Abstract: UAV object detection technology holds great potential for ecological restoration monitoring in photovoltaic (PV) power stations. However, practical applications face challenges such as complex background interference, blurred features, and small object sizes. To address these issues, this paper proposes MDS-DETR, an improved object detection model based on RT-DETR. First, an improved backbone network named CSP-MambaVision is designed. By synergizing the gradient shunt characteristics of CSP with the linear global modeling capabilities of MambaVision, and introducing SFS-Conv and EMA for progressive feature optimization, this backbone significantly enhances the visual semantic modeling capacity in complex environments. Second, a lightweight DTAB is introduced to replace the native AIFI module. Relying on grouped channel control and masked spatial constraints, DTAB expands the receptive field to capture multi-scale contextual information while optimizing the model's perception and discriminative abilities for objects with ambiguous features. Finally, a small object detection module, SOEP-MFM, is proposed. Utilizing cross-scale feature recombination and a dynamic weight adjustment mechanism, this module achieves multi-level preservation of small object features within the network, effectively strengthening their representation and improving the detection accuracy of small objects.Experiments on public datasets demonstrate the significant advantages of MDS-DETR. Compared to the baseline model, Precision, Recall, mAP50, and mAP50-95 increase by 4.96%, 3.04%, 4.09%, and 3.58%, respectively. The model outperforms other mainstream algorithms. Furthermore, the study applies the optimized MDS-DETR to PV ecological restoration monitoring. The results align closely with actual measurements, indicating that the model provides reliable support for ecological restoration planning.

摘要： 无人机目标检测技术在光伏电站生态修复监测中的应用潜力巨大，但在实际应用中面临背景干扰、特征模糊及目标尺寸小等挑战。针对上述关键问题，本文提出一种基于改进RT-DETR(Real-Time DEtection TRansformer)的目标检测模型MDS-DETR(MambaVision driven Dilated-attention Small-object DEtection TRansformer)。首先，设计改进型主干网络 CSP-MambaVision(Cross-Stage Partial and MambaVision Hybrid Backbone Network)，通过将CSP的梯度分流特性与MambaVision的线性全局建模能力协同，并引入SFS-Conv和EMA对特征渐进式优化，增强对复杂环境的视觉语义建模能力；其次，将轻量化后的注意力转换块DTAB(Dilated Transformer Attention Block)替代原生AIFI(Attention-based Intrascale Feature Interaction)模块，依托分组通道控制与掩码空间约束，在扩大感受野捕捉多尺度上下文信息的同时优化模型对特征模糊类目标的感知与判别能力；最后，提出小目标检测模块SOEP-MFM(Small Object Enhance Pyramid with Modulation Fusion Module)，利用跨尺度的特征重组与动态调整权重机制，实现小目标特征在网络中的多层次保持，有效增强小目标的表征能力，提升模型对小目标的检测精度。在公开数据集上的实验结果表明，MDS-DETR各项指标上较现有算法具有显著优势，其中Precision、Recall、mAP50%及mAP50-95%较基线模型分别提高了4.96%、3.04%、4.09%与3.58%，优于其他主流算法。此外，将基于迁移学习优化的MDS-DETR模型应用于光伏生态修复监测任务中，结果表明其测量的覆盖度与实测数据具有高度一致性，可为光伏电站的生态修复规划提供可靠的支撑。