Siamese-Discriminant Deep Reinforcement Learning for Solving Jigsaw Puzzles with Large Eroded Gaps

Jigsaw puzzle solving has recently become an emerging research area. The developed techniques have been widely used in applications beyond solving. This paper focuses on Puzzles with Large Eroded Gaps (JPwLEG). We formulate the reassembly as a combinatorial optimization problem and propose Siamese-Discriminant Deep Reinforcement Learning (SD2RL) to solve it. A Q-network (DQN) is designed visually understand puzzles, which consists of two sets Siamese Discriminant Networks, one set perceive pairwise relations between vertical neighbors another for horizontal neighbors. proposed DQN considers not only evidence from incumbent fragment but also support its four trained using replay experience carefully rewards guide search sequence swaps reach correct solution. Two JPwLEG datasets are constructed evaluate method, experimental results show that SD2RL significantly outperforms state-of-the-art methods.