Dynamically Reconfigurable Hardware Architectures for Context-Based Statistical Compression of Visual and Data Content

这项工作的目的是研究基于上下文的视觉和数据内容统计无损压缩的算法和硬件架构，使用动态可重构硬件来支持每种数据和压缩类型的最佳建模策略。建模输出的熵编码将使用静态配置的算术编码引擎执行。当前网络融合的趋势是，视觉和数据内容沿着同一物理通道传输，这表明能够提供最佳压缩比和快速适应内容性质的技术将变得越来越重要。这是推动这项研究工作的两个关键概念。基于上下文的统计压缩与流行算法（如ZIP家族）中使用的基于字典的压缩有着根本的不同，它被认为能够提供比这些更好的压缩比。然而，这只能通过复杂的软件算法来实现，这些算法需要相当大的内存容量，并且每字节的吞吐量非常低，只有数千个CPU周期。这意味着需要耗电的奔腾4级微处理器以GHz速率运行，以提供运行这些高级统计算法所需的计算能力，甚至这些cpu也难以支持远程医疗等应用，这些应用需要高带宽的静态图像、视频和科学数据进行无损实时压缩。其他应用，如空间中的数据、视频和图像传输，需要在能源和硅高效平台上实现性能。为了实现这些应用的要求，我们提出了第一个通用的无损压缩硬件核心，结合了基于上下文的变阶统计建模和算术编码。目前，还没有这些技术的实际硬件实现，因为还没有为可行的体系结构提出令人满意的解决方案。

The purpose of this work is to investigate algorithms and hardware architectures for context-based statistical lossless compression of visual and data content using dynamically reconfigurable hardware to support optimal modelling strategies for each data and compression type. Entropy coding of the modelling output will be performed using a statically configured arithmetic coding engine. The current trend of network convergence where visual and data content are transmitted along the same physical channel suggests a technology capable of delivering optimal compression ratios and fast adaptation to the nature of the content will become increasingly important. These are the two key concepts that will drive this research effort. Context-based statistical compression differs fundamentally from dictionary-based compression as used in popular algorithms such as the ZIP family and it is recognised as being able to offer superior compression ratios to these. However, this has been only achieved with complex software algorithms that require considerable amounts of memory capacity and have very low throughputs in the range of thousands of CPU cycles per byte. This means that power-hungry Pentium 4 class microprocessors running at GHz rates are needed to provide the required computing power to run these advanced statistical algorithms and even these CPUs will find difficult to support applications such as telemedicine where still images, video and scientific data would require lossless real-time compression with high bandwidths. Other applications such as data, video and image transmission in space require the performance to be achieved in an energy and silicon efficient platform. To achieve the demands set by these applications we propose the first universal lossless compression hardware core combining context-based variable-order statistical modelling and arithmetic coding. At present, there are no practical hardware realisations of these techniques, since no satisfactory solutions have yet been proposed for a viable architecture.