面向无约束视频的时空显著性模型及其应用研究

项目名称： 面向无约束视频的时空显著性模型及其应用研究

项目编号： No.61471230

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 刘志

作者单位： 上海大学

项目金额： 80万元

中文摘要： 近年来用于视频的时空显著性模型已逐渐成为国际前沿的研究热点，但目前各类模型的普遍缺陷是难以有效处理无约束视频，其时空域特征的复杂性会导致现有模型的显著性检测性能严重下降。为有效克服现有模型的缺陷，本项目提出基于区域级时空域特征的时空显著性模型。首先，提取区域运动轨迹描述符及帧内/帧间区域相似性矩阵等时空域特征，以增强对象区域和背景区域的可区分度。然后，系统地提出区域级时域/空域显著性度量、显著性帧间传播与调整、基于置信度的显著性融合及显著性帧内扩散等方法来生成区域级及象素级时空显著性图，以有效提升对无约束视频的显著性检测性能。最后，提出适于无约束视频的时空显著对象检测与分割方法，利用窗口轨迹显著性、区域级对象分割与显著性修正的联合优化及象素级局部修正，来提升检测与分割性能并充分验证所提出模型的有效性。预期研究成果不仅将丰富并发展显著性模型的研究，而且将推动基于显著性的视频处理技术的发展。

中文关键词： 视频信息处理；时空显著性模型；无约束视频；区域级显著性图；基于显著性的视频处理

英文摘要： Spatiotemporal saliency model is becoming an international cutting-edge research topic in the recent years. However, the common drawback of the current spatiotemporal saliency models is their insufficiency for processing unconstrained videos. The complexity of spatial and temporal features in unconstrained videos results in a significant degradation on saliency detection performance of the current models. In order to overcome the drawback of current models, this project proposes a spatiotemporal saliency model based on region-level temporal and spatial features. First, region-level motion trajectory descriptors and intra-frame/inter-frame regional similarity matrices are extracted to enhance the discrimination between object regions and background regions, in terms of region-level features. Then, region-level temporal/spatial saliency measure, inter-frame saliency propagation and adjustment, confidence based saliency fusion, and intra-frame saliency diffusion are systematically proposed to generate region-level and pixel-level spatiotemporal saliency maps, in order to effectively improve the saliency detection performance on unconstrained videos. Finally, this project proposes spatiotemporal salient object detection and segmentation approaches suitable for unconstrained videos, specifically, saliency of window's trajectory, joint optimization of region-level object segmentation and saliency boosting, and pixel-level local refinement, are exploited to improve the performance of object detection and segmentation, and to validate the effectiveness of the proposed spatiotemporal saliency model. The expected research results will not only enrich and promote the development of research on saliency model, but also advance the saliency-based video processing technology.

英文关键词： video information processing;spatiotemporal saliency model;unconstrained video;region-level saliency map;saliency-based video processing