Abstract: Aiming to the problem faced by the R matrix spectral decomposition method, a spectral decomposition method based on the camera response characteristics is proposed, a particle swarm optimization BP neural network model (PSOBP) is established for the inversion of the decomposed metameric black to realize the optimization of network training weights, and simulation experiments are conducted using the global training samples and local training samples quadratic spectral reconstruction method. The results show that under the D65 light source, using the PSOBP combined reconstruction method, the mean square error of the two test sets reconstructed by the RGB camera is reduced by at least 1.71% and 0.51%, respectively, compared with the other traditional method, and the maximum color difference is 3.5579 and 2.3776, basically meeting the requirements of the human eye color discrimination threshold. While the mean square error of spectral reconstruction accuracy of WorldView3 is less than 2% in bands of 410-510, 555-565, 590-685 and 705-740 nm, the proportion of acceptable samples represented by the fitness coefficient is 91.667%, and the maximum color difference is 1.6002 and 1.1177, respectively. In addition, the spectral reconstruction accuracy and chromaticity accuracy of the proposed method have been improved compared with other methods, and the 6-channel multi-spectral camera can meet the requirements of high precision spectral reconstruction.

Key words: remote sensing, spectral reconstruction, metamerism black, particle swarm optimization; neural network, homogeneous nonlinear extension