Denoising neural network for low-light imaging of acoustically coupled combustion

Abstract

This study investigates the use of a trained neural network (NN) to enable more efficient noise reduction in processing images associated with acoustically coupled combustion phenomena, as compared with more commonly used image processing techniques. The approach is applied to experiments involving high-speed imaging of a single and coaxial methane–air jet diffusion flames exposed to various acoustically resonant environments. Proper orthogonal decomposition (POD) analysis applied to the flame imaging may be used to capture characteristic signatures in the flame dynamics and in verification of the proposed approach in this investigation. The NN trains on low-exposure input images and high-exposure response images for a steadily burning fuel jet with no coaxial flow, yet is remarkably successful when applied to a range of coaxial flow and acoustic excitation conditions. The proposed neural network approach demonstrates a significant decrease in the preprocess time required in analyzing flame images, typically by over a factor of 5, and preserves image quality. The approach replicates POD-based flame dynamics very well, for both low-amplitude and high-amplitude flame responses, the latter involving transitions in the dynamics due to the introduction of multiple timescales. The relative simplicity and success of this NN approach thus show the potential for improved image processing for complex dynamical flows and their transitional features.