基于改进RT-DETR的草原鼠洞智能识别与检测模型设计与试验OA

In recent years,the rapid proliferation and expansion of rodent burrows have become one of the major causes of grassland degradation and ecological imbalance.The dense distribution of burrows destroys turf structure,weakens soil stability,reduces vegetation coverage and forage productivity,and seriously threatens the ecological security and sustainable development of animal husbandry in grassland areas.To address the problems of insufficient small-target recognition,inadequate feature extraction,and significant background interference in existing detection algorithms,this study proposes an improved real-time detection model,RT-DETR-ECG(efficient global perception and feature selection),based on the Real-Time Detection Transformer(RT-DETR)framework.The model first introduces the Efficient Vision Mamba(EfficientVIM)module,which achieves efficient fusion of global and local features through state-space modeling and multi-scale feature interaction,thereby enhancing the perception of small targets.Second,it constructs a Convolutional Gated Linear Unit(CGLU)module,which dynamically suppresses complex background noise via the synergy of convolution and gating mechanisms,effectively highlighting the edges and textures of burrows.Finally,it designs a small-object weighted loss function(CGLU-Loss),which introduces center-offset weighting and local-overlap gain terms in the regression stage to strengthen gradient contributions from small-scale samples and improve localization precision.Experiments based on a dataset of 2 397 cross-season UAV images of grassland burrows demonstrate that RT-DETR-ECG outperforms existing models in detection accuracy,real-time performance,and lightweight efficiency.With an IoU threshold of 0.5,the model achieves a mean average precision(mAP@0.5)of 96.3%,5.1%higher than the original RT-DETR,with a detection speed of 91.8 f/s and a 53.7%reduction in computational load.These results verify that RT-DETR-ECG effectively balances detection accuracy and efficiency,providing a reliable and efficient approach for intelligent grassland rodent burrow detection and ecological monitoring.