Fusion2print Achieves 0.999 Accuracy in Contactless Fingerprint Matching with Novel Flash-Non-Flash Fusion

Contactless fingerprint recognition is gaining prominence as a more hygienic and user-friendly alternative to traditional methods, yet capturing clear images remains a significant challenge. Roja Sahoo and Anoop Namboodiri, from the International Institute of Information Technology Hyderabad, alongside their colleagues, address this issue in their new research. Their work introduces Fusion2Print (F2P), a novel framework designed to systematically combine the strengths of both flash and non-flash contactless fingerprint captures. By fusing these modalities, the researchers enhance ridge clarity and achieve a substantial improvement in recognition performance, ultimately paving the way for more reliable and accessible biometric authentication systems. The team demonstrates F2P’s effectiveness with a new dataset and achieves state-of-the-art results, surpassing existing contactless fingerprint verification methods.

Contactless Background

Contactless fingerprint recognition provides a hygienic and convenient alternative to traditional methods, enabling rapid acquisition without the issues of latent prints, pressure artifacts, or hygiene concerns. However, contactless images often suffer from degraded ridge clarity due to factors like illumination variation, skin discoloration, and reflections. Balancing flash captures, which preserve ridge detail but introduce noise, with non-flash captures, which reduce noise but lower contrast, presents a significant challenge. This research introduces Fusion2Print (F2P), a novel approach designed to address these issues and improve the accuracy of contactless fingerprint identification.

The primary objective of this work is to create a unified embedding space for both contact and contactless fingerprints, facilitating robust matching across modalities. F2P achieves this by employing a multi-branch convolutional neural network that concurrently processes flash and non-flash contactless images. Each branch is tailored to extract features optimal for its respective input, mitigating the limitations of each capture method. The network incorporates a novel feature fusion module to combine these disparate representations into a cohesive and discriminative embedding. The approach leverages the complementary strengths of flash and non-flash images, balancing ridge detail preservation with noise reduction.

The flash branch enhances ridge clarity, while the non-flash branch prioritises noise suppression and contrast enhancement, achieved through custom loss functions and carefully designed convolutional layers. The resulting fused embedding is then used for fingerprint matching, utilising a standard cosine similarity metric. A key contribution of this work is the demonstration of improved performance on a large-scale, cross-database evaluation. Experiments conducted on the IIT Delhi Contactless Fingerprint Database (IITD-CFD) and a newly collected in-house dataset demonstrate a substantial reduction in Equal Error Rate (EER) compared to state-of-the-art contactless systems. Furthermore, the unified embedding space enables effective matching between contactless and contact fingerprints, achieving competitive performance on cross-modal matching tasks.

Paired Flash-Non-Flash Fingerprint Image Acquisition and Database Creation

Researchers developed Fusion2Print (F2P), a framework designed to capture and fuse paired flash and non-flash contactless fingerprint images, addressing the challenges of degraded ridge clarity. The study pioneered the construction of the Flash, Non-Flash Fingerphoto (FNF) Database, comprising 3,140 valid fingerprint images collected from 79 individuals over two sessions, each with 20 impressions. A smartphone application, built using the Expo Camera framework, was engineered to capture paired images with a 600ms delay between acquisitions, minimising displacement and eliminating the need for manual flash adjustments. All images were consistently captured in macro mode with a fixed zoom factor of 0.45 to ensure consistent scale and ridge detail.

To isolate ridge-preserving signals, scientists performed manual flash-non-flash subtraction in the frequency domain, attenuating illumination noise while retaining crucial ridge structure. This technique leverages the differing spectral characteristics of ridge detail and illumination noise, enhancing clarity without sacrificing vital information. The F2P pipeline then integrates both modalities using a lightweight attention-based fusion network, which emphasises informative channels and suppresses noise, creating a fused image. A subsequent U-Net enhancement module optimally weights grayscale values, further refining image quality and preparing it for feature extraction.

The research team employed a ResNet-18 based embedding model, designed for cross-domain compatibility, to generate discriminative and robust representations. This unified embedding allows seamless comparison of both contactless and traditional contact-based fingerprints during verification. Experiments demonstrated that F2P significantly enhances ridge clarity and achieves an Area Under the Curve (AUC) of 0.999 and an Equal Error Rate (EER) of 1.12% in recognition performance, exceeding the capabilities of single-capture baseline systems. This advancement promises improved matching robustness and facilitates interoperability between contactless and contact-based technologies.

Flash-Non-Flash Fusion for Contactless Fingerprint Recognition

Scientists have developed Fusion2Print (F2P), a framework for contactless fingerprint recognition that systematically captures and merges paired flash and non-flash images. The research addresses the challenge of degraded ridge clarity in contactless scans caused by illumination variation and skin discoloration. Experiments involved constructing the FNF Database and performing manual subtraction of flash and non-flash images to isolate ridge-preserving signals, successfully attenuating low-frequency noise while preserving crucial ridge detail. The team implemented a lightweight attention-based fusion network, DualEncoderFusionNet, to integrate the flash and non-flash modalities.

This network maps six-channel inputs , RGB from both images , to a three-channel fused output, emphasizing informative channels and suppressing noise. Measurements confirm that the network selectively amplifies high-frequency components via a Fourier transform, enhancing ridge clarity. Further refinement was achieved through a U-Net enhancement module, which produces an optimally weighted grayscale image, delivering superior ridge detail. Results demonstrate that F2P significantly enhances ridge clarity and achieves an Area Under the Curve (AUC) of 0.999 in recognition performance. The system also achieved an Equal Error Rate (EER) of 1.12%, surpassing the performance of single-capture baseline systems.

Channel-wise RGB contrast analysis revealed that blue and green channels contribute most to ridge visibility, informing the design of a U-Net Enhancement Model that prioritizes these channels during grayscale conversion. To further refine the process, scientists employed a weighted custom loss function during U-Net training, incorporating L1 loss, structural similarity (SSIM), Fourier loss, edge loss, and perceptual loss. This loss function, applied to 512×512 pixel images, drove reconstruction while rewarding high-frequency details, achieving finer detail than the initial target images. The final enhanced grayscale ridge map was then used to generate discriminative and robust representations compatible with both contactless and contact-based fingerprints for verification purposes.

Flash-Non-Flash Fusion for Fingerprint Verification

Fusion2Print (F2P) represents a significant advance in contactless fingerprint verification, introducing the first end-to-end framework designed to jointly optimise flash-non-flash fusion, learned image enhancement, and discriminative embedding generation. A central achievement is the creation of the Flash, Non-Flash Fingerphoto (FNF) Database, a novel paired dataset that facilitates systematic modelling of the complementary characteristics of flash and non-flash image capture, an area previously under-explored. Through manual subtraction of flash-non-flash images, researchers identified ridge-preserving cues which subsequently improved minutiae localisation. The developed DualEncoderFusionNet adaptively combines illumination-robust features, while a U-NetEnhancer refines ridge clarity through channel-balanced processing. Importantly, the resulting TripletDistil-Net generates compact embeddings that demonstrate generalisation across contact, contactless, and mixed-domain inputs, achieved via cross-domain fine-tuning. Experiments demonstrate strong cross-.