Abstract

Broadband Emission (BEMI) pyrometry is a low-cost technique for indirect characterization of soot temperature fields in laboratory flames using images captured with an RGB camera. However, retrieving temperature from color images through classical deconvolution techniques requires solving ill-posed inverse problems, producing results that are highly sensitive to signal noise and the choice of regularization parameters. This paper proposes using Convolutional Neural Networks (CNNs) to improve the accuracy of estimated 2D soot temperature fields from images of canonical axisymmetric laminar flames. Using a dataset of physically-grounded simulated images of temperature fields in the flame and their corresponding convoluted projections in the camera plane, we trained a CNN to learn the relationship between the reference temperature and the measured signals. Experiments over simulated and experimental images show that the trained CNN outperforms classical inversion methods when retrieving temperature from noisy images, especially in areas of interest such as the center of the flame. Resilience to noise makes CNNs attractive for implementing low-cost techniques for soot pyrometry using equipment of different quality.