Open Access

ARTICLE

An Improved Transfer-Learning for Image-Based Species Classification of Protected Indonesians Birds

Chao-Lung Yang¹, Yulius Harjoseputro^2,3, Yu-Chen Hu⁴, Yung-Yao Chen^2,*

1 Department Industrial Management, National Taiwan University of Science and Technology, Taipei, 106, Taiwan
2 Department Electronics and Computer Engineering, National Taiwan University of Science and Technology, Taipei, 106, Taiwan
3 Department Informatics, Universitas Atma Jaya Yogyakarta, Yogyakarta, 55281, Indonesia
4 Department Computer Science and Information Management, Providence University, Taichung, 433, Taiwan

* Corresponding Author: Yung-Yao Chen. Email:

Computers, Materials & Continua 2022, 73(3), 4577-4593. https://doi.org/10.32604/cmc.2022.031305

Received 14 April 2022; Accepted 12 June 2022; Issue published 28 July 2022

Abstract

This research proposed an improved transfer-learning bird classification framework to achieve a more precise classification of Protected Indonesia Birds (PIB) which have been identified as the endangered bird species. The framework takes advantage of using the proposed sequence of Batch Normalization Dropout Fully-Connected (BNDFC) layers to enhance the baseline model of transfer learning. The main contribution of this work is the proposed sequence of BNDFC that can be applied to any Convolutional Neural Network (CNN) based model to improve the classification accuracy, especially for image-based species classification problems. The experiment results show that the proposed sequence of BNDFC layers outperform other combination of BNDFC. The addition of BNDFC can improve the model’s performance across ten different CNN-based models. On average, BNDFC can improve by approximately 19.88% in Accuracy, 24.43% in F-measure, 17.93% in G-mean, 23.41% in Sensitivity, and 18.76% in Precision. Moreover, applying fine-tuning (FT) is able to enhance the accuracy by 0.85% with a smaller validation loss of 18.33% improvement. In addition, MobileNetV2 was observed to be the best baseline model with the lightest size of 35.9 MB and the highest accuracy of 88.07% in the validation set.

---

Keywords

---