Denoising and fuel spray droplet detection from light-scattered images using deep learning

Abstract

A deep learning-based method for denoising and detecting the gas turbine engine spray droplets in the light-scattered image (Mie scattering) is proposed for the first time. A modified U-Net architecture is employed in the proposed method to denoise and regenerate the droplets. We have compared and validated the performance of the modified U-Net architecture with standard conventional neural networks (CNN) and modified ResNet architectures for denoising spray images from the Mie scattering experiment. The modified U-Net architecture performed better than the other two networks with significantly lower Mean Squared Error (MSE) on the validation dataset. The modified U-Net architecture also produced images with the highest Power Signal to Noise Ratio (PSNR) compared to the other two networks. This superior performance of the modified U-Net architecture is attributed to the encoder-decoder structure. During downsampling, as part of the encoder, only the most prominent features of the image are selectively retained by excluding any noise. This reconstruction of the noise-free features has produced a more accurate and better denoised image. The denoised images are then passed through a center predictor CNN to determine the location of the droplets with an average error of 1.4 pixels. The trained deep learning method for denoising and droplet center detection takes about 2.13 s on a single graphics processing unit (GPU). This study shows the promise for real-time processing of the experimental data using the well-optimized network.