结合感知分组和视觉注意特性的显著目标检测研究

Aiming at the problem that the computer’s visual perception ability cannot meet the needs of today' s society, this project starts from the cognitive characteristics of human vision and studies salient object detection method in complex natural scenes, which helps to improve the performance of computer vision. By modeling and combining the perceptual grouping and visual attention characteristics of human vision, this project enables the computer to quickly and effectively detect the salient image region containing the key information of a complex natural scene. This project includes the following research methods and goals: (1) The image is grouped into multiple perceptual consistent regions by exploiting PRSVM and multi-label graph based ranking algorithms and solving the confliction between multiple perceptual grouping characteristics; (2) Extract the image regions which grab the visual attention of humans based on complementary low-level and high-level cognitive properties of visual attention; (3) Through the reasonable construction of hypergraph based regularization framework, the perceptual grouping and the visual attention characteristics are merged to obtain the final result of visual salient object detection in complex scenes. This project tries to effectively model and fuse the perceptual grouping and visual attention characteristics of human vision for the first time, aiming at improving the robustness of salient object detection method in complex natural scenes. This project will introduce new ideas for the research on salient object detection in complex natural scenes, and provide better technical support for the follow-up computer vision tasks.

针对计算机视觉感知能力不能满足当今社会需求的问题，本项目从人类视觉的认知特性出发，研究有助于提高计算机视觉性能的复杂自然场景显著目标检测方法。本项目通过结合人眼的感知分组和视觉注意特性，使计算机能够在复杂自然场景中快速有效地检测出包含关键信息的显著区域。具体的研究内容包括：(1)基于PRSVM和多标签图排序算法对图像进行感知分组，解决多感知分组特性间的冲突问题和多图像区域间的分组关系描述问题；(2)提取和学习具有互补作用的低层和高层视觉注意特性，提高复杂自然场景图像中显著区域的定位精度；(3)通过合理构造适用于显著目标检测问题的超图正则框架，解决感知分组和视觉注意特性的有效融合问题。本项目首次结合人眼的感知分组和视觉注意认知特性，旨在提高显著目标检测模型对复杂自然场景的鲁棒性。本项目的开展将为复杂自然场景中的显著目标检测研究引入新的思路,为后续的计算机视觉任务提供更好的理论和技术支持。

视觉显著性检测通过从图像中提取包含关键信息的显著区域来提高计算机视觉和图像处理的性能和速度，是一种强有力的数据分析工具。视觉显著性检测可以广泛地应用于计算机视觉和图像处理领域。本项目对人眼的感知分组特性进行建模，解决了复杂自然场景中多个图像区域之间分组关系的表示问题。对多种感知分组特性的整合方法进行研究，解决了人类视觉多种感知分组特性之间的冲突和竞争问题。研究基于低层和高层认知特性的视觉注意建模方法，提高了复杂自然场景图像中显著区域的定位精度。本项目进一步研究适用于显著目标检测问题的整合框架，解决了感知分组和视觉注意这两类认知特性的有效融合问题。本项目完成了基于人眼的感知分组特性和视觉注意特性的复杂自然场景显著目标检测的研究与开发。在MSRA10K和ECSSD等显著目标检测数据库上进行了测试，验证了本项目提供的基于视觉认知特性的复杂图像显著目标检测方法有利于在复杂自然场景图像中检测出显著目标。本项目的研究提高了显著区域的检测性能，可以为机器视觉其他任务提供应用和技术支持。