Voltage over Scaling for Approximate Newton Method in Total Variation-based Image Denoising

Approximate computing is an energy efficient technique that allows acceptable errors to improve circuits' energy and latency performance in image processing. This paper presents a runtime-based approximate integrated method by applying the voltage over scaling (VOS) technique to the approximate Newton method for total variation (TV)-based image denoising. The proposed technique reduces the energy dissipation of the approximate Newton method by decreasing the supply voltage of adder cells. The output delay thus increases beyond the worst-case delay, thereby causing errors in the output; the number of output errors varies without changing the circuit structure. The results illustrate that the output image quality is kept constant when the number of approximate bits (NAB) is smaller than 15 for both exact and approximate cell using VOS; meanwhile, the energy consumption decreases gradually. When the NAB value is high at around NAB=17 with 50% noise, the exact full adder shows better performance than the approximate cell (AMA1) when the approximation is low (less than five errors); the maximum number of iterations and energy consumption for AMA1 (16.2 and 14.33nJ) is larger than for exact full adder (8.6 and 7.26nJ).