Development of learning-based noise reduction and image reconstruction algorithm in two dimensional Rayleigh thermometry

Abstract

Consistent research efforts have been invested to deal with particles’ Mie scattering interference in Rayleigh measurements. This work reports an alternative learning-based image denoising and reconstruction algorithm to remove high value Mie scattering noise in Rayleigh thermometry. The algorithm, named denoising generative adversarial network (DNGAN), consists of a generator and a discriminator to learn the texture features and infer denoised image with a given noisy counterpart. Training data of the neural network come from artificially generated noise-free Rayleigh image superimposed with experimentally acquired Mie scattering image. Extensive evaluations are then conducted to prove the superiority of the proposed DNGAN network, compared with several other neural networks reported in literature. Quantitative results show that a convincing denoising with an overall reconstruction error of ~ 0.6% a peak signal-to-noise ratio of ~ 43 dB can be reached. Moreover, the pre-trained network is able to denoise image of the jet flame with a different Reynolds number without retraining the network parameters.