Image segmentation and pigment mapping of cultural heritage based on spectral imaging

The goal of the work reported in this dissertation is to develop methods for image segmentation and pigment mapping of paintings based on spectral imaging. To reach this goal it is necessary to achieve sufficient spectral and colorimetric accuracies of both the spectral imaging system and pigment mapping. The output is a series of spatial distributions of pigments (or pigment maps) composing a painting. With these pigment maps, the change of the color appearance of the painting can be simulated when the optical properties of one or more pigments are altered. These pigment maps will also be beneficial for enriching the historical knowledge of the painting and aiding conservators in determining the best course for retouching damaged areas of the painting when metamerism is a factor. First, a new spectral reconstruction algorithm was developed based on Wyszecki’s hypothesis and the matrix R theory developed by Cohen and Kappauf. The method achieved both high spectral and colorimetric accuracies for a certain combination of illuminant and observer. The method was successfully tested with a practical spectral imaging system that included a traditional color-filter-array camera coupled with two optimized filters, developed in the Munsell Color Science Laboratory. The spectral imaging system was used to image test paintings, and the method was used to retrieve spectral reflectance factors for these paintings. Next, pigment mapping methods were brought forth, and these methods were based on Kubelka-Munk (K-M) turbid media theory that can predict spectral reflectance factor for a specimen from the optical properties of the specimen’s constituent pigments. The K-M theory has achieved practical success for opaque materials by reduction in mathematical complexity and elimination of controlling thickness. The use of the general K-M theory for the translucent samples was extensively studied, including determination of optical properties of pigments as functions of film thickness, and prediction of spectral reflectance factor of a specimen by selecting the right pigment combination. After that, an investigation was carried out to evaluate the impact of opacity and layer configuration of a specimen on pigment mapping. The conclusions were drawn from the comparisons of prediction accuracies of pigment mapping between opaque and translucent assumption, and between single and bi-layer assumptions. Finally, spectral imaging and pigment mapping were applied to three paintings. Large images were first partitioned into several small images, and each small image was segmented into different clusters based on either an unsupervised or supervised classification method. For each cluster, pigment mapping was done pixel-wise with a limited number of pigments, or with a limited number of pixels and then extended to other pixels based on a similarity calculation. For the masterpiece The Starry Night, these pigment maps can provide historical knowledge about the painting, aid conservators for inpainting damaged areas, and digitally rejuvenate the original color appearance of the painting (e.g. when the lead white was not noticeably darkened)