Abstract
The psycho-visual nature of images and iterative nature of processing algorithms make vision and image processing suitable applications for approximate computing. State-of-the-art research in this area examines application resilience to approximation while assuming a uniform distribution for the information source. In this paper, we demonstrate that data-driven analysis can provide better insight into approximation requirements for image processing applications. Furthermore, this analysis is leveraged to design the multi-stage adaptive approximation control (m-SAAC) methodology that can save compute power by utilizing approximate computing, without compromising on image quality. The results demonstrate the efficacy of the proposed methodology for a variety of test cases.