Efficient Methods for Automatic Recognition With Application to Target Identification

研究人员的研究目标是学习如何对复杂结构的组织进行编码，以实现有效的分类、识别和浏览。考虑了复杂的结构，例如对象、图像、3D扫描和网络文档。调查的重点是非常大的数据集(例如，数百万个)。由于计算问题，现有的浏览方法无法处理如此大的尺寸。作为这项研究的一部分，正在开发解决这些问题的理论工具和方法；然后将这些工具和方法应用于从三维成像激光雷达(激光雷达)测量中自动识别目标的问题。为基于几何/结构的查询索引大型数据库的困难主要是由于两个因素：1)速度和精度之间的内在冲突，2)维度的诅咒。研究人员正在开发的基于不变统计学的结构表示法解决了这些困难。首先，它们是完全不变的，因此它们绕过了寻找两个结构之间的最佳映射的问题，从而允许快速数据比较。其次，对于类属结构，它们是无损的，因此它们不会影响准确性。此外，它们允许低复杂性度量，从而产生几乎肯定是100%准确的快速比较算法(与总是近似的近似算法相反)。此外，由于它们不包含歧义，因此可以使用解决维度诅咒的高维索引技术对它们进行索引。最后，它们是浮点运算和底层数据友好的。

The objective of the investigator's research is to learn how to encode the organization of complex structures for efficient classification, recognition and browsing. Complex structures such as objects, images, 3D scans and web documents are considered. The focus of the investigation is on very large datasets (e.g., several millions). Because of computational issues, such large sizes cannot be handled by existing browsing methods. As part of this research, theoretical tools and methods for addressing these issues are being developed; these tools and methods are then applied to the problem of automatically identifying a target from 3-D imaging Laser Radar (Ladar) measurements. The difficulty of indexing a large database for geometry/structure-based queries is mostly due to two factors: 1) The inherent conflict between speed and accuracy, and 2) The curse of dimensionality. The structure representations that are being developed by the investigator, which are based on invariant statistics, address these difficulties. First of all, they are fully invariant and so they allow fast data comparison by bypassing the problem of finding the best mapping between two structures. Secondly, for a generic structure, they are lossless, and so they do not compromise accuracy. Moreover, they allow for low complexity metrics, thus yielding fast comparison algorithms that are almost surely 100% accurate (as opposed to approximate algorithms, which are always approximate.) In addition, as they contain no ambiguity, it is possible to index them with high-dimensional indexing techniques that address the curse of dimensionality. Finally, they are floating-point arithmetic and low-level data friendly.