Convolutional neural network model for fingerprint-based gender classification using original and degraded images

Abstract

Fingerprint-based gender classification is a crucial component of soft biometrics, providing valuable additional information to narrow the search space in forensic investigations and large-scale identification systems. Although deep learning models, particularly convolutional neural networks (CNNs), have demonstrated significant potential, performance validation is typically performed on high-quality fingerprint images. This creates a gap between laboratory results and real-world applications, where fingerprint evidence is often found in a degraded state, such as smudged, distorted, or partially damaged. This study attempts to bridge this gap by proposing a more realistic training approach. We design a lightweight and computationally efficient CNN and train it on a comprehensive combined dataset. The main contribution of this study lies in the data training strategy, which explicitly combines real and synthetically modified fingerprint images from the Sokoto coventry fingerprint (SOCOFing) dataset into a single, unified training set. Experimental results show that the proposed model achieves very high classification accuracy (97.39%) on a test set that also includes a combination of original and degraded images. This finding not only confirms the effectiveness of diverse data-based training to produce more robust models but also establishes a new benchmark for fingerprint based gender classification research under conditions more representative of practical scenarios.