视觉判别先验建模及面向识别理解的视频编码研究

Nature images adding prior knowledge gives us perception. The processing of vision information in the human brain depends not only on the characters of the vision information, but also on the prior knowledge, i.e. the general regulations of the nature images generalized and memorized by the human brain. Based on the understanding of the mechanism of human vision in the pathway of vision information transmission and human vision characters, our project tries to give summarization the general laws of the nature images corresponding to the human vision perception process, especially the discriminative prior model as well as the corresponding key techniques in the video coding. The dominant contents include: developing pervasive and effective prior modeling and the image prior models fusion strategy; developing the theory and algorithms for the discriminative prior model and then compromising the discriminative vision prior and compressed video reconstruction, proposing the methods of vision information prediction and reconstruction. Accordingly, we will focus on the problems in the current video coding frameworks, and propose the method of high quality images reconstruction and reliable quality for analysis and classification in the condition of low ration; these measures can be applied to break through the independence of video coding and video analysis, which are expected to make contribution to the future applications.

自然图像加先验知识才能产生认知：大脑获取先验知识，归纳与记忆自然图像的一般性规律，为判别和理解提供根本保障。该项研究从视觉通路及人类视觉特性两个方面，理解与总结人类视觉系统的机理；并以此为基础，研究视觉判别先验建模及面向识别理解的视频编码理论与方法，完成以下目标：建立并完善普适、高效的自然图像先验模型学习和融合策略，面向计算视觉层面的识别与理解，挖掘自然图像的判别性信息及其优化表达，探索并建立视觉信息判别先验建模理论和方法；进而面向兼容智能分析的视频编码，研究视觉信息感知、语义层面上失真度量方法，以此为基础构建基于视觉特性的率失真模型；最终综合以上，选择合理的视觉先验模型介入的途径与方式，设计融合判别先验信息的视频表示与重建策略，研究出在较高压缩效率下如何兼顾图像的高质量重构以及后续分析分析与理解。本项目预期改变目前视觉信息分析判别与编码彼此孤立的现状，为视频智能应用奠定理论和技术基础。

自然图像加先验知识才能产生认知：大脑获取先验知识，归纳与记忆自然图像的一般性规律，为判别和理解提供根本保障。该项研究从视觉通路及人类视觉特性两个方面，理解与总结人类视觉系统的机理；并以此为基础，研究视觉判别先验建模及面向识别理解的视频编码理论与方法，完成了目标：建立并完善普适、高效的自然图像先验模型学习和融合策略，面向计算视觉层面的识别与理解，挖掘自然图像的判别性信息及其优化表达，探索并建立视觉信息判别先验建模理论和方法；进而面向兼容智能分析的视频编码，研究视觉信息感知、语义层面上失真度量方法，以此为基础构建基于视觉特性的率失真模型；最终综合以上，选择合理的视觉先验模型介入的途径与方式，设计融合判别先验信息的视频表示与重建策略，研究出在较高压缩效率下如何兼顾图像的高质量重构以及后续分析分析与理解。