A vision transformer-based automated human identification using ear biometrics

Abstract

Recent years Vision Transformers (ViTs) have gained significant attention in the field of computer vision for their impressive performance in various tasks, including image recognition and machine translation tasks, question answering, text classification, image captioning. ViTs performs better on several benchmark image datasets such as ImageNet with fewer parameters and computation compared to CNN-based models. The self-attention part performs the feature extraction component of the convolutional neural network (CNN). The proposed model provides a framework on vision transformer-based model for 2D ear recognition. The self-attention part is jointly applied with Convolutional Neural Network (CNNs) in the proposed model. Adjustments and fine-tuning has been done based on the specific characteristics of the ear dataset and the desired performance requirements. In the field of deep learning, the application areas of the CNNs have been proven to be de-facto mainly due to its learning capability of spatially local representations based on their inductive biases, learning the global representation further enhances the recognition accuracy through self-attention mechanism of vision transformers (ViT's). This has been made possible by direct applications of transformer on to the sequence of image patches for better performance in classifying the images. The proposed work utilizes various patch size of images during the model training. From the experimental analysis, it has been observed that with patch size 16 × 16 it achieves highest accuracy of 99.36%. The proposed model has been validated with the Kaggle and IITD-II data set. The efficiency of the proposed model over the existing models has been also reported in the present work.