Design space exploration for energy-efficient approximate Sobel filter

Abstract

Approximate computing (AC) is an emerging computing paradigm for energy efficiency. AC is most suitable for error-tolerant applications, e.g., image processing. The Sobel filter is an edge detector which is used heavily in image processing. One of the basic blocks in the hardware implementation of the Sobel filter is the full adder (FA), which approximation can greatly reduce the energy consumption of the filter. In this paper, we propose three new Non-exact FAs (NeFAs) that are suitable for image processing. The proposed NeFAs along with existing approximate FAs are used to create a library of approximate FAs. We use this library to perform a design space exploration (DSE) of the approximate Sobel filter, which is an essential step when searching for an optimized implementation. Experiments have shown that the executed DSE was able to achieve a target reduction of up to 75% in area and power. We analyzed the generated designs objectively and subjectively. Using the subjective assessment, we defined two Pareto optimal criterion where we found that the implementations based on the proposed NeFA are in the Pareto optimal for high target reduction, i.e., most efficient designs. Based on the objective assessment, we found that the NeFA-based designs achieve outstanding quality and produce finer edges than the exact design in some cases.