Complexity-quality tradeoffs for real-time signal compression

The growth of computer networks in the past decade has been accompanied by a growth in the types and complexity of media-related applications that run over these networks. During the same time, the computing platforms that form the nodes of these networks have grown in complexity, partially in order to handle the difficulty of these applications. At any given time, the latest applications have provided a challenge for the latest computing platforms. These applications have several characteristics: many of them need to complete certain tasks by real-time deadlines, and also admit some sort of quality-complexity tradeoff. This tradeoff means that the tasks of the applications can be completed faster and/or with fewer computational resources for a possible sacrifice in the quality of result, for example a lower signal fidelity and/or an increased bitrate in an encoded audio or video bitstream. The nature of these tradeoffs are usually dependent upon the input data to the tasks. We are concerned about situations where the computing platform is handling many such tasks. More specifically, we address the allocation of computational resources in such situations, in order to get the best overall quality among all tasks currently running. We first discuss the quality-complexity tradeoffs inherent in media applications, both from a theoretical perspective and a practical perspective. We propose a methodology and framework for transferring computational resources between particular tasks in an application. We evaluate the input and output data of each task to assess its quality of result, and make adjustments to its complexity. The goal of these adjustments is to improve the application's overall quality of result while keeping its overall complexity under a given constraint. We then adapt our framework for implementation with a conventional operating system and a simulator for SCORE, a computational model for reconfigurable/highly parallel computing systems. We apply our framework to H.264 video compression and demonstrate that we can eliminate 20-30 percent of certain intensive calculations with less than 5 percent increase in bitrate for the same fidelity of decoded video.