数据驱动的高效视频编码理论与方法

High definition videos play an important roles in multiple areas, including national and metropolis security, military simulation, movie and entertainment, education and medication, etc., which leading to the video data volume doubles every year. Highly effective video compression for these large scale videos are highly desired, and it is also one of the hot topics in worldwide research. Since knowing the visual factors of human vision is still limited, the traditional video coding system still adopts the Peak Signal-to-Noise Ratio (PSNR) to measure the video quality. However, the PSNR cannot precisely predict the real perceptual quality, thus, the visual redundancies in videos encoded by traditional video codec are not sufficiently eliminated. On the other hand, the current coding standard significantly increases the number of mode candidates to improve the coding efficiency. The gain is limited but the computational complexity increases significantly. To solve these problems, this project will construct the subjective video dataset and the coding parameter dataset, and investigate the relation among the video representations, visual perceptions and parameters in video coding systems based on data-driven analytic tools. Then, we will model the regression problems in the coding process, investigate perceptual rate-distortion model , and establish the objective function and relation functions. Finally, we are going to design data-driven high efficient video coding algorithm and find solutions for the specific regression and classification problems. Coding efficiency and effectiveness can be improved by jointly and reasonably allocate the rate, perceptual distortion and complexity. This project will improve the theories and technologies in data-driven video coding, which is thus boost the development of signal processing community and coding technologies.

高清视频在国家与城市安全、军事模拟、影视文化、医疗教育等领域应用广泛，以致视频数据总量约每两年翻一番。海量视频的高效压缩是急需解决难点问题，是当前国内外研究热点。由于感知因素尚不明朗，编码系统仍以峰值信噪比度量视频质量，难以真实反映主观感知，导致视觉冗余消除效果有限；另一方面增加备选模式以计算量换取增益，代价巨大且收益小。本项目将构建主观感知和编码参数集，以视频内容、视觉特性以及编码系统等联合数据为驱动，研究视频内容的有效表示及其与高效编码和视觉感知的关联性模型，发掘海量数据中隐含知识和规律；然后，建模编码中回归问题，研究感知-率失真优化模型，构建一般性目标和关系函数；最后，设计数据驱动的高效视频编码方法，求解回归的优化目标，合理配置码率、质量和计算等资源，实现高效压缩。本项目可实现数据驱动的高效视频编码理论的创新与技术突破，促进信号处理学科和编码技术的发展。

高清视频在国家与城市安全、军事模拟、影视文化、医疗教育等领域应用广泛，海量视频的高效压缩是急需解决难点问题，是当前国内外研究热点。本项目构建主观感知和编码参数集，以视频内容、视觉特性以及编码系统等联合数据为驱动，研究视频内容的有效表示及其与高效编码和视觉感知的关联性模型；然后，研究感知-率失真优化模型，构建一般性目标和关系函数；最后，设计数据驱动的高效视频编码方法，合理配置码率、质量和计算等资源，实现高效压缩，并降低计算复杂度。项目研究成果已在本领域国内外代表性学术期刊发表(录用)SCI期刊学术论文11篇，包括IEEE Trans. Image Process 3篇，ACM TOMM 1篇，INS 1篇, Neurocomputing1篇，其中JCR1区8篇，二区3篇，发表EI会议论文2篇，公开数据测试集3个，申请中国/PCT发明专利7项，课题组如期顺利完成研究任务。