All mineral pigments and protein binders used in this work were the same as traditional painting materials. The mineral pigments were chalk (No.A38, particle size: 10#) and azurite (No.A29, particle size: 10#) (Tianya Mineral Pigment, China), and the protein binders were animal glue (Beijing Dongxin Culture Co. Ltd., China), pure milk and egg white were collected from supermarket.

Three sets of samples were prepared according to old recipes^[7] and the analysis results of art samples (see Table 1). Although the FTIR spectra of different substrate materials of paintings are different, and the depth of penetration of ATR-FTIR is less than 1 μ m, which is much more than the depth of color layers, so the model of samples can be simplified, and all samples were prepared on 30 mm× 30 mm× 1 mm glass slides in this study. The first set of samples, three groups of single protein binder samples without pigment were obtained by drying each protein binder at room temperature. The second set of samples contained different amounts of chalk in protein binders. For example, the chalk-milk samples were prepared by mixing chalk with milk at the chalk-milk ratio from 1% to 100% with an increment of 10%, the chalk-animal glue samples and chalk-milk samples were prepared in a similar mode. The last set of samples containing different amount of azurite in protein binders were prepared in a similar mode with the former two sets.

Table 1

Table 1

Table 1 Formulations of three sets of samples simple pigment protein binder chalk /g azurite /g milk /g animal glue/g egg white/g deionized water/g single protein binder milk 1 100 animal glue 1 100 egg white 100 chalk-protein binder chalk-milk binder 0.01~1 1 100 chalk-animal glue binder 0.01~1 1 100 chalk-egg white binder 0.01~1 1 100 azurite-protein binder azurite-milk binder 0.01~1 1 100 azurite-animal glue binder 0.01~1 1 100 azurite-egg white binder 0.01~1 1 100

Table 1 Formulations of three sets of samples

Other same three sets of fresh samples were exposed to UV light to provide artificial art samples.The accelerated UV light aging process was carried out by irradiation with light from a UV lamp emitting over 310 nm at 765 W· m^-1. FTIR spectra were registered after 0 hand 888 h of exposure.

The FT-IR spectra of all samples were collected by a Nexus 670 FTIR spectrometer equipped with an attenuated total infrared (ATR) assembly with a ZnSe crystal supplied by Thermo-Nicolet (MA, USA). All spectra were the average of 32 scans in the range of 4 000~650 cm^-1 with a resolution of 4 cm^-1. During the experiment, the influence of water vapor and carbon dioxide was subtracted in situ by the spectrometer.

The software OPUS 7.2 (Bruker, DEU) was used to process the spectra. The ordinate of all spectra was transformed into absorbance before the automatic baseline correction. All spectra were normalized and then smoothed with a window size of 13 points. The SD-IR was calculated with a window size of 13 points and used to establish ATR-FTIR spectral pattern recognition models for the protein binders in color layers.

PCA is a convenient and common tool to establish ATR-FTIR spectral pattern recognition models. In this research, PCA of the ATR-FTIR spectra was achieved using the PLS_toolbox 6.2.1 (Eigenvector Research Inc, USA) and Matlab (Mathworks Inc, USA). In order to reduce the interference of UV light and improve the applicability of pattern recognition model, the SD-IR spectral of the all samples in the region of 1 700~1 000 cm^-1 were analyzed by PCA, and SD-IR spectra were used after the standard normal variate (SNV) and mean-centered.

Fig.1 shows the ATR-FTIR spectra of fresh and artificial art samples. As is shown in Fig.1, there are two significant peaks in the regions of 1 644~1 638 and 1 548~1 540 cm^-1 assigned to the amide Ⅰ and Ⅱ bands of proteins^{[8, 9, 10]}, which means these three kinds of fresh samples contain plenty of proteins. The significant peaks of three kinds of protein samples are so similar that it’ s difficult to identify protein species in painting materials. As is shown in Fig.2, the amide Ⅰ and Ⅱ bands in the SD-IR spectra may be separated into individual peaks, which correspond to the α -helixes, β -sheets, β -turns and γ -randoms of secondary structures of the protein.

Fig.1

	Figure Option ViewDownloadNew Window
	Fig.1 ATR-FTIR spectra of fresh and artificial art single protein binder samples (a): Milk; (b): Animal glue; (c): Egg white

Fig.2

	Figure Option ViewDownloadNew Window
	Fig.2 The SD-IR spectra of fresh single protein binder samples. Each spectrum is the mean of all samples of the same class

The asymmetrical and symmetrical stretching bands of CH₂ near 2 920 and 2 852 cm^-1, the stretching band of C=O near 1 740 cm^-1, in conjunction with the stretching bands of C— O near 1 255 and 1 171 cm^-1, indicate the existence of aliphatic acids and esters in milk binder samples. The asymmetrical and symmetrical stretching bands of CH₂ near 2 923 and 2 853 cm^-1, in conjunction with the stretching band of C=O near 1 746 cm^-1, indicate the existence of aliphatic acids and esters in animal glue binder samples. Due to the degreasing of treatment of animal glue, the relative absorbance of significant peaks of the stretching band of C=O in milk binder samples is sharper and stronger than that in animal glue. According to the relative absorbance of significant peaks, there are more aliphatic acids and esters in milk binder samples, and no aliphatic acids and esters in egg white binder samples. The relative absorbance of significant peaks of aliphatic acids and esters is one of the most obvious differences in three single protein binder samples.

The stretching bands of C— O in the region of 1 200~1 000 cm^-1 is another obvious difference in the fresh single protein binder samples. The stretching band of O— H near 3 298 cm^-1, and the stretching bands of C— O near 1 091, 1 072 and 1 035 cm^-1, indicate the existence of lactose in milk binder samples^[11]. The skeletal band of C— C near 1 085 cm^{-1 [12]}, indicate that there is no saccharide in animal glue. The stretching bands of O— H near 3 285 cm^-1, in conjunction with the stretching bands of C— O near 1 076 and 1 156 cm^-1, indicate the existence of galactose in egg white binder samples. According to the difference of significant peaks of saccharides, milk binder samples can be identified from three protein binder samples. But the peaks of animal glue binder samples and egg white binder samples in the regions of 1 200~1 000 cm^-1 are quite similar so that it’ s difficult to distinguish them.

As is shown in Fig.1, in fresh classes and artificial art classes, respectively, both the original spectra of animal glue binder samples and egg white binder samples are quite similar in intra-class. However, due to the uneven spatial distribution of the single protein binder samples, the spectra of the same milk samples are different. Moreover, the spectra of the animal glue samples and the egg white samples both provide a visualization of separation in the region of 1 800~1 000 cm^-1 between the artificial art samples and fresh samples. But in the same region, there is no obvious difference in the spectra of the milk samples. The reason is that the intra-class variances caused by uneven spatial distribution are much higher than that caused by the aging of milk.

Fig.3 shows the ATR-FTIR spectra of fresh samples with different content of pigment. The ν ₃ antisymmetric stretching band of C O32-near 1 403 m^-1, the ν ₂ out-of-plant bending band of C O32-near 874 cm^-1, and the ν ₄ in-plant bending band of C O32-near 711 cm^-1, are the significant peaks of chalk^[13]. As the chalk-protein binder ratio increases, the relative absorbance of significant peaks of chalk near 1 403, 874 and 711 cm^-1 also increases, but the relative absorbance of significant peaks of protein binder decreases, and the signifi-cant peaks of protein binders are gradually overlapped. The bending band of O— H near 951 cm^-1, the asymmetric stretching band of C O32-near 1 410 cm^-1, the out-of-plant band of C O32-near 827 cm^-1, and the in-plant band of C O32-near 769 cm^-1, are the significant peaks of azurite. As the azurite-protein binder ratio increases, the relative absorbance of significant peaks of azurite near 1 410, 827 and 769 cm^-1 also increases, but the relative absorbance of significant peaks of protein binder samples decreases, and the significant peaks are gradually overlapped.

Fig.3

	Figure Option ViewDownloadNew Window
	Fig.3 ATR-FTIR spectra of fresh samples as the content of pigment increased in a range of 0.1%~100% with an increment of 10% (a1): Chalk-milk binder; (b1): Chalk-animal glue binder; (c1): Chalk-egg white binder; (a2): Azurite-milk binder; (b2): Azurite-animal glue binder; (c2): Azurite-egg white binder

Fig.4 shows the ATR-FTIR spectra of artificial art samples with different contents of pigment. Since the intra-class variances caused by the irregular pigments content are much higher than the intra-class variances caused by the aging degree in the same class. In the same region of 1 800~1 000 cm^-1, there is no obvious difference between the spectra of fresh samples and artificial art samples.

Fig.4

	Figure Option ViewDownloadNew Window
	Fig.4 ATR-FTIR spectra of artificial art samples as the content of chalk increased in a range of 0.1%~100% with an increment of 10% (a1): Chalk-milk binder; (b1): Chalk-animal glue binder; (c1): Chalk-egg white binder; (a2): Azurite-milk binder; (b2): Azurite-animal glue binder; (c2): Azurite-egg white binder

In the region of 4 000~3 000 cm^-1, there is obvious difference between the spectra of fresh samples and artificial art samples in Fig.1. On the other hand, in the region of 3 000~1 700 cm^-1, there are the absorption peaks of fats in some art samples. So the SD-IR spectral of the artificial art samples in the region of 1 700~1 000 cm^-1 were analyzed by using PCA with the SD-IR spectral pattern recognition models established by the SD-IR spectra of the fresh samples. The clustering is evident in the score plot, and the loading plot gives information on what absorption bands or wavenumber mostly contribute to the separation of all fresh samples according to the PC1 and PC2 values.

Fig.5 shows the PCA results of SD-IR spectra of fresh and artificial art single protein binder samples. Fig.5(a) shows the scores plot in the plane of the first and second principal components (PCs), which together account for 63.3% of the total variability, and Fig.5(b) shows the loading s plot using PC1 and PC2 versus the variable number.

Fig.5

	Figure Option ViewDownloadNew Window
	Fig.5 PCA results of SD-IR spectra of fresh and artificial art single protein binder samples (a): Score plot of the first two PCs of the spectra; (b): Loadings plot of the first two PCs of the spectra

As is shown in Fig.5(a), the plot of PC1 versus PC2 provides a better visualization of separation among all fresh samples of single protein binder allowing the qualitative recognition of three classes by using SD-IR spectra than the original spectra. But there is significant difference between the spectra of fresh and artificial art single protein binder samples, so that the artificial art samples and fresh samples in the same class, such as the animal glue binder samples and egg white samples, can be classified into two different groups. In milk binder samples, since the intra-class variances caused by spatial distribution are much higher than that caused by UV light ageing, the fresh samples and artificial art samples can be classified into the same group.

As is shown in Fig.5(b), most variables with high loading coefficients for the PC1 are close to 1 633, 1 533, 1 190, 1 133, 1 105, 1 080 and 1 052 cm^-1, which means that the discriminated compounds are the secondary structures of the proteinand saccharides. On the other hand, most variables with high loading coefficients for PC2 are close to 1 685, 1 646, 1 583 and 1 512 cm^-1, which means that the discriminated compounds are the secondary structures of the protein.

Fig.6 shows the PCA results of the SD-IR spectra of fresh samples with different chalk-protein binder ratios. Fig.6(a) shows the scores plot in the plane of the first and second principal components (PCs), which together account for 61.2% of the total variability, and Fig.6(b) shows the loading’ s plot using PC1 and PC2 versus the variable number.

Fig.6

	Figure Option ViewDownloadNew Window
	Fig.6 PCA results of SD-IRspectra of fresh and artificial art samples with different chalk-protein binder ratio (a): Score plot of the first two PCs of the spectra; (b): Loadings plot of the first two PCs of the spectra

As is shown in Fig.6(a), the plot of PC1 versus PC2 provides a better visualization of separation among all fresh and artificial art samples of chalk-protein binder allowing the qualitative recognition of three classes by SD-IR spectra than the original spectra. Compared with the score plots of single protein binder samples in Fig.5(a), the artificial art samples and fresh samples in the same class can be classified into the same groups in Fig.6(a). The reason is that the intra-class variances caused by different chalk contents and spatial distribution are much higher than that caused by UV light aging.

As is shown in Fig.6(b), most variables with high loading coefficients for the PC1 are close to 1 194, 1 134, 1 084, 1 057 and 1 011 cm^-1, which means that the discriminated compounds are saccharides in the PC1. On the other hand, most variables with high loading coefficients for PC2 are close to 1 633 and 1 548 cm^-1, which means the discriminating compounds are the secondary structures of the protein in the PC2.

Fig.7 shows the PCA results of SD-IR spectra of fresh samples with different azurite-protein binder ratios. Fig.7(a) shows the scores plot in the plane of the first and second principal components (PCs), which together account for 64.0% of the total variability, and Fig.7(b) shows the loading’ s plot using PC1 and PC2 versus the variable number.

Fig.7

	Figure Option ViewDownloadNew Window
	Fig.7 PCA results of SD-IRspectra of fresh and artificial art samples with different azurite-protein binder ratio (a): Scores plot of the first two PCs of the spectra; (b): Loadings plot of the first two PCs of the spectra

As is shown in Fig.7(a), the plot of PC1 versus PC2 provides a better visualization of separation among all the fresh samples of azurite-protein binder allowing the qualitative recognition of three classes by SD-IR spectra than the original spectra. Compared with the scores plot of single protein binder samples in Fig.7(a), the artificial art samples and fresh samples in the same class can be classified into the same groups in Fig.7(a). The reason is that the intra-class variances caused by different azurite content and spatial distribution are much higher than that caused by UV light aging.

As is shown in Fig.7(b), most variables with high loading coefficients for the PC1 are close to 1 633, 1 527, 1 124, 1 109, 1 082 and 1 053 cm^-1, which means the discriminating compounds are the secondary structures of the protein and saccharides in the PC1. On the other hand, most variables with high loading coefficients for PC2 are close to 1 653 and 1 527 cm^-1, which means the discriminating compounds are the secondary structures of the protein in the PC2.

According to the result of clustering of fresh samples by using PCA, PCA can finds out the critical variables from large amounts of data in SD-IR spectral to discriminate the samples and review the classification.

In original FTIR spectra, uneven spatial distribution and irregular content of pigments are unfavorable to identify protein in color layers. But in PCA, these difficulties will help identify protein binders in the artificial art samples using the pattern recognition models established by fresh samples, so considerable time will be saved in protein binder identification.The pattern recognition models with general applicability and high reliability can be established since the inter-classes variance and intra-class varianceare both high. The high inter-class variance contributes to discriminating the samples in different class, and the high intra-class variance contributes to improving the applicability of models. In this study, due to the fact that inter-class variances are much higher than the intra-class variances in all models, three protein binders can be identified under different conditions, and because the intra-class variances caused by different pigments content and spatial distribution are much higher than that caused by UV light aging, the artificial art protein binders can be identified by the pattern recognition models established by the fresh samples in Fig.6(a) and Fig.7(a).