Ensemble of transfer learning and light-weight convolutional neural network model for an effective ear recognition system

• Published in: Evolving systems Vol. 15; no. 1; pp. 115 - 131
• Main Authors: Mehta, Ravishankar, Singh, Koushlendra Kumar
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01-02-2024; Springer Nature B.V
• Subjects: Accuracy; Artificial Intelligence; Artificial neural networks; Biometrics; Classification; Complex Systems; Complexity; Datasets; Deep learning; Ear; Engineering; Learning; Machine learning; Neural networks; Original Paper; Pattern recognition; Wavelet transforms; Weight reduction; Biometrics; Deep learning; Accuracy; Ear recognition; Ensemble; ROC
• ISSN: 1868-6478; 1868-6486
• DOI: 10.1007/s12530-023-09561-6

The transfer learning and deep convolutional neural network-based recognition models have been merged to develop an efficient model in the contemporary time. The ear recognition system has major challenges in terms of the number of sample images, illumination changes, and other environmental challenges. To make the recognition system free from these challenges, the authors utilized the strength of both the existing three pre-trained models (e.g. VGG19, VGG16, DenseNet201) and three simple light-weight Convolutional Neural (CNN) models. The proposed method encompasses the transfer learning as well as simple deep convolution neural network model to design the ensemble model. The first three models are based on the transfer learning approach which uses VGG19, VGG16, and DenseNet201 whereas the last three models are light-weight Convolution Neural Networks (CNNs) that consist of comparatively less number of convolutional layers. The use of transfer learning in the proposed approach overcomes the limitation of small datasets whereas the use of lightweight CNN models reduces the overhead of the training time of the model. The proposed model is validated with the IITD-II ear dataset in which we achieved recognition accuracy of 95.96% and 93.08% through weighted ensemble and average ensemble techniques. The combined approach of transfer learning and deep CNN show improvements in performance accuracy by 2–4% when compared to individual models. It achieves a precision of 96.74, recall of 96.37, and f-score of 95.96 using the weighted ensemble method which is an improvement over the other state-of-the-art methods.