Classification of pigments using hyperspectral imaging

Abstract

Hyperspectral imaging (HSI) is used in art conservation to capture signal differences in wavelength not visible to the human eye. This work investigates the use of HSI in the range of 400nm to 1000nm, to segment an image of a painting into regions consisting of different paint characteristics and identify the paints within the region. Image segmentation algorithms often need to balance the trade-off between undersegmnetating the image, that is, grouping dissimilar regions, and over-segmentation, that is, fragmenting a region into separate regions. This dissertation proposes a Spectral Similarity Merging algorithm which prioritises homogeneity of regions, by keeping a low intra-class variance, while reducing over-segmentation. This was possible by modifying the Kernel Spectral Angle Mapper (KSAM) similarity metric into Kernel Spectral Correlation Mapper (KSCM) which enhanced the merging properties with regards to homogeneity. In fact, Spectral Similarity Merging (SSM) yielded an F-score of 48.5%, 39.8% higher than Simple Linear Iterative Clustering (SLIC), and an over-segmentation of 67.3%, 27.5% less than SLIC. Following the segmentation problem is the pigment classification method is carried out to identify the paints that constitute a painted region. This work proposes three methods, namely (i) a Hierarchical Paint Analysis (ii) a Global Non-negative Matrix Factorisation (NNMF) Paint Analysis, both of which were developed to attempt paint mixture identification and which are limited to two paints per mixture, and (iii) a Direct Classification of Paint Mixtures, which was developed to compare the performance of KSCM and NNMF when the mixtures are present in the reference library. The results obtained show that the Global NNMF Paint Analysis performs better than the Hierarchical Paint Analysis, producing an average accuracy of 76%. For the third method, The Direct Classification of Paints by KSCM yielded an accuracy of 84%, 14% more accurate than NNMF, meaning that KSCM is better when reference spectra are available, with the additional advantage of being less susceptible to spectral variance.