Facial Image-Based Gender Identification Using MobileNetV2

Abstract

Face-based gender classification has become an important component in security, demographic analysis, and personalized digital services, where rapid and automated facial analysis is required. This study proposes a lightweight gender classification framework based on MobileNetV2 with ImageNet transfer learning for deployment in resource-constrained environments. Experiments were conducted using a public Kaggle dataset containing 15,934 facial images (7,884 male and 8,050 female). The preprocessing pipeline consisted of image resizing, Haar Cascade face detection, pixel normalization, and brightness–contrast augmentation. To prevent data leakage, the dataset was split into training, validation, and testing subsets (70:20:10) before augmentation, ensuring that augmented versions of the same image were confined to the training subset. The model was trained using a two-stage transfer learning strategy in which the classification head was trained first, followed by fine-tuning of the last 30 layers of MobileNetV2. Performance was evaluated on a held-out test set comprising 707 unaugmented facial images using accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC. The proposed model achieved a test accuracy of 94.91%, with precision, recall, and F1-score of 96.42%, 94.49%, and 95.44% for the female class, and 93.04%, 95.45%, and 94.23% for the male class, demonstrating balanced classification performance across both classes. A qualitative evaluation using ten independently sourced facial images correctly classified seven samples, suggesting that further investigation is required to improve generalization beyond the training data distribution. The results indicate that MobileNetV2 provides an effective balance between classification accuracy and computational efficiency, making it suitable for deployment on smartphones, embedded systems, and other resource-constrained computer vision applications.