Statistical Optimization of Image Processing Architectures for Maximum Resource Utilization and Throughput

Abstract

With the evolution of modern image processing languages, hybrid architectures and easily replieatable hardware architectural components, it becomes feasible to undertake practical steps toward realizing the long-desired goal of algorithm-driven architectures. The dimensional complexity of the prerequisite computations needed to accomplish global image processing architectural optimization surely remains formidable. However, significant enhancement of the computing-power-to-cost ratio of real-world image processor systems is feasible through the statistical analysis of an application algorithm or generic sets of algorithms so as to determine the optimum connectivity and quantities of a specified repertoire of functional elements at the levels of a processor, memory control, data paths, and input/output devices. A methodology for statistical image processor architecture optimization will be described. Examples of automated statistical optimization across particular generalized architectural forms will be presented. The importance of a generalized image algebra to the long-range potential for more general application-driven architectural optimization will be discussed.