Approximate Computing based on Latest-result Reuse for Image Edge Detection

This paper describes a memo-table-free approximate computing technique for an image edge detector to reuse the computed result only in the closest past. Decision of reuse is made based on tolerance-level controllable similarity checking for two consecutive 3x3 input pixels. We designed and implemented the proposed approximate edge detector in an FPGA and evaluated with 6 benchmark images. Results have demonstrated that by controlling the tolerance level the execution time is reduced by 15% and the energy dissipation is saved by 3-12% compared with the precise computation at acceptable quality in edge detected images.