Enhancing Fingerprint Gender Classification Using VGG19 Transfer Learning with Image-Based Synthetic Oversampling

Acute class imbalance and feature degradation in biometric fingerprint datasets significantly degrade gender classification accuracy, causing a shift of the decision boundary towards the majority class and compromising forensic reliability. The objective of this research is to develop and validate a forensically robust deep learning framework capable of improving the performance of fingerprint-based gender classification under severe class imbalance while preserving biologically meaningful fingerprint characteristics. The research method uses a quantitative experimental approach by combining fixed-feature Transfer Learning based on the VGG19 architecture with a Spatial Topology-Preserving Image-Based Synthetic Oversampling technique. Unlike standard data augmentation applied uniformly across all classes or feature-space oversampling that distorts high-dimensional manifolds, IBSO is a targeted image-domain protocol executed exclusively on minority-class training samples prior to feature extraction to preserve biological ridge topology without introducing synthetic artifacts. Using the Sokoto Coventry Fingerprint Dataset (SOCOfing), consisting of 6,000 real fingerprint images (4.770 males and 1.230 females), the dataset was strictly partitioned into Training (70%), Validation (15%), and Testing (15%) subsets before augmentation to prevent data leakage. IBSO balanced the training set to 6,678 images (1:1 ratio), while the validation and testing subsets remained unaugmented. The results show that, when evaluated on an independent testing dataset, the proposed framework achieves 99.92% Accuracy, 99.86% Precision, 100.00% Recall, and 99.93% F1 score, significantly outperforming the baseline ResNet50 (94.12% accuracy) with statistical significance (p < 0.0001). The conclusion of this study is that combining a locked 3 × 3 linear convolution layer with spatial IBSO effectively stabilizes minority class learning while maintaining fine dermatoglyph line density, thus providing an accurate, scalable, and forensic-compliant framework for biometric gender identification and supporting its application in real-world forensic biometric systems.