COMPPACT: Compression of Video using Perceptually Optimised Parametric Coding Techniques

目前，视频压缩是一个非常令人兴奋和具有挑战性的时期。预计带宽需求的增长，尤其是对移动服务的需求，很大程度上是由视频应用驱动的，现在可能比以往任何时候都要大。这有四个原因：(i)最近推出的3D和多视图格式，加上不断增加的动态范围、空间分辨率和帧率，都需要提高比特率来提供更好的沉浸感；以视频为基础的网络流量继续增长并在互联网上占主导地位；用户的期望继续推动灵活性和质量，从线性交付转向非线性交付；（iv）最后是新服务的出现，特别是通过4G/LTE向智能手机提供移动服务。虽然网络和物理层技术的进步无疑将有助于解决这一问题，但视频压缩的作用也至关重要。这个研究项目的基础假设是，在大多数情况下，视频压缩的目标是提供良好的主观质量，而不是最小化原始图像和编码图像之间的误差。因此，可以设想一种压缩方案，其中分析/综合框架取代了传统的能量最小化方法。这种方案可以通过减少残差和运动矢量编码来提供更低的比特率。提出的方法将使用波形编码和纹理替换的组合来建模场景内容，使用计算机图形模型来替换解码器中的目标纹理。这不仅提供了显著提高性能的潜力，而且还提供了与内容相关的固有参数化，这将用于分类和检测任务，并促进与CGI的集成。这有可能为视频压缩创建一个新的内容驱动框架。在这种情况下，我们的目标是将视频编码范式从率失真优化转变为率质量建模，其中基于区域的参数与感知质量指标相结合，以通知和驱动编码和合成过程。然而，很明显，为了充分利用该方法的潜力并产生稳定和有效的解决方案，需要进行大量的研究。例如，均方误差不再是有效的目标函数或质量度量，而新的嵌入式感知驱动的质量度量是必不可少的。纹理分析和合成模型的选择也很重要，对长期图像依赖关系的利用也很重要。

It is currently a very exciting and challenging time for video compression. The predicted growth in demand for bandwidth, especially for mobile services is driven largely by video applications and is probably greater now than it has ever been. There are four reasons for this: (i) Recently introduced formats such as 3D and multiview, coupled with increasing dynamic range, spatial resolution and framerate, all require increased bit-rate to deliver improved immersion; (ii) Video-based web traffic continues to grow and dominate the internet; (iii) User expectations coninue to drive flexibility and quality, with a move from linear to non-linear delivery; (iv) Finally the emergence of new services, in particular mobile delivery through 4G/LTE to smart phones. While advances in network and physical layer technologies will no doubt contribute to the solution, the role of video compression is also of key importance.This research project is underpinned by the assumption that, in most cases, the target of video compression is to provide good subjective quality rather than to minimise the error between the original and coded pictures. It is thus possible to conceive of a compression scheme where an analysis/synthesis framework replaces the conventional energy minimisation approach. Such a scheme could offer substantially lower bitrates through reduced residual and motion vector coding. The approach proposed will model scene content using combinations of waveform coding and texture replacement, using computer graphic models to replace target textures at the decoder. These not only offer the potential for dramatic improvements in performance, but they also provide an inherent content-related parameterisation which will be of use in classification and detection tasks as well as facilitating integration with CGI. This has the potential to create a new content-driven framework for video compression. In this context our aim is to shift the video coding paradigm from rate-distortion optimisation to rate-quality modelling, where region-based parameters are combined with perceptual quality metrics to inform and drive the coding and synthesis processes. However it is clear that a huge amount of research needs to be done in order to fully exploit the method's potential and to yield stable and efficient solutions. For example, mean square error is no longer a valid objective function or measure of quality, and new embedded perceptually driven quality metrics are essential. The choice of texture analysis and synthesis models are also important, as is the exploitation of long-term picture dependencies.