2D Irregular Fragment Reassembly with Deep Learning Assistance

Fragment reassembly is widely used in fields such as archaeology and forensics. This paper introduces an algorithm for reassembling irregular fragments, enabling the reconstruction of arbitrarily segmented irregular fragments without any prerequisites. The reassembly procedure encompasses feature extraction, local pairwise matching, and global composition. We design a classification network to assess the compatibility of fragment pairings. In order to alleviate the impact of unavoidable errors in the local matching phase, we put forth an error assessment criterion calculated from the contour differences at the junctions, a two-stage reassembly strategy involving both initial and candidate phases, and a region completeness evaluation standard based on the size of the background region to gauge the ultimate composition outcomes. Additionally, due to the limited availability of publicly shared datasets at present, we created a dataset comprising 1000 sets of randomly divided irregular fragments for training and testing. The experimental results show the excellent performance of this algorithm in terms of accuracy and completeness in fragment reassembly for both public and self-constructed datasets.