面向人类视觉感知的高分辨率遥感图像检索研究

With rapid development of space exploriation technologies， the acquisition of massive high resolution remote sensing data comes true, while during the application we still face to effective information deprivation. The key problem lies in the limited data processing and analysis capability, especially lack of effective retrieval means, so that vast quantities of information are submerged in the ocean of data. This project takes aim at extracting interesting scene which should be according with the characteristics of human visual perception from complex high resolution satellite data, main research content includes semantic representation mechanism based on multi-channel visual features, visual perception model with personalized information, task-driven complex scene retrieval algorithm and so on. As high resolution data is charateristiced with high dimension and diverse land cover, we firstly analysis the visual stimulation ability and assitant interpretation capacity of local visual features, then give a complementarity feature set to describe the image comprehensive semantic. with this basis, personalized concepts such as attention focus, objective spatial object are imported to hierarchical visual attention model which contains three layers: pixel, element and concept, with assitant of this model the mapping of image, feature and concept knowledge is established. At last, we concern on bridging the gap between perception model and actual visual attention process, under the leading of visual expection, the modeling of complex scene feature space and measurement of visual similarity are studyed, finally a more precise retrieval algorithm which faces to human visual perception is proposed.

空间探测技术的飞速发展使得海量高分辨率遥感数据的获取成为可能，应用实践中却仍然面临着信息匮乏的窘境，关键原因在于数据处理能力滞后，特别是缺乏有效的检索手段，使得大量有效信息淹没在数据海洋中。本课题以从复杂高分辨率遥感数据数据中提取出符合视觉感知的感兴趣场景为目标，研究视觉空间中多通道特征的语义表达机制、包含个性信息的视觉感知模型、任务驱动的复杂场景检索算法等内容。首先针对高分辨率卫星数据特征空间维度高、覆盖类型复杂的特点，分析局部特征的视觉刺激能力和辅助解译能力，研究能够表征综合语义信息的互补性视觉特征集合；在此基础上，引入注视焦点、地理对象等个性化概念信息，提出包含像素、基元及概念的层次化视觉感知模型以建立图像、特征及概念知识之间的映射；最后研究视觉期望引导下，研究复杂场景的视觉特征空间建模及视觉相似性度量，缩小视觉感知模型与实际视觉注意方式的差距，提出更为逼真的面向人类视觉感知的检索方

本课题主要从视觉显著性的角度研究了针对高分辨率遥感图像检索的技术和方法，重点研究了视觉空间中遥感图像显著特征的表达方法、显著度模型以及在图像检索中的应用，研究的主要成果主要包括研究了采用注意焦点以及边缘点两类兴趣点描述个性化信息融合方法，提出了以点特征为基础的局部显著特征遥感影像检索方法；研究了以图像块为注意单元的对象级特征提取方法和显著性度量方法，提出了一种既强调了主要对象又兼顾图像背景信息的遥感影像检索算法；研究了通过图像中底层特征、中层对象及高层语义信息之间的关联关系，提出了一种基于视觉关键词模型的遥感图像检索方法，该方法可以融合具有不同类型的属性如点特征、纹理和颜色等，可有效提高图像检索的准确率。