On the Design of Approximate Sobel Filter

Abstract

Approximate computing (AC) is an emerging computing paradigm for energy efficiency. Typically, AC is implemented at the primary arithmetic level, e.g., addition, multiplication, and division, and its performance is evaluated by integration within an application. However, the achieved design efficiency may not be satisfactory. Therefore, for a specific approximate application, we need to study the most suitable settings of its basic approximate component. In this paper, we investigate several approximate designs of the Sobel filter, which is used for image edge detection. We consider different target designs, e.g., for 25% area reduction, we determine the various types of the used full adders and the number of components for each type. For an approximate Sobel filter with 15% to 55% area and power reduction compared to the exact design, we determine the settings for each target design. The obtained Sobel designs are evaluated for different benchmark images, i.e., Cameraman, Lena, and Bikesgray, and show a highly acceptable quality for edge detection. The average multiscale structural similarity (MSSSIM) index for all evaluated designs on the three benchmark images was 0.73.