时空轮廓编组计算模型及其在目标检测中的应用

Contour grouping groups the edge fragments to segregated groups with explicit visual meanings.It is one of the fundamental and most difficult problems in the field of computer vision, which connects low-level vision and high-level vision. Inspired by the mechanism of perceptual organization of human being, we'll study the spatio-temporal salient contour grouping model and its applicatin for object detection. This project is focused on: 1)A new long curve grouping element structure will be proposed by analyzing the spatio-temporal characteristics of the input, which are image sequences captured from natural scene.2) The quantification models for grouping rules with multi-features and multi-element will be built.And the approach for the fusion of grouping cues and the computing models for the unified grouping cost will be proposed by use of manifold learning. 3)We use the global saliency information and priori knowledge of shape to restrict the process of contour grouping on the basis of the research achievements about the attention role in perceptual organization in psychology.Based on the measure of global ideality for the grouping results, a contour grouping algorithm with feedback and reasoning will be proposed by studying the cooperation between global grouping cues and local grouping cues. 4) The computing model will be applied for specific object detection tasks in video. The latest research progress in the fields of biologic vision and machine learning are combined with this project and the research results can be applied in the domains of image understanding, image retrieval, and so on. This project is of significance both in theory and practice.

轮廓编组将边缘片段组织成具有明确视觉意义的独立目标轮廓，搭建了从初级视觉特征提取到高级视觉任务之间的桥梁，是计算机视觉领域中极其重要的研究课题。本项目以人类知觉组织机理为指导，深入研究时空轮廓编组计算模型，并将其应用于目标检测任务中。主要研究内容为：1）分析图像序列的时空域特性，提出一种新的基于时空信息的长曲线轮廓编组元结构；2）采用非监督流形学习理论研究多特征、多元编组线索的量化和融合，提出统一编组代价计算方法；3）建立以显著性注意和形状先验为约束的优化目标函数，以时空一致性为编组理想性度量，利用推理和反馈机制构建协同全局性和局部性的轮廓编组优化算法；4）在特定的目标检测任务中对构建的计算模型进行应用研究和测试验证。研究方案结合了生物视觉、机器学习等多领域的最新研究进展，研究成果可应用于视频内容理解与分析、基于内容的视频检索等领域,具有重要的理论意义和实用价值。

从单帧图像/图像序列中提取具有明确视觉意义的独立目标轮廓编组，不仅可作为原始视觉输入的简洁语义表达，而且搭建了从初级视觉特征提取到高级视觉任务之间的桥梁，是计算机视觉领域中极其重要的研究课题。本项目在时空轮廓计算理论研究基础上，在算法和模型层面展开深入研究，在生物视觉处理机制的启发下，探索了基于编组的时空轮廓计算模型优化，重点研究了基于学习的目标轮廓计算模型，以目标轮廓（语义边界）的视觉意义为基础，开展了基于目标轮廓和区域边界的视觉感知增强研究，在图像/视频目标检测、多视角分析、立体视觉应用等领域建立了若干应用系统。.项目组在四年的执行期内，针对研究目标，按照研究计划开展研究工作，取得了一些重要的理论研究成果，并成功地应用到了基于视觉的高速铁路基础设施服役状态智能监测等任务中，具有很好的应用前景。共发表学术论文26篇，包括期刊论文21篇，会议论文5 篇；其中，SCI检索论文7篇，EI期刊检索论文5篇，CCF A类会议1篇，B类会议1篇。申请专利7项，其中已授权4项，申请软件著作权2项，其中已授权1项。毕业博士生4名，毕业硕士生13名，获国家科技进步一等奖1项，中国专利优秀奖1项。