Person Identification and Gender Classification Based on Vision Transformers for Periocular Images

ABSTRACT

Simple SummaryThe idea of identifying persons using the fewest traits from the face, particularly the area surrounding the eye, was carried out in light of the present COVID-19 scenario. This may also be applied to doctors working in hospitals, the military, and even in certain faiths where the face is mostly covered, except the eyes. The most recent advancement in computer vision, called vision transformers, has been tested for the UBIPr dataset for different architectures. The proposed model is pretrained on an openly available ImageNet dataset with 1 K classes and 1.3 M pictures before using it on the real dataset of interest, and accordingly the input images are scaled to 224 × 224. The PyTorch framework, which is particularly helpful for creating complicated neural networks, has been utilized to create our models. To avoid overfitting, the stratified K-Fold technique is used to make the model less prone to overfitting. The accuracy results have proven that these techniques are highly effective for both person identification and gender classification.AbstractMany biometrics advancements have been widely used for security applications. This field's evolution began with fingerprints and continued with periocular imaging, which has gained popularity due to the pandemic scenario. CNN (convolutional neural networks) has revolutionized the computer vision domain by demonstrating various state-of-the-art results (performance metrics) with the help of deep-learning-based architectures. The latest transformation has happened with the invention of transformers, which are used in NLP (natural language processing) and are presently being adapted for computer vision. In this work, we have implemented five different ViT- (vision transformer) based architectures for person identification and gender classification. The experiment was performed on the ViT architectures and their modified counterparts. In general, the samples selected for trainvaltest splits are random, and the trained model may get affected by overfitting. To overcome this, we have performed 5-fold cross-validation-based analysis. The experiment's performance matrix indicates that the proposed method achieved better results for gender classification as well as person identification. We also experimented with train-val-test partitions for benchmarking with existing architectures and observed significant improvements. We utilized the publicly available UBIPr dataset for performing this experimentation.