Answers are grounded in this paper and cite their sources.
Gender classification from eye images is a soft-biometric task that may be relevant when facial information is partially occluded. This study evaluates a hybrid framework that uses pretrained AlexNet, VGG16 and MobileNetV2 networks as fixed feature extractors and a support vector machine (SVM) with a radial basis function (RBF) kernel for classification. The framework was tested on the MMU2 iris dataset, comprising 995 images from 100 volunteers. Contrast-limited adaptive histogram equalisation (CLAHE) was applied to enhance image contrast. The dataset was divided at participant level, with 74 volunteers (737 original images) assigned to training and 26 volunteers (258 images) reserved for testing. Horizontal flipping, rotations of up to ±20°, and random zooming and slight shifting were applied only to the training images, increasing the training set to 2,211 images. The three CNNs were used without fine-tuning, and the SVM was trained with C = 1.0 and gamma = “scale”. On the held-out test set, AlexNet–SVM, VGG16–SVM and MobileNetV2–SVM achieved accuracies of 85.66%, 98.45% and 99.22%, respectively, with corresponding F1-scores of 85.49%, 98.50% and 99.24%. MobileNetV2–SVM produced the highest observed accuracy with the fewest backbone parameters. However, these findings are preliminary because they derive from one participant-independent partition of a small dataset captured under controlled conditions. Repeated participant-level validation, cross-dataset testing and direct computational-cost measurements are required to assess generalisability and deployment suitability.
No takes yet. Share an insight, caveat, or question.
Masaod et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: