Performance comparison of four pre-trained deep learning models for fingerprint-based gender classification

Fingerprints Recognition System is known to be the best tool for verification, and identification because it adopts the use of human biological traits that are unique and permanent throughout one’s lifetime. It is widely used in surveillance, forensic investigation, border control access and criminal investigation. This paper aims to develop a model that is capable of classifying an individual’s gender from fingerprint images using four pre-trained Convolutional Neural Networks architectures. The goal is to evaluate their performances, comparing their results, and also comparing their results with other state-of-the-art models using the Sokoto Coventry Fingerprint (SOCOFing) dataset. To achieve our aim, an experimental approach was adopted where features were extracted, and different pre-processing techniques such as cropping, resizing, and rotation were done; the dataset was balanced with down sample technique. The dataset was split into a ratio of 80:10:10 for the training, validation, and testing set respectively. That implies that for 1,000 males, we randomly split to 800 for training, 100 for validation, and 100 for testing. The same was repeated for females. So, a total of 1,600 data was used for training, 200 for validation, and 200 for testing. We employed four deep learning transfer models, which are ResNet50, DenseNet121, EfficientNet-B0, and VGG19. EfficientNet-B0 emerges as the best model, obtaining the best results across all metrics with 74.69% accuracy and 82.79% Area Under Curve (AUC), closely followed by ResNet50. The consistent balance between sensitivity and specificity across all models indicates robust learning without gender bias, while the strong AUC scores confirm meaningful classification capability.