# Performance evaluation of enhanced deep learning classifiers for person identification and gender classification

## TL;DR

This paper introduces an enhanced deep learning classifier for improving accuracy and efficiency in person identification and gender classification using periocular images.

## Contribution

The paper proposes a novel EDLC paradigm with a hexagon-shaped ROI extraction and adaptive optimization for improved biometric classification.

## Key findings

- SSA-RTNet achieved 99.8% and 99.67% accuracy for person identification on UBIPr and UFPR datasets.
- SSA-RTNet achieved 98.4% and 99.68% accuracy for gender classification on UBIPr and UFPR datasets.
- The proposed EDLC models showed significant improvement over existing methods in accuracy and efficiency.

## Abstract

Person authentication using periocular images is a prominent research domain. Although the biometric identification systems have advanced, the existing approaches still struggle with accuracy, overfitting issues and computational efficiency, especially when utilizing periocular images for person identification and gender classification. In order to overcome these limitations, this paper proposes an enhanced deep learning classifier (EDLC) paradigm to recognize a person based on the periocular region within a face. A novel Hexagon-shaped ROI extraction is performed in the localization phase to extract the periocular ROIs. Following that, the feature extraction mechanism is accomplished utilizing the Laplacian transform. Finally, three distinct custom EDLCs are employed, such as dilated axial attention convolutional neural network, self-spectral attention-based relational transformer net, parameterized hypercomplex convolutional Siamese network for classification. Further, an adaptive coati optimization algorithm is used to adjust the hyperparameters of the classification models. The efficacy of the model is assessed concerning different outcome indicators. It is also compared with the recent competitive models. For person identification, the SSA-RTNet has achieved a maximum accuracy of 99.8% and 99.67% using the UBIPr and UFPR datasets respectively. Similarly, for gender classification, an accuracy of 98.4% and 99.68% using SSA-RTNet was obtained for the UBIPr and UFPR datasets. As a result, it is perceived that a considerable improvement was observed using the enhanced models.