Low-power and variation-aware approximate arithmetic units for Image Processing Applications

Abstract In applications such as image processing and machine learning, imprecision can be tolerated because of the nature of the application itself or the limitation of human senses. By using the approximate computation in parts of imprecision-tolerant applications, where the output quality can be slightly degraded, significant power, delay, or area reductions can be achieved. In this paper, three approximate full adders with reasonable accuracy, low power, and low delay are proposed. The effects of die-to-die (D2D) process variation on the threshold voltage of approximate full adders have been evaluated, and a method has been proposed to reduce the effects of variability. For evaluating the accuracy and the variability, these approximate full adders have been used and analyzed in the ripple carry adder structure and image Sharpening algorithm. In terms of power-delay-product (PDP), accuracy, and area for uniformly distributed inputs, one of the presented approximate full adders exhibits the best performance, and another one shows the best peak-signal-to-noise ratio (PSNR) for real images.