A non-overlapping image stitching method for reconstruction of page in ancient Chinese books

Abstract

Automatic stitching of page images in ancient Chinese books plays an important role in preservation and transmission of cultural heritage, significantly diminishing the need for manual intervention. Current methods accomplish image stitching based on their overlapping area and struggle with the ancient book pages without overlapping areas. To overcome this hurdle, this study proposes a novel deep learning based method to accurately stitch ancient book pages, which contains three key steps. Firstly, aiming to locate stitching seams precisely, a semantic segmentation model is exploited to predict the thickness masks of page images, and the non-overlapping pages can be stitched by cropping the thickness areas. Secondly, a novel multi-rule page stitching module with two creative page alignment methods is designed to align elements along the stitching seams. Lastly, the proposed method encompasses a self-assessment module, which judiciously selects the optimal stitched outcome from the multiple probable outputs generated by the multi-rule page stitching module. Experimental results demonstrate that the proposed method achieves superior performance in automatic stitching of ancient book pages. The stitching results on over 140 pages from three different ancient books show an accuracy of 82.18%, with 37.75% improvements over existing methods. This method provides a foundation for the automatic digitization of ancient Chinese books, showing significant potential applications in the field of automatic character recognition for historical documents.