Accurate and Low-cost Fingerprint Classification via Transfer Learning

Fingerprint classification is highly required in reducing the search time and computational complexity of a 1:N fingerprint-based human identification problem. In this respect, the development of an accurate and low-cost fingerprint classification algorithm is highly desirable. To meet this demand, we propose a fingerprint classification algorithm, which can perform classification from raw fingerprint images. The proposed algorithm utilizes the characteristic of deep learning to handle all tasks in the classification pipeline, such as pre-processing, feature extraction, and classification. The low computational complexity of the proposed algorithm is ensured by using transfer learning (GoogLeNet) rather than training a deep learning architecture from scratch. The performance of the proposed algorithm is tested on the NIST-4 database. In the experimental part, the proposed algorithm achieves a better set of accuracy levels than other methods; 94.7% and 96.2% for the five-class and four-class classification problems, respectively. Given such a performance, it is desirable to adopt the proposed algorithm in practical applications.