In the inspection of transmission lines by unmanned aerial vehicles (UAVs), the detection task of corroded bolts and nuts is significantly affected due to the influence of complex background occlusion, small size, and density.In order to solve the above problems, this paper introduces a multi-scale DETR framework, termed MS-DETR, incorporating a multi-scale feature fusion mechanism and an enhanced attention mechanism. Initially, a feature extraction module is developed using a cross-stage partial connection network (CSPDarknet) as the backbone, which effectively integrates local perception, global attention, and multilayer perceptrons to achieve robust multi-scale feature aggregation. Subsequently, an advanced intra-scale feature interaction module, leveraging an improved attention mechanism, is proposed to dynamically select key points, reducing computational complexity while preserving global information interaction. Experimental results on a custom rust bolt-nut dataset demonstrate that MS-DETR enhances the mAP50 metric by 2.3% over the baseline model, showcasing superior detection accuracy for small-sized objects in complex environments. These findings suggest that MS-DETR offers an effective solution for bolt-nut detection in transmission line inspections.