As multicore processors are becoming commonplace, machine-vision software vendors are re-engineering their products, allowing their customers to leverage the power of multiprocessing. However, the task is not trivial since the programmer must optimize the power of image-processing functions across multiple processors. While the maximum expected performance increase of multiprocessing can be predicted to some extent by Amdahl’s law, the total performance increase will be limited by any sequential part of the program that cannot be optimized to run in parallel.
Many image-processing functions such as point processing or neighborhood operators such as convolution do lend themselves to optimization. However, other techniques—such as adaptive image processing—that require some type of decision-making process and rely heavily on branching conditions are not easily optimized.
