A Novel Fingerprint Recognition Strategy Driven by Convolutional Autoencoders and Attention Mechanisms

This paper investigates a fingerprint recognition method based on Convolutional Autoencoder, aimed at integrating the advantages of fingerprint ridge and minutiae information to enhance the efficiency and accuracy of fingerprint feature extraction and matching. To achieve this, we propose a novel fixed-length local fingerprint embedding extraction method. This method is modified based on the Convolutional Autoencoder framework and incorporates an attention mechanism, significantly boosting the model's ability to capture latent information and detailed features beneficial for fingerprint matching. Furthermore, we employ fingerprint masks to refine local descriptors, thereby mitigating the adverse impact of nonoverlapping fingerprint regions on matching. Through experiments conducted on multiple datasets, our method surpasses the performance of the currently advanced fingerprint recognition system AFRNet, excelling particularly in handling low-quality and highly similar fingerprints. Additionally, the method demonstrates remarkable performance in fingerprint reconstruction tasks, providing a new solution for enhancing low- quality fingerprints. Finally, the paper concludes with a summary of the research findings and an outlook on future research directions.