Comparison of Approximate Computing with Sobel Edge Detection

Abstract

In this study, a Sobel filter with approximate computing was used to detect the edge of an image and compare the quality of the output images and error metrics of various approximate adders. Approximate computing is performed by applying the logic of various approximate adders to the lower-bit part of the addition operation of the Sobel filter. The image quality of the edge detection result and error metrics are analyzed by varying approximating bits. The approximate adders used include an Approximate Mirror Adder (AMA), Approximate XOR/XNOR-based Adder (AXA), and Inexact Adder (InXA). A simulation was performed in Quartus using Verilog HDL. The precise adder and the approximate adder logic were applied to the low part, synthesized, and compared. The two filters were compared through various error metrics for approximate computing and logic utilization indicating how much of the logic module has been used. In the simulation, the Sobel filter with the AMA4 showed the best error metrics and performance logic among the 10 types of approximate adders. The results show that approximate adder logic with the appropriate approximate bits can be applied to an edge-detecting filter and has an acceptable error and image quality.