Gender Classification from Facial Images under Illumination and Head-Pose Variations Using AlexNet Features and a Grasshopper-Optimized Multilayer Perceptron

Abstract

Automatic binary gender-label classification from facial images is a well-established face-analysis task. Although current systems perform well under controlled conditions, their accuracy can deteriorate in the presence of illumination changes, head-pose variation, age differences, facial expression, and partial occlusion. This study addresses the design of a framework that can extract stable facial representations from a limited training set and learn the classifier decision boundary without relying exclusively on gradient-based optimization. The proposed method comprises three components: the frozen convolutional part of an ImageNet-pretrained AlexNet used as a feature extractor, a multilayer perceptron with one hidden layer of 128 neurons used as the classifier, and the Grasshopper Optimization Algorithm (GOA) used to search for the classifier weights and biases. After resizing, illumination normalization, and limited data augmentation, input images are passed through the convolutional backbone, and the Pool5 output is flattened into a 9,216-dimensional feature vector. The features are normalized using parameters estimated exclusively from the training set and are then supplied to the MLP. The main evaluation uses the official identity-disjoint split of GENDER-FERET, comprising 474 training images and 472 test images. The reported results indicate a test accuracy of 98.94%. For the male class, precision, recall, and F1-score are 99.15%, 98.73%, and 98.94%, respectively; for the female class, the corresponding values are 98.73%, 99.15%, and 98.94%. The small difference between class-specific results indicates balanced errors on this split. Nevertheless, the limited dataset size, controlled image acquisition, absence of cross-dataset testing, and the substantial computational cost of directly optimizing more than one million parameters constrain the external validity of the results. From the perspective of reusing deep representations while separating feature extraction from classifier optimization, the proposed framework is a viable approach for further investigation in low-data settings.