the 1D mapping more reliable, the project will study the optimal way to employ redundant (> 2) radial lines to increase the amount of image information content available to mapping. At the same time, the project will study the method of using the projective invariance property of radial lines to detect and correct the false matches. Another issue with the 1D mapping is that it is still not computationally fast enough for practical use. To make the 1D mapping more time-efficient, the project will study the method used to derive the least number of radial data samples, which is, however, large enough to guarantee an accurate 1D mapping. Also the project will study the method of fast locating the target image position guided by the beneficial image features. Second, the project will study the method for the registration between wide-baseline images. Consider that in reality the overlap in input images is complicated in its shape and appearance. So, the project will study, via the use of the efficient 1D mapping, first how to determine a local matching region in small size, and then gradually expand the region outwards until it covers the entire overlap, which could be irregular in shape or has textured/structural similarity. Finally, the project will study the method for the dense stereo matching of wide-baseline images. According to the technique requirements of dense stereo matching, the project will first study the methods of using the efficient 1D mapping to match an entire image scanline segment by segment and to detect the endpoints of each line segment, then the methods of optimizing the derived dense-stereo-matching results, and finally the methods of handling occluded region. The proposed study can provide the scientific methods and technique support to the wide-baseline image matching, and promote the technological advances in the field of image processing.

本项目针对图像匹配实际需要，重点研究宽基线图像匹配的三个方面的问题：一是研究一维匹配的高效算法。针对一维匹配可利用数据少和速度不够快的的问题，研究用冗余径向线来增加匹配可用信息量的最佳途径，以及用射影不变性质来查错与纠错的方法，以提高其可靠性；研究匹配所需径向数据最少的模式，以及由图像特征引导的匹配块快速定位方法，以提高其时效性。二是研究宽基线图像二维全图配准方法。针对实际图像重叠区形状及形貌的复杂性，研究以高效一维匹配引导局域图像园块或特征点匹配、向外扩张并最后覆盖整个不规则的、纹理相似和结构相似兼有的图像重叠区的匹配方法。三是研究宽基线图像密集三维匹配方法。根据密集三维匹配的技术要求，研究由高效一维匹配来执行图像线分段匹配的模式，分段起始点的检测方法，对匹配结果优化的方法，以及处理不能匹配的遮挡区方法。研究成果可为宽基线图像匹配提供科学方法与技术支撑，推进图像处理领域的技术进步。

本项目针对宽极线图像匹配的实际需要，研究了在三种不同维度下实现宽基线图像匹配的方法，取得了可靠且有效的匹配结果，达到了预期的研究目的，具体内容如下。第一，针对原一维匹配方法缺乏健强性及时效性的问题，提出了一个新的基于冗余解投票的匹配查错方法，以及一个新的基于查错多尺度搜索的目标定位方法。据此提出了一个新的可靠且快速的宽基线图像一维匹配算法。同时，也开发出了相应的软件系统，并用INRIA数据组对该系统进行了评价，证实了该方法较原一维匹配方法新增的两个特点，即查错键强性与高时效性（平均要快三千倍以上），同时也保持了亚像素的配准精度（最大配准误差只为1.05像素）。第二，在宽基线图像二维全图配准时，针对实际图像重叠区的形貌复杂性，提出了一个新的在对数极坐标系下尺度不变特征点的检测方法。进一步，以这类特征对应点作为种子、向外扩张长大并覆盖整个重叠区的匹配模式，实现了宽基线图像二维全图配准。同时，也开发出了相应的软件系统，并用Rensselaer数据组对该系统进行了评价，证实了该方法的配准成功率比国际主流的基于尺度不变特征点的图像配准方法（如SIFT方法和GDB-ICP方法）的要高出13.64%。第三，根据密集立体（三维）匹配的技术要求，提出了一个新的基于形貌自适应窗口的立体匹配方法，该方法由一个形貌自适应的支持窗口划分，结合一个由区域边界约束强化的扫描线优化方法，以及系统的适于平行计算的视差查错与纠错方法，实现了准确的宽基线图像立体匹配。同时，也开发出了相应的软件系统，并按Middleburry基准检测对该系统进行了评价。结果表明，该系统对不同的宽极基线图像均能给出准确的视差结果，同时在平均视差错误上比同类型的Middleburry算法（如Cross方法）要低出25.05%。总之，系统的测试结果表明，三方面主要研究的技术指标已接近或达到了当前国际主流技术的水平。这些研究成果为推进图像匹配领域的发展提供了科学方法与人才支撑，也为今后进一步的开发应用打下了基础。