Research on color image-based multispectral reconstruction methods for candle flames.

A multispectral reconstruction method for flame color images based on K-means clustering and backpropagation neural networks (BPNN) is proposed to overcome the low spectral resolution of temperature measurement using color RGB three-band radiation images. A synchronized imaging system with an RGB camera and a 25-band multispectral camera was built to capture candle flame images. Image partitioning created a training set linking the three-band RGB and 25-band multi-spectral responses. Neural network training established a mapping between them. Spectral reconstruction of the candle flame images achieved an average relative error below 5%. The temperature inversion yielded an average error of 31.5 K, with a mean error of 1.79% in the error distribution, respectively, with test set R2 values of 0.97-0.99, confirming high model accuracy. This work demonstrates the feasibility of merging the spatial advantages of RGB images with the spectral advantages of multispectral data, offering a new approach for dynamic flame temperature field monitoring.