Evaluation of Multi-core Architectures for Image Processing Algorithms

Diverse application scenarios and real-time constraints on computer-vision applications have motivated numerous explorations of computer architectures that provide more efficiency through hardware scalability by exploiting the characteristics of image processing and computer vision algorithms. The growing and computational power and programmability of the of multi-core architectures provide great prospects for acceleration of image processing and computer vision algorithms which can be parallelized. This thesis undertakes a novel study to find unique attributes of three widely used algorithms in computer vision, and identifies computer architecture(s) best suited for each algorithm. Significant acceleration over standard CPU implementations is obtained by exploiting data, thread and instruction parallelism provided by modern programmable graphics hardware. We test the following architectures most used for graphics and imaging applications: Intel Pentium 4 HT, Intel Core 2 Duo, NVidia 8 Series GPU and Sony PlayStation3 (PS3) CellBE. Additionally, we have optimized two image processing and computer vision algorithms, namely Canny edge detection and KLT tracking for the PS3. The architectures’ capabilities of handling three image processing algorithms of varying complexity were evaluated over standard inputs. The results are then tabulated for comparison.