Parallel content-based image analysis on PIM processors

We study the problem of retrieving images from a large digital library, by searching the image contents with a specified pattern, using a low cost processing-in-memory (PIM) parallel processor architecture called Shamrock. We consider large databases with images that are rich in details, such as satellite and medical images. For the algorithmic part we focus on Fourier transforms and analyze the time complexities for performing the required operations on the Shamrock architecture. From the architectural point of view, we use a parallel processor architecture called Shamrock, which contains large amounts of memory and logic on the same chip, organized in a "mesh-like" structure. We conclude that content-based image processing, under some relaxed conditions, such as the existence of descriptors and special I/O organizations, can be done at high rate on a low cost PIM architecture, down to a single chip, by the year 2010.