교차 도메인 지문 인식을 위한 효과적인 데이터 증대법

Fingerprints are used as a means of personal identification because they are highly specific and persistent biometric data. These characteristics can be crucial clues in identifying suspects at crime scenes and establishing their connection to the crime. In the process of acquiring fingerprint images at crime scenes, various sensors, lighting conditions, etc. can create images with different characteristics. In this situation, a cross-domain fingerprint recognition problem may arise. Cross-domain fingerprint recognition problems are caused by the technical differences of various sensors and changes in the experimental environment, which reduce the accuracy of the fingerprint recognition model. In this paper, we propose an effective data augmentation method to solve this cross-domain fingerprint recognition problem. Experimental results show that the Sharpen augmentation method improves the accuracy by more than 2% when EfficientNet-B0 is used as a backbone, and the MedianBlur augmentation method improves the accuracy by more than 3% when AFR-Net is used as a backbone. The proposed augmentation technique will contribute to enhancing the capabilities of digital forensics by securing interoperability between heterogeneous fingerprint systems, improving recognition accuracy, enhancing the utilization of existing DBs, and expanding the scope of application.