Recognition of Local Bird Species of Yunnan Using Transfer Learning Based on ResNet

This study created the GMBID dataset for locally protected bird species in Yunnan and explored the application of transfer learning with different ResNet architectures (ResNet18, ResNet34, ResNet50) on this dataset. By comparing the performance of these models in terms of accuracy, precision, recall, and F1 score, the study found that ResNet34 performed the best, achieving a good balance between high accuracy, model complexity, and computational efficiency. Although ResNet50 has a deeper architecture, its performance improvement over ResNet34 was limited, and it incurred higher computational costs. ResNet18, while training faster, had slightly lower overall performance compared to ResNet34. This study highlights the practical significance of applying transfer learning in data-scarce regions, particularly in the fields of bird conservation and ecological monitoring. Future work could focus on expanding the dataset, experimenting with more advanced model architectures, and optimizing transfer learning strategies to further enhance model performance, providing technical support for wildlife conservation.