A TRANSFER LEARNING–BASED FRAMEWORK FOR GENDER RECOGNITION UNDER POSE AND ILLUMINATION VARIATIONS USING PRE-TRAINED CONVOLUTIONAL NEURAL NETWORKS

Abstract

Accurate gender classification from facial images is a fundamental task in computer vision with applications in security, human-computer interaction, and intelligent systems. In this study, we propose a robust framework that combines deep feature extraction using a pre-trained convolutional neural network with a multilayer perceptron classifier optimized via a metaheuristic Grasshopper Optimization Algorithm. The CNN extracts discriminative features from facial images, while the optimized MLP ensures efficient and accurate classification by overcoming local optima issues commonly encountered in conventional training methods. The proposed method was evaluated on the GENDER-FERET dataset, achieving a test accuracy of 98.94% and demonstrating balanced performance across male and female classes, with precision, recall, and F1-scores exceeding 98% for both categories. The experimental results highlight the framework's robustness against variations in facial pose, illumination, and expression. This approach provides a reliable and efficient solution for gender recognition and offers a flexible foundation for future extensions to real-time applications and multi-task facial analysis systems.