Fingerprint Classification for AFIS Search-Space Reduction Using Deep Learning

Thumbnail Image

Date

2026

Journal Title

Journal ISSN

Volume Title

Publisher

University of Mohamed El Bachir El Ibrahimi - Bordj Bou Arréridj

Abstract

Fingerprint-based Automatic Fingerprint Identification Systems (AFIS) perform 1:N matching by comparing each query against every enrolled template in the gallery, a procedure whose computational cost scales linearly with gallery size and becomes prohibitive at operational scale. This thesis addresses that bottleneck by developing a deep learning-based Henry-class pre-classification stage that partitions the enrolled gallery into five class-labelled bins before matching, restricting each identification query to the relevant subset and reducing the active comparison set without discarding correct candidates. The classification system was developed and evaluated on NIST Special Database 4 (NIST SD4), a 2 000-image benchmark comprising matched F/S impression pairs across all five Henry classes. Three fingerprint input representations were investigated — raw CLAHE-enhanced images, Orientation Field Enhancement Maps (OFEM), and HSV-encoded orientation fields — and five ImageNet-pretrained convolutional backbone architectures were evaluated: EfficientNet-B3, ResNet50, MobileNetV3-Small, VGG16-BN, and ConvNeXt-Tiny. Orientation-field encoding consistently outperformed raw and OFEM representations across all architectures, establishing input representation quality as the dominant performance factor. ConvNeXt-Tiny, trained on a dual raw-and-orientation-field input, achieved 92.50% standalone accuracy. A binary VGG16-BN specialist, trained with Focal Loss and controlled oversampling to address the chronic underperformance on the Tented Arch minority class, raised the Tented Arch F1-score from 0.606 to 0.808. Integration of the specialist through a confidence-gated veto mechanism produced the final dual-architecture system, reaching 93.25% overall classification accuracy. The operational contribution of the classification system was validated by integrating it with the SourceAFIS 3.18.1 minutiae-based matching engine within a complete 1:N identification pipeline. Three search-space scenarios were evaluated: full-gallery baseline (Scenario A, 2 000 comparisons), single-bin classification-guided search (Scenario B, ≈400 comparisons), and dual-bin search (Scenario B+, ≈800 comparisons). The B+ strategy preserved near-complete identification accuracy — 87.50% Rank-1 — while reducing active comparisons by 2.61× relative to the full-gallery baseline. A Gabor score-fusion stage, applying Gabor-enhanced template matching in parallel with raw matching and combining scores via a conservative max-fusion rule, lowered the Equal Error Rate from 13.36% to 8.58% by recovering 153 identification failures attributable to poor-quality probe impressions, while breaking only 14 previously passing matches. The principal finding of this thesis is the empirical demonstration that a deep learning classification front-end, integrated with a traditional minutiae matcher, delivers measurable and reproducible search-space reduction without sacrificing identification accuracy. The dual-architecture design, the Gabor fusion strategy, and the honest characterisation of the conditions under which the approach succeeds and fails together constitute a reproducible and extensible foundation for future work in adaptive AFIS search-space reduction.

Description

Keywords

Fingerprint classification · Henry system · AFIS · Deep learning · ConvNeXt · Transfer learning · Orientation field · OFEM · SourceAFIS · Search-space reduction · Score fusion · NIST SD4

Citation

Endorsement

Review

Supplemented By

Referenced By