Accurate detection of small transmission towers in remote sensing images is critical for low-altitude aviation safety and power-grid monitoring, yet it remains challenging due to complex backgrounds, weak target responses, and severe feature submersion. To address these issues, we propose a Semantic Boundary-Guided Dynamic Detection Network (SBG-DDN) that integrates a semantic boundary-guided representation framework with a dynamic detection and localization optimization scheme. Specifically, the proposed method combines a frozen DINOv3 backbone to provide global semantic priors and a CSPDarknet backbone to capture local boundary-sensitive details while further enhancing their interaction through the Semantic-Boundary Fusion Module (SBFM). In addition, a Dynamic Semantic-Boundary Optimization Head (DSBOH) and an Adaptive Structure-Aware Transmission Tower IoU (ASTIoU) loss are introduced to improve multi-scale feature adaptation and geometry-aware localization for sparse and elongated transmission towers. To support broader evaluation in this area, we constructed the Power Transmission Tower Object Detection (PTTOD) dataset, which covers multiple countries and diverse geographic environments. Experimental results on the public SRSPTD dataset and the proposed PTTOD dataset demonstrate the effectiveness of the proposed method. On SRSPTD, SBG-DDN achieves 74.3% mAP@0.5 and 35.1% mAP@0.5:0.95, outperforming existing state-of-the-art detectors.