ENHANCING WILDLIFE CONSERVATION: A DEEP LEARNING FRAMEWORK FOR ACCURATE AND REAL-TIME AMUR TIGER IDENTIFICATION

Abstract

Modern wildlife conservation efforts are hampered by a lack of non-invasive monitoring methods for endangered species, which has generated a need for automated species identification. In this paper, we present a novel deep learning framework that integrates EfficientNetB3 with YOLOv8 for real-time detection of Amur tigers, which would improve automated detection over traditional manual tracking methods. The framework applies transfer learning to improve EfficientNetB3 to recognize tigers by their unique fur patterns and other distinctive morphological features. We generated a dataset of 1,886 images of tigers for training, and then applied multiple preprocessing techniques to increase the efficiency of the training phase (e.g., to improve the robustness of the model to variations in input data we applied resizing, normalization, and augmentation). Our model achieved a test accuracy (97.88%) and macro-average precision and recall (exceeding 95%) that demonstrates a general ability to accurately classify images in a wide range of natural environments. In addition, YOLOv8 real-time video captioning and detection functionality has been incorporated and deployed through a Streamlit web application. Our framework has the highest accuracy compared to traditional methods used for the non-invasive wildlife monitoring, and provides a new scalable approach in this field. We also support ecological research by providing a new reliable automated tool for conservationists that eliminates the need for field personnel to tag or mark animals. The system high potential for mass use in wildlife management and studying biodiversity can be seen from its high performance and ease of use.