高分辨率视频编码参数优化的关键技术研究

In video coding, many encoding parameters are adjustable, and different parameters sets bring different coding performance. In H.264 encoder, there are more than 100 encoding parameters. Over 10 additional encoding parameters would be included in the next-generation High Efficiency Video Coding (HEVC) standard under development. The encoding parameter space is also expanded with the increasing video definition. The selection of encoding parameters is critical to video encoders for optimal rate-distortion performance with videos of different contents and high-definitions. Traditionally, to choose the encoding parameter set of near-optimal performance, multi-pass encodings are generally required, which is quite time-consuming for encoding high-definition videos. Besides, many works focus on the selection of several encoding parameters by fixing other parameters, which would obviously affect the whole performance. The main constraint to the parameter selection is that the rate-distortion performance is obtained after video encoding. From our previous research on image coding, the compression quality could be predicted accurately with the image feature under a set of encoding parameters without actual image encoding. Based on previous work, a pre-encoding performance prediction model is presented for video encoding parameter space in this proposal. With the performance prediction model, the encoding parameters space would be explored with lower computation cost and time consuming. To further reduce the dimension of encoding parameter space, the parameters would be pre-selected and cropped according to encoding scheme. The proposed methods would improve encoding performance of H.264 and HEVC with lower computation cost. Furthermore, the theory and methods would be implemented on the real-time high-definition video encoder IP, which targets on the design and application of the home-made processors.

视频编码有大量可调的编码参数，不同的参数组合会带来不同的性能。随着视频分辨率的提高和H.264、HEVC的制定，编码参数空间成指数级增长。如何从中选出合适的参数组合获得较高的率失真性能，已经成为高分辨率视频编码的核心难题。传统上，编码参数组合的选择需要多次全部或部分编码，否则只能调节个别参数，极大地影响了编码速度和性能。申请者在图像编码的前期研究发现，事实上不需要编码，仅结合图像特征，就可能以极低代价准确地预测出一组参数组合的编码性能。以此为基础，本项目提出建立视频编码的编码前性能预测模型。通过该预测模型，无需编码即可快速准确地得到一组视频编码参数组合的性能，可望对编码参数空间搜索带来变革。再结合参数空间的预选、裁剪技术，即可实现一个实用高效的H.264 和HEVC编码参数优化方法。本项目的研究成果最终将实用于国产实时高分辨率视频编码IP研发中，并力争为同行提供一个开源的编码参数优化软件。

面向高分辨率视频编码应用，与性能相关的参数空间成指数级增长。如何选出合适的参数组合来获得较高的率失真性能、减少遍历参数空间时带来的高计算访存开销，是高分辨率视频处理的核心难题。本项目提出利用图像特征和机器学习/分析方法来建立性能评估/预测模型，从而避免遍历参数空间中大量参数组合来找出最优的参数组合；同时提出性能建模和参数空间裁剪机制，可以有效降低计算访存开销，提高高分辨率视频的处理速度。本项目通过提供给同行一套适用于高分辨率视频的高效低复杂度的参数优化算法及视频加速器架构，可望为国产多媒体处理器的设计提供借鉴。. 基于上述思想，本项目完成了以下工作：1.结合视频编解码的计算和访存特点，提出图像特征提取及性能优化方法。2.为了减少机器学习和统计方法中回归模型训练的大量仿真时间、并提高学习精度，提出了一套设计高鲁棒的设计空间建模方法。3.为了高效支持上述相关方法，设计了视频加速器架构。该视频加速器相对于CPU和GPU有数十倍的能效加速，可以支持视频理解应用。. 通过本项目的研究，本项目组发表了16篇论文（包括9篇SCI、16篇EI论文）和4项发明专利。相关工作发表在多种领域顶级会议（包括MICRO、ISCA、ASPLOS）以及多种IEEE/ACM Trans.（包括IEEE Trans. on Computers、IEEE Trans. on Parallel&Distributed Systems、ACM Trans. on Computer Systems、ACM Trans. on Design Automation of Electronic Systems）上。