Exploring the Impacts of Multiple Kernel Sizes of Gaussian Filters Combined to Approximate Computing in Canny Edge Detection

Abstract

Image processing applications are currently available on mobile devices, stressing the energy efficiency demands during the hardware design. In these applications, image edge detectors use filters as a preprocessing step to reduce undesirable artifacts and smooth the image. This work explores the combination of Multiplierless Multiple Constant Multiplication and Approximate Computing techniques on the Gaussian filter, investigating three different kernel size impacts on image processing. The impact of approximation is evaluated at different levels using the copy strategy technique on the LSBs of adders. The results show power and area reductions for the kernel sizes under evaluation. For instance, the approximate kernel $7\times 7$ kernel achieved reductions of up to 40% and 48% for the area and power consumption, respectively, compared to the exact version. It shows ample space for design exploration targeting different trade-offs of quality and power results.