PuzzleFlow

Abstract:

Jigsaw puzzles have been a popular recreational activity, challenging individuals to assemble various interlocking pieces to form a coherent image. With advancement in machine learning and computer vision, there has been growing interest in automating the puzzle solving process. This project “PuzzleFlow: Solving Jigsaw puzzles using flow based generative models” also presents a novel framework that utilizes flow-based generative model for solving spatial jigsaw puzzles. The proposed system will take a random arrangement of puzzle pieces as an input and reconstruct the original image by intelligently arranging the scattered puzzle pieces after comparison of compatibility between the puzzle patches. The proposed method involves two key stages. Firstly, a Flow-based model is trained to extract the knowledge from the underlying distribution of the puzzle’s image dataset. Then the model is responsible for generating realistic image samples from random noise vectors. Secondly, a puzzle assembly module that employs the trained model is introduced to predict the optimal placement of each puzzle pieces, ensuring a coherent and visually consistent final arrangement. Overall, the system will work by transforming the input arrangements density function into Gaussian density function, resembling the arrangement patches and then inverting the previous transformation. This project shall focus mainly on jigsaw puzzle pieces with identical squares.

Code Link:

Download Report