Patch‐based latent fingerprint recognition: A novel approach for reliable identification of partial prints

• Published in: Journal of forensic sciences
• Main Authors: Dhaneshwar, Ritika, Kaur, Mandeep, Kaur, Manvjeet, Gupta, Varun
• Format: Journal Article
• Language: English
• Published: United States 26.10.2025
• Subjects: generative adversarial networks; deep learning; fingerprint recognition; latent fingerprints; convolution neural network; synthetic images
• ISSN: 0022-1198; 1556-4029; 1556-4029
• DOI: 10.1111/1556-4029.70204

Latent fingerprints, that are imperative for forensic investigations, are seldom uplifted perfectly. These unintentional impressions left at crime sites are mostly partial with insufficient features that are not suitable for recognition. Further, the existing acquisition approaches rely on the single‐shot touch‐based capturing mechanism wherein the reagents are physically applied to the crucial evidence for examination. The current paper presents an Automated Patch‐based Latent Fingerprint Recognition System for reliable recognition based on partial samples. The experiments were conducted on the samples digitally captured using the touchless Reflected Ultra Violet Imaging System (RUVIS) equipment that can uplift multiple instances of evidence with high resolution. The proposed patch estimation algorithm identifies features to counter manual minutiae matching for estimating optimal patch size. Classical and Generative Adversarial Networks‐based augmentations were applied to simulate prints from a realistic crime site and deep feature extraction, respectively. The recognition capability of partial samples is then evaluated for different shallow and deep learning models, where the VGG16 and ResNet50 architectures outperformed. After fine‐tuning, the configured model achieved the maximum accuracy of 96% with ResNet50 as the backbone architecture and multiclass SVM as the subject classifier. Weighted average fusion further improved the accuracy by ~2%. The existing patch‐based recognition approaches cite accuracy between 68% and 84% on different benchmark data sets. However, the proposed model achieved an accuracy of 98% on the RUVIS data set and 96% when validated on the standard NISTSD27 data set, indicating better generalizability.