<p>Transmission line fitting detection is vital for the safety and reliability of overhead transmission lines, and unmanned aerial vehicles (UAVs) have become indispensable tools for automated inspection. However, the small size, dense distribution, and extreme scale variations of power fittings in aerial images pose significant challenges for conventional object detectors, often resulting in missed or inaccurate detections. To robustly tackle these issues, we propose STF-Net, a deep learning-based high-precision detection network specifically designed for small target fitting detection in UAV imagery. STF-Net introduces a skip-layer and cross-scale three-path feature fusion network (SCTN) to enhance multi-scale feature representation and preserve fine-grained spatial details across network layers. Furthermore, a context channel and position attention (CCPA) module is incorporated to improve the network's focus on semantically relevant regions, especially for small and visually similar fittings. In addition, we improve the region proposal network to generate more precise and compact candidate regions by adapting to fitting-specific structural cues. Experimental results on our self-constructed Transmission Line Fitting Inspection (TFI) dataset demonstrate that STF-Net significantly outperforms representative state-of-the-art detectors. Notably, it achieves a 3.9% improvement in mAP50 over Faster R-CNN with FPN and delivers superior detection robustness under multi-scale and complex background conditions, making it a promising solution for real-world UAV-based inspection systems.</p>