Advancing 3D finger knuckle recognition via deep feature learning
Contactless 3D finger knuckle patterns have emerged as an effective biometric identifier due to its discriminativeness, visibility from a distance, and convenience. Recent research has developed a deep feature collaboration network which simultaneously incorporates intermediate features from deep neural networks with multiple scales. However, this approach results in a large feature dimension, and the trained classification layer is required for comparing probe samples, which limits the introduction of new classes. This paper advances this approach by investigating the possibility of learning a discriminative feature vector with the least possible dimension for representing 3D finger knuckle images. Experimental results are presented using a publicly available 3D finger knuckle images database with comparisons to popular deep learning architectures and the state-of-the-art 3D finger knuckle recognition methods. The proposed approach offers outperforming results in classification and identification tasks under the more practical feature comparison scenario, i.e., using the extracted deep feature instead of the trained classification layer for comparing probe samples. More importantly, this approach can offer 99% reduction in the size of feature templates, which is highly attractive for deploying biometric systems in the real world. Experiments are also performed using other two public biometric databases with similar patterns to ascertain the effectiveness and generalizability of our proposed approach.