Solving Jigsaw Puzzles Using Variational Autoencoders

Machine learning has recently occupied a remarkable position due to the ability of engagement in various systems and applications. As a result, a significant effort has been directed to enhance the existing techniques and present a critical assessment of applying these techniques in different applications. However, there is a wide room of improvements, especially for solving common problems using simpler architectures such as variational autoencoders. The simplicity of variational autoencoders makes it suitable for wide range of applications and systems. Among these applications is solving Jigsaw problem, which is based on reconstructing images from shuffled image tiles. Many articles addressed the Jigsaw problem and proposed different solutions, however the presented techniques suffer from high complexity and long training time. In this work, we explore the use of variational autoencoders to learn high level features to reconstruct images from shuffled image tiles. We also explore the use of the learnt features in transfer learning to adapt the trained model to other tasks such as classification or detection. To the best of the authors knowledge, this type of problems has never been addressed using variational autoencoders technique before. We obtained around the bar results with the more complex CNN-based models.