Multi-class classification framework for ear imaging using attention-based hybrid dense CNN and non-linear manifold approximation
Mohammed Alezzi · Samarra Journal of Engineering Science and Research · 2025
The ear identification problem is the task of recognizing or authenticating an individual's identity based on the unique characteristics of their ear. Like fingerprints or face recognition, ear identification tasks mainly depend on the unique shape, structure, and geometry of the ear for accurate classification. Moreover, realizing high classification accuracy is challenging for both security systems and forensic investigations since it boosts the robustness of security systems and favors law enforcement efforts. In this work, we develop a novel deep learning-based multi-class classification approach for ear images that uses a hybrid-dense CNN architecture combined with a self-attention mechanism and uniform manifold approximation and projection (UMAP) for dimensionality reduction. We conducted our experiments with ear images labeled for multi-class classification, using a Dense CNN network built from Xception-Net and Dense-Net architectures to capture intricate patterns. The self-attention mechanism enhances feature representation, while UMAP optimizes feature space by simplifying high-dimensional data. The model processes low-dimensional features through fully connected layers and softmax for classification.The experimental findings of the proposed methodology showed classification accuracies of 98.24% on the first dataset and 94.66% on the second. When compared to cutting-edge models like VGG-Net, Inception Networks, and ResNet-50, our technique outperformed them in terms of classification accuracy, precision, recall, and F1 score. This framework has great potential for use in a variety of applications, including biometric identification, medical diagnostics, forensic science, security systems, and assistive technology.