Abstract

This paper proposes a multiscale residual framework for gender classification on fingerprint images. It addresses the tough problem where most gender-related features in fingerprint ridges are subtle, and previous methods relied mainly on spatial features and failed to utilize the multi-scale features, attention, and dimension reduction. Therefore, morphological preprocessing, discrete wavelet-based multiscale feature extraction, Siamese network-based dimensionality reduction, and residual neural network classification, involving attention and non-local block, is proposed to effectively solve this problem. This framework performs morphological preprocessing on fingerprint images and decomposes the image into low-, mid-, and high-frequency channels by discrete wavelet transform; then, these channels are embedded into feature vectors by a Siamese ResNeXt-SE structure, and their cross-attention is fused to retain inter-/intra-class discriminability. Then, it applies a residual neural network with attention and a non-local block on the features for classification. On the SOCOFing dataset, which has 6,000 fingerprint images, the accuracy, precision, recall, and F1-score are 99.50%, 99.68%, 98.82%, and 0.99, respectively, and AUC is 0.99 for both genders, which achieves excellent results. Ablation studies have proven that morphological operators, wavelet decomposition, and Siamese compression do contribute to the framework individually. Experiments against two common types of noises, Gaussian and salt-and-pepper, also demonstrate the robustness against noise with trivial loss of accuracy, and comparing with prior CNN and CNN-SVM models, the method shows superiority by utilizing multiscale and frequency domain information comprehensively. However, it is constrained to use high-quality grayscale images and is not tested on adversarial robustness. The framework shows high potential for applications such as automatic fingerprint identification systems, forensics, and large biometric databases since the gender classification is effective, scalable, and discriminative.