Analysis Prediction Template Toolkit (APTT) for real-time image processing

Image processing applications often must not only provide accurate results but also meet real-time exigencies. This suggests a sensible division of labour, since in practice algorithmic designers are firmly wedded to workstations or PCs. Real-time acceleration in machine vision can be provided either by specialist hardware such as field-programmable gate arrays (FPGAs) or by parallel processing, neither of which are convenient for algorithmic development. Radar, vision, and some varieties of speech processing all commonly have a strong pipelined structure. The Analysis Prediction Template Toolkit (APTT) provides a seamless way to model graphically an image processing pipeline before purchase of the target hardware, and subsequently construct a parallel application from the developer's code, without significantly compromising algorithms. Testing on a handwritten postcode recognition application has confirmed agreement to within 10% between simulation and target system for pipeline traversal latency and throughput, at the same time allowing the designer to gain an intuitive feel for the behaviour of the application. Analytic results are available to refine the prediction, though the simulation is already suitable for cross-architectural comparisons of asynchronous pipelines.