Human identification system based on ear shape using convolutional neural network
Randy Antonio, Nadya Tyandra, Tiffany Angela Indryani, Ivan Sebastian Edbert, Alvina Aulia · Procedia Computer Science · 2024
Convolutional Neural Networks (CNN) represents a class of Deep Neural Networks frequently employed for image and video recognition, image classification, medical image analysis, and numerous computer vision tasks. In human identification systems, fingerprint and facial recognition are widely used. However, fingerprints are subject to deterioration, and both face and finger could become less recognizable in the presence of a cut or a wound. The ear, a less-known component in the identification system, eliminates the less advantageous traits of fingerprint and facial recognition. The methodology employed encompasses the systematic collection of datasets, model development, and model evaluation. Evaluation procedures involve user assessments conducted objectively about findings from analogous studies. The EarVN1.0 dataset serves as the foundation, and five distinct models, namely MobileNetV3-Large, VGGNet-19, ResNet50V2, EfficientNetV2-Small, and a modified version of AlexNet, have been meticulously developed. The outcomes of this research are delineated by the accuracies achieved by each model, which stand at 79%, 22%, 79%, 91%, and 52%, respectively. An in-depth exploration involving hyperparameter tuning of the EfficientNetV2-S model, focusing on key aspects such as batch size, pooling layer, dropout, and maximum epoch limit. This meticulous tuning resulted in a notable accuracy improvement, elevating the performance of the EfficientNetV2-S model to 92%.