多态图像环境下的图嵌入算法及应用基础研究

With the rapid development of the technologies, such as video surveillance and robot vision, the traditional technologies of image analysis, understanding and recognition are confronting much more challenges. In general, object image possesses multi-states property in these fields listed above. And in recent years, the research on graph embedding methods, based on Spectral Graph Theory, is always the hot topic in manifold learning and shows many advantages in fields of image recognition and classification. Considering that the graph embedding technology can meet the requirements of robustness and real time in the scenario of multimodal images, we aim to conduct a further study on manifold learning theory together with the existing multi-manifold cluster and supervised graph embedding algorithms, then design a more robust, real-time and effective multimodal graph embedding algorithm, which will be optimized by sparse representation in later work, for the scenario of multi-states image recognition. Our research will show much theoretical value and a broaden application prospect, as the work of graph embedding algorithms for multi-states image recognition is still at an incipient stage at present.

随着视频监控及机器人视觉等新型技术的迅速发展，传统的图像分析、理解、识别技术面临新的挑战。对于视频监控及机器人视觉感知这些应用领域来说，获得的目标图像一般具有多态性的特点。基于谱图分析理论的图嵌入技术（Graph embedding）研究一直是流形学习研究领域的热点。近年来，关于图嵌入技术的一些深入研究显示了该技术在图像识别、分类等应用领域的比较优势。本项目注意到了基于图嵌入框架将可能满足多态图像环境对健壮性、实时性的较高要求。计划在理论上深入分析流形学习理论并对已有的多流形聚类、监督图嵌入算法作进一步研究，在此基础上针对多态图像环境的特点设计出可以健壮、实时、准确地处理多态图像的图嵌入算法，并进一步实现基于稀疏表示的优化算法及其它优化研究。目前国际上对多态图像的图嵌入技术研究尚处于起步阶段，本项目的研究工作将具有较大的理论价值和广阔的应用价值。

本项目按预定计划，完成了文献调研、理论研究、算法设计、数据获取及原型系统实现等工作。首先，我们完成了适合多态图像环境的近邻图设计方案，并针对交通信号识别、目标追踪等领域设计并实现了具有比较优势的图嵌入算法。其次，我们完成了结合稀疏表示技术的多态图像图嵌入算法的优化设计，并在特定领域证明了其高效、实时、准确的比较优势。另外，针对实时交通信号识别领域实现了基于移动终端的交通信号识别原型软件系统，申请并获得一项国家发明专利授权。最后，我们还开展并完成了多态图嵌入在迁移学习框架下的扩展研究，设计了基于图嵌入的迁移学习算法。本项目圆满完成了预定的研究目标及成果目标，还在成果应用转化及与迁移学习结合的方向上比预定目标有所突破。