Accelerating a modified Gaussian pyramid with a customized processor

Image pyramids are multi-scale representations of images, and their calculation is computationally intensive. They can be a main bottleneck in image processing and computer vision tasks such as edge detection and feature extraction. Thus, high speed computation of image pyramids is necessary. Moreover, when these algorithms are intended for embedded systems, other requirements such as area and energy consumption need to be satisfied. This paper presents a customized processor design to accelerate the execution of a stack of Gaussian low-pass filters. Using an instruction extension language, we added custom instructions to a 32-bit RISC-based configurable processor. We use three techniques to improve performance: operator fusion, single-instruction multiple-data vectorization and data reuse. The proposed processor achieves 12.3× speedup compared to the base processor, with 19% hardware overhead. The estimated improvement in energy consumption is 10.3×. The paper also presents the implementation results for the computation of a modified Gaussian pyramid in a tone mapping algorithm.