DSCF-DET: An RT-DETR-Based Framework for Fine-Grained Detection of Musk Deer and Visually Similar Artiodactyls
Open Access
Abstract
Musk deer are forest-dwelling artiodactyls of high conservation value, but their wild populations remain under severe conservation pressure due to poaching driven by the demand for natural musk, together with habitat fragmentation and habitat loss. Efficient non-invasive image-based monitoring is therefore important for musk deer conservation; however, fine-grained detection of musk deer and visually similar artiodactyls in ecological images remains difficult because of background camouflage, vegetation occlusion, and high inter-class similarity. In this study, a fine-grained wildlife image dataset was constructed, and an RT-DETR-based framework, termed DSCF-DET, was proposed for automated detection in complex natural scenes. DSCF-DET integrates three task-oriented modules: DRPBlock for receptive-field-aware feature extraction, SASTE for sparse spatial encoding, and CBAFusion for cross-level feature fusion. On the constructed dataset, DSCF-DET achieved 91.4% precision, 86.2% recall, 88.7% F1-score, and 86.4% mAP50. Compared with RT-DETR-r18, it improved these metrics by 8.6, 12.1, 10.5, and 12.4 percentage points, respectively, while maintaining moderate model complexity. Visualization results showed more target-focused feature responses and reduced background-related activations. Cross-dataset experiments on an independent public wildlife dataset further suggested potential applicability to broader wildlife detection scenarios. These results indicate that DSCF-DET provides a computationally balanced approach for ecological image screening and intelligent musk deer monitoring.Keywords
Funding Information
- the national key Research and Development Program of China (2022YFC2601205)
This publication has 45 references indexed in Scilit:
- UAV-based deep learning for biodiversity monitoring: Advances, applications, and future directionsEcological Informatics, 2026
- MobileNetV4: Universal Models for the Mobile Ecosystem2024
- YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information2024
- Lightweight CNNs for Advanced Bird Species Recognition on the Edge2024
- WilDect-YOLO: An efficient and robust computer vision-based accurate object localization model for automated endangered wildlife detectionEcological Informatics, 2022
- An overview of remote monitoring methods in biodiversity conservationEnvironmental Science and Pollution Research, 2022
- Chemical compositions and pharmacological activities of natural musk (Moschus) and artificial musk: A reviewJournal of Ethnopharmacology, 2022
- End-to-End Object Detection with Transformers2020
- CBAM: Convolutional Block Attention Module2018
- Complete mitochondrial genome of the Himalayan Musk Deer, Moschus leucogaster, with phylogenetic implicationConservation Genetics Resources, 2018