Robust Solvers for Square Jigsaw Puzzles

Abstract

A jigsaw puzzle solver reconstructs the original image from a given collection of non-overlapping image fragments using their color and shape information. In this paper we introduce new techniques for solving square jigsaw puzzles (with no prior knowledge of the initial image) that improves the accuracy of the state-of-the-art jigsaw puzzle solvers. While the current puzzle solving techniques are based on finding enhanced compatibility metrics across piece boundaries, we combine the existing techniques to achieve higher accuracy and robustness, i.e., our solver outperforms the known solvers even when the piece boundaries are imprecise. Unlike the most successful puzzle solvers that use greedy pairwise compatibility metrics among puzzle boundaries, we incorporate global information that enhances performance. As a step towards the future goal of developing an automated assembler for real-life corrupted image fragments or shredded documents, we examine puzzles that are corrupted by noise. Our proposed compatibility metrics shows robustness even in such scenarios.