Giải pháp nhận diện loài chim nguy cấp, quý hiếm dựa trên học sâu

Abstract

Accurate identification of endangered, rare, and precious bird species plays an essential role in biodiversity monitoring and conservation. This study aims to develop a reliable identification model for 26 priority conservation species by integrating deep learning with specific processing mechanisms. The proposed methodology consists of three components: pre-processing using YOLOv12 to discard non-bird images; species classification based on a transfer learning and fine-tuning ResNet model on a in-house dataset (27,457 images of 171 species); and post-processing using OpenMax to reduce confusion with species outside the training set. Experimental results show that the model achieves a Macro F1-Score of 83.54%, demonstrating its strong potential for application in biodiversity monitoring systems.