A Study On Digital Oracle Bone Rejoining Based On Contour Matching

Oracle fragments are usually manually rejoined by experts,which is a very challenging task.The huge number of oracle bone fragments and the broken edges lead to inefficient manual rejoining and poor accuracy of the results,and existing methods usually rely on rule-based methods to match contour’s shapes and textures.To address these various problems,this paper provides an in-depth investigation of computer-aided techniques for oracle bone rejoining.Our paper proposes two works as follows: firstly,a complete process for oracle rejoining which is built into an interactive user system for oracle embellishment;Secondly,a 2D contour matching and search model based on deep learning which is derived based on the goal of oracle fragment rejoining.For the interactive user system,this paper designs a complete process and interactive UI for oracle bone rejoining.By integrating a series of processes such as contour extraction,matching,searching and fragment merging of oracle bone topologies,this paper adopts a rule-based heuristic matching algorithm,which matches contours by analyzing their similarity,curvature degree and external angles.With the evaluation of the proposed method,experiments are conducted on a real oracle dataset in this paper,and the results show that the matching algorithm for oracle rubbings is able to provide correct matching pairs over a large range and improve the matching efficiency for a large number of oracle bones.For the learning-based contour matching and searching model,we investigate a more generalized deep learning model on the matching and reassembly of fragments,but not only for the oracle fragments.The model combines the geometry and texture information of the connected contour of two fragments to efficiently search and match2 D fragment pairs.We first use patch-based neighborhood encoding to obtain the encoding of the geometry and color information of each contour point,and then we use a graph convolutional network to aggregate the information of neighboring points to extract the local contour features and local texture features of each contour point.After that,we directly concatenate these two features for feature fusion.Finally,for computing the matching scores,a similarity matrix is generated for each pair of fragments to represent the correspondence of contour points,and the Hough transform is used to detect the straight lines on the matrix and compute the matching scores.In this paper,a color image dataset with irregular curve segmentation is generated to simulate the fracture of oracle bone fragments.Experiments have been conducted on the generated dataset,and the results show that our proposed model achieves good accuracy both in pairwise matching and in global search for the correct fragment pairs.