Parameter Optimization of Approximate Image Processing Algorithms in FPGAs

Approximate computing has been proposed as a paradigm for contexts where resilience of applications to errors can be exploited, e.g. allowing to trade quality off for power/energy or hardware resources. Numerous approximation methodologies have been introduced in the literature and combining several of them can result in improved benefits. However, as approximation techniques require to be parametrized to control the loss of accuracy, using multiple ones implies to explore larger parameter sets. Furthermore, combined approximation methods can interact and influence the error propagation, adding to the optimization complexity. In this work, we propose an optimization model, targeted for a multi-objective genetic algorithm, to perform design space exploration simultaneously on all the parameters for each of the approximation techniques used in a system. We tailor the encoding and genetic operations for an image color processing application so that the genetic algorithm can converge properly to a Pareto front with good diversity. The optimization is carried out for trade-offs between image quality, FPGA hardware resource, and power. The results show that the proposed model can achieve the design space exploration and converge to a Pareto front that offers a wide range of trade-offs to choose from, while taking into account the potential interactions between the combined approximation techniques.