Knowledge Based Lacunas Detection and Segmentation for Ancient Paintings

Abstract

Lacunas are a common form of the damage that can occur to paintings and more often to murals. Taking Dunhuang murals as research background, a new algorithm to detect and segment the lacuna area from mural images is proposed, which consists of a training phase and a runtime phase. In the training phase, a Bayesian classifier is trained. At runtime, the Bayesian classifier is first applied to perform the rough lacuna regions detection. Then, a graph representing the mural image is built with output of the Bayesian classifier. The domain knowledge of murals is incorporated into the graph in this step. At last, the image segmentation using graph cut is done based on the minimal cut/maximal flow algorithm. The outputs of the image segmentation are lacuna regions and background regions. About 250 high resolution Dunhuang mural images are collected to test the proposed method's performance. Experimental results have demonstrated its validity under certain variations. This research has the potential to provide a computer aided tool for mural protectors to restore damage mural paintings.