|M.Sc Student||Gilad-Glickman Danit|
|Subject||Color Restoration of Scanned Archaeological Artifacts|
with Repetitive Patterns
|Department||Department of Electrical Engineering||Supervisors||Professor Lihi Zelnik-Manor|
|Professor Ilan Shimshoni|
|Full Thesis text|
In this work, we address the problem of virtually restoring archaeological artifacts. Virtual restoration is the process of creating a noise-free model of a degraded object, to visualize its original appearance. Our work focuses on restoring the coloring of the object. We considered both 2D and 3D objects, including scans of ancient texts and 3D models of decorated pottery.
Our denoising method exploits typical characteristic of archaeological artifacts, such as repetitive decoration motifs and a limited palette of colors. A special case of such characteristics are ancient texts, where a single color is commonly used, and reoccurring letters constitute as repetitions. Assuming a limited number of colors allows us to use classification of the model's colors, in order to revert the degradation process. Repetitive patterns are used to deduce the original coloring at damaged regions.
Our classification method is based on minimization of an energy function. Similar to previous methods, the energy function contains a data term and a smoothness term. But, in addition to those, we add a correspondence term, which encourages consistent labeling of matching regions. The new term can be considered a generalization of the smoothness term; while the original smoothness term imposes similarity on geometrically close pixels, the suggested correspondence term enforces similarity on semantically close pixels. The energy function is minimized using Graph-Cut algorithm.