基于超像素网格的通用物体定位算法研究

Image and video object localization algorithm has important applications in video surveillance, image understanding, robot navigation and other aspects. However, existing object localization algorithm is difficult to meet the actual needs, mainly reflected in: (1) the existing algorithms is designed for clear media. For degraded media, one should restore it and then perform localizing, (2) the algorithm is inefficient and difficult to achieve real-time localizing; (3) does not utilize the tags fully. To solve the above problem, we will (1) extend the existing feature extraction theory. During object localization, we will extract and match features directly from the degraded media to avoid introducing additional noise and improve computational efficiency; (2) propose super pixel grid structure. We can regularize arbitrary superpixels with irregular structure to a grid structure. When using the sliding window schema to perform object localization, it can improve the computational efficiency; (3) propose performing clustering on sub-window. Due to the introduction of weakly-supervised label, cluster sub-window can further improve the efficiency and reduce the workload of user interaction during localization. The research result of this proposal will greatly increase the computation efficiency and accuracy of object localization algorithm. Besides, it can provide a firm foundation for intelligent image analysis and application in theoretical and practical aspects.

图像及视频的物体定位算法在视频监控、图像理解、机器人导航等方面有着重要的应用。 现有的物体定位算法难以满足实际的需要，主要体现在：（1）现有算法面向清晰媒体，对于降质媒体，需要先恢复再定位；（2）算法效率低下，难以做到实时定位；（3）训练过程中需要过多标记，无法利用现有标签直接进行训练。针对以上问题，本课题将（1）扩展现有的特征提取理论，在进行物体定位时，直接从降质媒体上进行特征提取及匹配，避免引入额外噪声并提高计算效率； （2）提出超像素网格结构，将任意不规则的超像素结构进行规整化处理，得到网格结构，并在之上进行物体定位，提高计算效率；（3）提出针对由于标签而引入的弱监督约束进行子窗口聚类，减少训练过程中用户交互的工作量并进一步提高物体定位的执行效率。本课题的研究将大幅度提高物体定位算法的计算效率和精度，并在理论和实际两方面为智能图像分析与应用打下坚实基础。

图像及视频的物体定位算法在视频监控、图像理解、机器人导航等方面有着重要的应用。现有的物体定位算法难以满足实际的需要，主要体现难以在图像上提取具有高表现力的特征，无法对目标进行高精度分割及快速定位。针对以上问题，本课题进行了以下研究（1）扩展现有的特征提取理论并将其应用于进行物体定位时、图像分割、显著性区域检测等方面；（2）提出超像素网格结构，将不规则的超像素结构进行规整化处理，得到网格结构；（3）深入研究了通用物体定位、图像分割、保持边界的图像平滑算法之间的内在共通性。本课题的研究为大幅度提高物体定位算法、图像分割算法的计算效率和精度，并在理论和实际两方面为智能图像分析与应用打下坚实基础。