A Self-Supervised Deep Metric Learning Approach for Jigsaw Puzzle Reconstruction

Abstract

In this study, we focus on solving jigsaw puzzles and introduce a novel approach using self-supervised deep metric learning to analyze the adjacency relationships between puzzle tiles and arrange them in the correct order. Our methodology involves constructing a Siamese Neural Network (SNN) and exploring various configurations to capture the compatibility between image tiles. Initially, we treat the task as a supervised learning problem to identify the optimal configuration for our model. Subsequently, we leverage self-supervised learning, a subtype of unsupervised learning, to enhance the model's capability without the need for labeled data. Our objective is to train the network exclusively on the particular puzzle we aim to solve. This approach allows the network to grasp the intrinsic information from the specific problem. Finally, we contrast two methods: Relaxation Labeling (ReLab) and Puzzle Solving by Quadratic Programming (PSQP), and assess the performance of our model by testing it against some of the most effective hand-crafted compatibility metrics designed for puzzle solving. These evaluations are conducted on publicly available datasets, demonstrating the practicality and effectiveness of our proposed methodology.