Collaborative Research: DDDAS-SMRP: Optimizing Signal and Image Processing in a Dynamic, Data-Driven Application System

该项目将在资源有限的情况下开发一个动态的、数据驱动的信号和图像处理应用系统。该系统将为基础信号和成像处理计算的高度优化实现奠定基础，这些计算出现在许多科学和工程问题中，包括图像识别、通信分析、语音处理、从多媒体数据库中查询、索引和检索，以及航空、卫星和天文图像的图像分割。提出的多学科方法根据数据的特性和环境和目标硬件设备的独特要求，从算法规范、数学表示到软硬件(FPGA)实现进行优化。该系统的新颖性有两个方面：(1)它在数学、软件和硬件(片上系统)领域进行联合优化；(2)它是一个动态的、数据驱动的系统，因为信号处理变换根据算法要求和输入信号量身定做，以减少失真和增加压缩，并且可以在执行期间查询和控制系统。实施的基础是问题的最佳数学公式，再加上通过将数学特性与实施行为联系起来的模型的集成，在备选方案空间中自动选择最佳实施。硬件和软件优化都以统一的方式处理。与以传统方式派生的实现相比，预计使用这些方法设计时间将减少两个数量级或更多。由于建议的系统可以探索超出人类设计者能力的广泛实现，因此项目中所追求的方法所派生的实现甚至可能表现出更低的资源成本和更高的性能。这项研究将为所有规模的信号处理提供基础，为工程师构建复杂的、分布式的自适应信号处理传感器网络提供关键构建块。

This project will develop a dynamic, data-driven application system for signal and image processing under resource constraints. This system would lay a foundation for highly optimized implementations of fundamental signal and imaging processing computations that arise in many science and engineering problems, including image recognition, communications analysis, speech processing, querying, indexing, and retrieval from multimedia databases, and image segmentation of aerial, satellite, and astronomical images. The proposed multidisciplinary approach optimizes from algorithm specification, to mathematical representation, to software and hardware (FPGA) implementation, based on properties of data and unique requirements of the environment and the target hardware device. The novelty of the system is twofold: (1) it performs joint optimizationacross mathematical, software and hardware (system-on-a-chip FPGA) domains; and, (2) it is a dynamic, data-driven system in that signal-processing transforms are tailored to algorithm requirements and input signals, for reduced distortion and increased compression, and the system can be queried and steered during execution. Implementations are based on the best mathematical formulation of the problem coupled with automated selection of the best implementation among a space of alternatives, through the integration of models relating mathematical properties to implementationbehavior. Both hardware and software optimization are treated in a unified way. It is anticipated that with these methods the design-time will be decreased by two orders of magnitude or more, compared to implementations derived in a traditional way. Because the proposed system can explore a broad range of implementations that exceed the capabilities of a human designer, the implementations derived by the approach pursued in the project may even exhibit lower resource costs and higher performance. This research will provide a foundation for signal processing at all scales, providing key building blocks for engineers to build complex, distributed networks of adaptive signal processing sensors.