M. Bryan Wheeler and Dr. Gene Ware, Anthropology
In the Middle Ages, the cost of writing materials was so prohibitive that scholars often scraped the text from the parchment of a book and copied new writings over the old text. Centuries later, modern scholars often have more interest in the barely discernible, earlier texts. Multispectral approaches to image acquisition have shown promise in restoring the earlier texts. Presently, the primary need of multispectral analysis techniques is the development of effective post processing algorithms. Although many multispectral datasets have already been collected, there remains much work to be done in the image post processing realm.
Dr. Gene A. Ware and his colleagues have collected numerous multispectral images of archaeological artifacts. The dataset used in this project is a set of images taken of a Greek palimpsest housed in the Vatican Library. This dataset was obtained in conjunction with the Brigham Young University Institute for the Study and Preservation of Ancient Religious Texts (ISPART). Each of the 11 images of the palimpsest represents a band limited slice of the electromagnetic spectrum ranging from 400nm to 1000nm.
One of the primary advantages of multispectral imaging is the ability to classify the features of an image according to spectral characteristics. In research published by Ware et. al, Curtis Martin proposed using an adaptation of an unsupervised vector quantization algorithm to perform spectral classification of multispectral images [1]. Essentially, the algorithm groups pixels according to their spectral characteristics. The pixels in one group are more similar to all the pixels of that group than to the all pixels of any other group. For presentation purposes, each of these groups is assigned an arbitrary color and the results of the classification are displayed as an image.
The initial application of the LBG algorithm revealed that the spectral images suffered from misalignment which caused misclassification of pixels on the boundaries between features. In figure 1, the edges of the later text are misclassified into their own cluster. To overcome these limitations, subsequent analyses employed a mask that excluded these pixels from calculations. Although this masking technique provided better classification results (Figure 2), a more refined technique was needed to correctly classify the palimpsest’s features.
Kirkland noted that principle component analysis (PCA) is an effective method of feature discrimination in multispectral images [2]. PCA is a multivariate analysis technique used to project a dataset onto a set of orthogonal planes in a new data space. Because each principle component is orthogonal, most of the dataset’s variance is contained in the first few principle components. Higher order principle components are almost entirely noise and can be discarded from the analysis. By analyzing the variance of the dataset after PCA, it was found that the inherent dimensionality of the data set was reduced from the ten original images to three principle component images.
After applying PCA techniques to the dataset and removing the high order principle components, another round of unsupervised, spectral classification was performed. The results, presented in Figire 3, show a marked improvement over the results in figures. The background is uniformly classified and shown in dark blue and the effaced text is shown in light blue. Of particular interest are the yellow pixels in Figure 3. These regions are areas where ink bled through the palimpsest page. In Figures 1 and 2, these pixels are misclassified as part of the background or as erased text. In Figure 3, these pixels are correctly classified as their own cluster.
The ultimate thrust of this research was to create a representation of the data so that scholars could more easily interpret the erased text. The methods used were successful in making the original palimpsest text more readable. From Figure 3, one can easily discern the erased text represented in light blue. We can also see the smudging that occurred as the text was erased. In the future, I plan to extend these techniques to other regions of this palimpsest as well as other ancient documents.