Model-based Hardware Design for Image Processing Systems

Abstract

Model-based design has been touted as the most viable design methodology of the future for the design of embedded hardware/software systems. Due to the large complexity of modern embedded systems, it is more and more error-prone to design systems without having a formal model to support and verify the application at design time. Also, formal models generally capture broad classes of applications, and thus any innovation on a modeling technique has the potential to enhance every individual application in the associated class. Often, a formal model captures the high-level abstraction of an application, which is lost in the final implementation, and thus modeling gives an effective platform to perform high-level design optimizations. Dataflow graphs have been widely used as formal models in the signal processing domain for a long time, and various commercial tools have adopted dataflow semantics for model-based design methodology.

In this thesis, we develop a new dataflow meta-modeling technique, called homogeneous parameterized dataflow (HPDF). HPDF is a meta-modeling technique in that it can be applied to a variety of underlying dataflow models of computation to enhance their expressive power, while maintaining much of the useful structure of the underlying models. HPDF addresses an important range of applications, especially in the image processing domain. We present various properties and capabilities of HPDF, including the notions of repetitions vector, valid schedule, derivation of looped schedules, single-rate equivalent graphs, and HPDF graph transformation methods. We also give three in-depth examples of complex systems that we have studied to demonstrate the capabilities of HPDF -- a gesture recognition application, an image registration application, and a gait-DNA application. For hardware implementation, we target our applications onto Xilinx and Altera field programmable gate arrays (FPGAs), and we present results from the hardware mapping of the gesture recognition and the image registration application.

To build a foundation for further broadening the impact of HPDF modeling, we present initial work on applying cyclo-static dataflow as an intermediate representation for mapping MATLAB programs into hardware implementations. Because of the compatibility between cyclo-static dataflow and the HPDF meta-modeling approach, which we demonstrate in Chapter 3 of this thesis, this is an important first step to exploiting HPDF techniques in the context of MATLAB-to-hardware synthesis. In particular, we focus on relating cyclo-static dataflow to Compaan process networks, which is a variant of the Kahn process network model of computation that has been shown to be useful in representing concurrency in MATLAB programs.

In summary, this thesis develops a useful new meta-modeling approach for implementing an important class of image processing applications, and develops and extensively demonstrates a methodology for efficient hardware implementation from representations in the proposed new meta-model.