基于视频内容智能分析的双域信息提取与任意尺寸重构算法研究

In this research, we focus on how to sample natural scene more effectively and how to achieve any size reconstruction (including downsized reconstruction and expansion or the original size reconstruction). The study includes: ( 1 ) Establishing a video compression sensing sampling model based on analysis of the importance of video features to make more efficient sampled data ; ( 2 ) Deeply mining sampled data’s features in spatial- time - frequency domain. Reconstructing high-quality spatial domain information of any size video frames; ( 3 ) Establishing a compression scheme based on the transmission of very low bit-rate video compression reconstruction. In traditional compressed sensing reconstruction to achieve multi-resolution video processing, video frequency- spatial domain characteristics are always ignored to perceived compressing sampling. In the reconstruction process, the video’s spatial - time domain correlations have been completely ignored. Only the solving of convex optimization problem is applied to achieve reconstruction. In this proposed research, we aim at integrating videos’ spatial-time - frequency domain features extraction / analysis with compressing sensing and any size reconstruction. Compared with traditional methods, it shall greatly reduce the computation, and improve the reconstruction subjective and objective qualities.

本课题主要研究如何对原始视频进行更佳的压缩采样并利用采样后的数据实现空域任意尺寸重构（包括缩小重构、原始尺寸重构和扩展重构）。研究内容主要包括：(1)建立基于视频频率重要性分析的压缩采样模型，将视频特征提取分析与压缩感知采样有机结合起来使得采样数据更加有效；（2）深入挖掘采样数据的空-时-频域信息重构高质量的空域任意尺寸（缩小重构、原始尺寸重构和扩展重构）视频帧；（3）建立一种基于压缩重构的极低码率视频压缩传输方案。在传统的压缩感知实现空域多分辨率重构的视频处理中，人们忽略了视频的频域-空域等特征进行感知采样；在重构过程中又完全忽略了视频图像特有的空域-时域相关性而单纯研究求解凸优化问题实现重构。本课题所研究的方法是将视频的空-时-频域特征提取/分析、感知采样和空域任意尺寸重构融为一体，在算法复杂度上较传统方法有大幅度的降低，在重构视频的主观和客观质量上较传统方法有极大的提高。

经过课题组的共同努力，本课题建立基于视频频率重要性分析的压缩采样模型，将视频特征提取分析与压缩感知采样有机结合起来使得采样数据更加有效；深入挖掘采样数据实现了在空-频域重构高质量的空域任意尺寸（缩小重构、原始尺寸重构和扩展重构）视频帧；建立无参考帧质量估计模型有效实时控制重构过程、调整参数使重构获得更高质量。本课题所研究的方法是将视频的空-频域特征提取/分析、感知采样和空域任意尺寸重构融为一体，克服了在传统的压缩感知实现空域多分辨率重构的视频处理中，忽略了视频信号的频域-空域等特征进行感知采样；在重构过程中又完全忽略了视频信号特有的空域-时域相关性而单纯研究求解凸优化问题的缺陷，实现重构在算法复杂度上较传统方法有大幅度的降低，在重构视频的主观和客观质量上较传统方法有极大的提高。本研究基本完成了项目计划的主要研究内容，达到了预期目标。