Abstract
Crested ibis (Nipponia nippon), as a rare and endangered species, requires accurate monitoring for effective biodiversity conservation and ecological management. However, in complex natural environments, automatic detection and segmentation of crested ibis remain challenging due to factors such as pose variation, background clutter, occlusion, and high inter-individual similarity. To address these challenges, this study proposes an improved model, named YOLO-FSD, based on YOLOv11n for robust crested ibis detection and segmentation. Specifically, the proposed model incorporates a C3K2 module and an adaptive fine-grained channel attention (FCA) module into the backbone to enhance fine-grained feature representation. A semantic-enhanced attention module (SEAM) is introduced in the neck to improve multi-scale feature fusion under occlusion conditions. In addition, GIoU-based non-maximum suppression (GIoU-NMS) is employed to optimize candidate box selection in dense scenes. Experimental results demonstrate that YOLO-FSD achieves a detection mean average precision (mAP) of 92.5%, with a detection precision of 92.8%, recall of 91.2%, and F1-score of 92.0%, while retaining a compact parameter count of 3.83 million. Ablation studies further confirm that the incorporation of FCA, SEAM, and GIoU-NMS contributes substantially to performance improvement in natural environments. Heatmap visualizations suggest that the YOLO-FSD model focuses more strongly on crested ibis regions under occlusion and in complex backgrounds. Compared with mainstream detection and segmentation models, YOLO-FSD achieves superior accuracy while maintaining relatively low computational complexity. Furthermore, when deployed on the HZHY-AI300G edge device, YOLO-FSD achieves an mAPBox of 84.9% and an mAPMask of 84.5%, demonstrating its effectiveness for edge-device deployment. These results demonstrate that the proposed model provides an efficient and deployable solution for crested ibis monitoring in the wild, offering valuable technical support for intelligent monitoring of endangered bird species.