MuralCut: Automatic character segmentation from mural images

Abstract

Segmenting characters from mural images is a crucial basic operation for other tasks whose common goal is to preserve the mural art, one of the important Thai cultural heritage. The problem is effective segmentation algorithms, which are used at present, are semi-automatic and general purpose; so users have to put heavy effort to get the satisfied result. Hence this paper proposes the automatic segmentation algorithm to segment characters from mural images automatically. The algorithm is divided into two main parts: automatic selection part and segmentation part. In automatic selection, we applied spectral residual to play a key role in selecting regions of interest, i.e., an object region and a background region to be inputs of the segmentation part. In segmentation part, an iterated graph-cut is used as the main mechanism of the segmentation in this work. Besides, In order to improve performance of the iterated graph-cuts, a superpixels algorithm is applied. Result of the algorithm from ordinary background images is 7.49% misclassified pixels with precision 73.02% and recall 94.64%.