FeatherYOLO: A novel method for detecting pheasant feathers in complex cultural imagery for avian conservation monitoring

Monitoring avian-derived materials in complex visual environments remains a methodological challenge in ecological informatics and bird conservation research. Feathers often appear not as isolated biological specimens, but as narrow, deformable, partially occluded, or culturally embedded objects in costumes, performances, online images, and other visually cluttered scenes. To support image-based screening of such materials, we developed FeatherYOLO, a lightweight object-detection framework for identifying pheasant tail feathers in complex cultural imagery. FeatherYOLO was built on YOLO11n and incorporated three task-oriented modules: HDRAB2 for multi-scale contextual feature extraction, V7DownSampling for preserving positional and textural information during downsampling, and LTND for improving the detection of damaged, deformed, or partially visible feather targets. We constructed MF-2025, a curated dataset containing 2662 annotated images of four pheasant feather classes: Reeves's pheasant ( Syrmaticus reevesii ), Lady Amherst's pheasant ( Chrysolophus amherstiae ), Golden pheasant ( Chrysolophus pictus ), and Common pheasant ( Phasianus colchicus ). FeatherYOLO achieved strong overall performance, reaching an mAP@0.5 of 0.692 with 1.938 M parameters and 5.7 GFLOPs, showing a favourable balance between detection accuracy and model compactness. Additional comparisons with mainstream detectors, including YOLO26n, RT-DETR, and D-FINE-N, further clarified its performance relative to recent baselines. Diagnostic evaluation using precision-recall curves, multi-class confusion matrices, multi-model visual comparisons, and real-world image-level error analysis showed both the strengths and limitations of the model under practical visual conditions. To extend the framework beyond object detection, we also developed a regression model linking transverse band count to Reeves's pheasant tail feather length and implemented a web-based prototype for low-volume interactive screening. Together, the detection model, regression component, and web prototype provide a compact workflow for screening visually embedded pheasant feathers in avian-derived material research.