Fig. 1. Basic structure of CNN

Fig. 2. Convolution process

Fig. 3. Designed CNN model

Fig. 4. Principles of dropout. (a) Network without dropout; (b) network with dropout

Fig. 5. Multispectral images of the pigment true silver

Fig. 6. Flow chart of spectral feature reorganization

Fig. 7. Flow chart of classification experiment for mural pigments

Fig. 8. Multispectral images of standard mural paint board

Fig. 9. Multispectral images of simulated mural

Fig. 10. Sample units after spectral feature recombination of standard pigment

Fig. 11. Classification renderings of different models. (a) Original mural; (b) statistical manifold-SVM model; (c) CNN model (without dropout); (d) CNN model (with dropout)