A lightweight small-target detection algorithm based on improved RT-DETR was proposed to address the challenges of small target size and difficult feature extraction in fault detection of railway overhead catenary dropper bolts. The CG-EfficientFormerV2, incorporating context-guided blocks, was adopted as the feature extraction network, maintaining multi-scale feature representation capabilities while reducing model complexity. The DySample dynamic upsampling strategy and the context-guided-block downsampling mechanism were integrated to achieve adaptive distribution of sampling points within target regions and enhance the model’s capability to capture detailed bolt features. The MASAG attention mechanism was introduced to enable effective fusion of local-global contextual information through a spatially selective modulation mechanism. An Inner-MPDIoU joint loss function was designed by combining internal keypoint distance metrics and aspect ratio constraints to optimize small-target localization accuracy. Experiments conducted on a custom-built dataset demonstrated that the proposed algorithm achieved 97.2% and 69.2% in terms of mAP@0.5 and mAP@0.5:0.95, respectively, with improvements of 4.0 and 6.6 percentage points over the baseline. Additionally, the number of parameters was reduced by 23.4%, computational complexity was decreased by 31.53%, and an FPS of 40.6 was achieved, meeting real-time requirements. In generalization experiments on the VisDrone2019 dataset, the proposed algorithm achieved 48.6% and 29.0% in terms of mAP@0.5 and mAP@0.5:0.95, respectively, with improvements of 5.8 and 3.0 percentage points over the baseline, thereby validating the cross-domain generalizability. Experimental results verified the effectiveness of the proposed algorithm, which meets the requirements for lightweight small-target detection of dropper bolts in overhead catenary systems.