EffRes-Bio: Lightweight Cross-Sensor Fingerprint Verification with a Controlled Evaluation of Frequency-Aware Attention

Fingerprint verification on resource-constrained devices requires compact models that generalise across heterogeneous capture sensors. This paper presents EffRes-Bio, a fingerprint feature extractor of 4.34 million parameters and 0.39 GMACs that combines an ImageNet-pretrained EfficientNet-B0 backbone with generalised-mean pooling and an ArcFace embedding head, and uses it to test whether a frequency-aware channel attention (FAA) module improves transfer to unseen sensors. Models are evaluated on the three real-sensor FVC2002 databases under an open-set protocol; cross-sensor results use a leave-one-sensor-out design repeated over three random seeds, giving nine paired observations per model. Because the databases differ in fingers as well as sensors, this protocol measures generalisation to unseen sensors rather than strict interoperability. On unseen sensors, EffRes-Bio attains a mean equal error rate (EER) of 19.55 %, statistically indistinguishable from ResNet-50 (18.70 %, p = 0.39), which has 5.5 times more parameters and 10.3 times more computation, while running 3.4 times faster on a single CPU thread. It outperforms MobileNetV3-Small (22.47 %) by 2.91 percentage points (95 % CI 1.15–4.67, p = 0.005), a difference that survives Holm correction. Within a database, the three networks do not differ significantly. FAA did not improve cross-sensor performance in any of the three tested forms, and a sampling analysis explains the null result at the final feature map, where the ridge period falls to 0.13–0.21 pixels. Finally, deriving a baseline's first layer from its pretrained weights rather than re-initialising it lowered its cross-sensor EER by 2.8–3.5 points, more than any architectural difference observed.

Authors

Institutions

Publication Details

Journal
Dasinya Journal for Engineering and Informatics
Published
2026-11-13
DOI
https://doi.org/10.65542/djei.v2i4.104
Primary Topic
Biometric Identification and Security
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

EffRes-Bio: Lightweight Cross-Sensor Fingerprint Verification with a Controlled Evaluation of Frequency-Aware Attention

Wafaa Mustafa Abduallah, Khalid Ibrahim Khalaf Jajan
5 links
Dasinya Journal for Engineering and Informatics
Biometric Identification and Security
article

EffRes-Bio: Lightweight Cross-Sensor Fingerprint Verification with a Controlled Evaluation of Frequency-Aware Attention

Wafaa Mustafa Abduallah, Khalid Ibrahim Khalaf Jajan
article en

Abstract

Fingerprint verification on resource-constrained devices requires compact models that generalise across heterogeneous capture sensors. This paper presents EffRes-Bio, a fingerprint feature extractor of 4.34 million parameters and 0.39 GMACs that combines an ImageNet-pretrained EfficientNet-B0 backbone with generalised-mean pooling and an ArcFace embedding head, and uses it to test whether a frequency-aware channel attention (FAA) module improves transfer to unseen sensors. Models are evaluated on the three real-sensor FVC2002 databases under an open-set protocol; cross-sensor results use a leave-one-sensor-out design repeated over three random seeds, giving nine paired observations per model. Because the databases differ in fingers as well as sensors, this protocol measures generalisation to unseen sensors rather than strict interoperability. On unseen sensors, EffRes-Bio attains a mean equal error rate (EER) of 19.55 %, statistically indistinguishable from ResNet-50 (18.70 %, p = 0.39), which has 5.5 times more parameters and 10.3 times more computation, while running 3.4 times faster on a single CPU thread. It outperforms MobileNetV3-Small (22.47 %) by 2.91 percentage points (95 % CI 1.15–4.67, p = 0.005), a difference that survives Holm correction. Within a database, the three networks do not differ significantly. FAA did not improve cross-sensor performance in any of the three tested forms, and a sampling analysis explains the null result at the final feature map, where the ridge period falls to 0.13–0.21 pixels. Finally, deriving a baseline's first layer from its pretrained weights rather than re-initialising it lowered its cross-sensor EER by 2.8–3.5 points, more than any architectural difference observed.

Dasinya Journal for Engineering and InformaticsVol. 2(4)
Duhok Polytechnic University (IQ), Akre University for Applied Sciences (IQ)
Biometric Identification and Security
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon