Parallel image understanding on a multi-DSP system

A course-grain multiprocessor architecture, based on an array of digital signal processors (DSPs), is presented to demonstrate the possibility of a parallel implementation of image-understanding algorithms. Aerial image understanding is investigated as an application. The system is designed to exploit temporal and spatial parallelism. A good speed-up was obtained for low- and intermediate-level operations. However, the speed-up for high-level operations was poor because of processor idle times, as the number of objects to be processed at higher level tends to be small. DSPs performed well as processing elements for number-crunching operations, but the performance was not so good for implementing symbolic operations.