From Blurs to Birds: Localization and Classification of Hard-to-See Bird Species in Norwegian Wilderness Camera Trap Images

Camera trap imagery datasets present unique challenges, including weather conditions, motion blur, occlusions from vegetation or other animals, and instances where the animal is partially or completely outside the frame due to the camera’s fixed position. This challenge is exacerbated when the target animal species are very small, highly similar in appearance, and fast-moving, such as different species of birds. Although deep learning has been extensively applied to classify and detect animal species, bird species are either captured in static, high-quality images under optimal conditions, with no motion blur or other disturbances, or they are grouped into a single class. Real camera trap images, however, are not optimal nor of high quality. Furthermore, recognizing bird species is more detailed and intricate than identifying larger animals such as deer, moose, or reindeer, making manual annotation more time-consuming and susceptible to errors. This study utilizes state-of-the-art deep-learning neural networks to localize and classify hard-to-see bird species in Norwegian wilderness camera trap images. Experimental results show that our fine-tuned YOLOv8x architecture achieved the highest performance for localization, outperforming YOLOv9, yielding a mAP@50 score of 0.92. For the classification of hard-to-distinguish bird species, EfficientNetB7 achieved the highest accuracy and F1-score of 0.87. Grad-CAM, an explainable AI technique, was employed to identify discriminative regions within the images.