Limited Training Approach To Model Latent Fingerprint Data For Time-Constrained Solutions
Kiran Kumar Ravulakollu, Gottam Sai Nithin Reddy, Mayur Jadhav, Varshith Batti, Varshtih Kumar Vupputuri, Rohith Kumar Polishetty · 2025
The primary objective of this research work is to design a speedy and efficient methodology for latent fingerprint recognition that addresses the problem of long processing times inside the current approaches. The research here proves the potential for adopting one-shot and few-shot learning techniques in an accurate and computationally efficient form of latent fingerprint recognition. To this end, a lightweight architecture has been integrated with one-shot and few-shot learning techniques in the proposed methodology to recognize latent fingerprints. The research used the IIIT-D latent fingerprint dataset. The research, in particular, adopted the few-shot learning approach with the prototypical network trained on a pre-existing model, DenseNet121. With the few-shot learning of the prototypical network by using the pre-trained DenseNet121 model, the research produced an outstanding test accuracy at 91.66%. Both the F1 score and precision metrics are 93.32% and 93.93%, which reflects the good performance of the developed methodology in latent fingerprint recognition. This research introduces an effective approach for latent fingerprint recognition, properly balancing highly accurate results and computational expense. The proposed methodology enforces promising advances toward the precise identification and classification of latent fingerprints by incorporating one-shot and few-shot learning technique into a light-weight architecture.