Under the background of continuous breakthroughs in the spatial resolution of satellite remote sensing technology, high-resolution remote sensing images have become a frontier data source for intelligent inspection research of power infrastructure. To address existing issues in remote sensing image application algorithms such as difficulties in power target feature extraction, low detection accuracy, and false positives/missed detections, this paper proposes the YOLOv9-GDV power tower detection algorithm specifically for power tower detection in high-resolution satellite remote sensing images. Firstly, under high-resolution imaging conditions where transmission tower features are prominent, a Global Pyramid Attention (GPA) mechanism is proposed. This mechanism enhances global representation capabilities, enabling the model to better understand object–background relationships and effectively integrate multi-scale spatial information, thereby improving detection accuracy and robustness. Secondly, a Diverse Branch Block (DBB) is embedded in the feature extraction–fusion module, which enriches the feature space by enhancing the representation capability of single-convolution operations, thereby improving model feature extraction performance without increasing inference time costs. Finally, the Variable Minimum Point Distance Intersection over Union (VMPDIoU) loss is proposed to optimize the model’s loss function. This method employs variable input parameters to directly calculate key point distances between predicted and ground-truth boxes, more accurately reflecting positional differences between detection results and reference targets, thus effectively improving the model’s mean Average Precision (mAP). On the Satellite Remote Sensing Power Tower Dataset (SRSPTD), the YOLOv9-GDV algorithm achieves an mAP of 80.2%, representing a 4.7% improvement over the baseline algorithm. On the multi-scene high-resolution power transmission tower dataset (GFTD), the algorithm obtains an mAP of 94.6%, showing a 2.3% improvement over the original model. The significant mAP improvements on both datasets validate the effectiveness and feasibility of the proposed method.

Drone-based image capture with deep learning techniques has become a prevalent approach for inspecting transmission lines, allowing for efficient detection of defects and anomalies. However, most existing algorithms rely on single-modality information, failing to fully exploit the textual modality inherent in labels or the unique characteristics of inspection images, such as sharply focused foregrounds and defocused backgrounds. To overcome these limitations, this paper proposes Vision-Language and High-Frequency Features DETR (VLF-DETR), a detection model that leverages a multi-stage training strategy within the Deformable DETR framework. In the first stage, to address the absence of domain-specific knowledge in the general-purpose vision-language model FLAVA, we fine-tune FLAVA to learn both textual and visual features relevant to the power industry. In the second stage, to better incorporate textual modality, we introduce conditional queries into Deformable DETR, effectively transferring knowledge from FLAVA into the defect detection model. In the third stage, leveraging the structural characteristics of inspection images, we apply a fast Fourier transform (FFT) to extract high-frequency edge features, suppress background noise, and provide spatial priors. Furthermore, an FFT-based loss function is introduced to further ensure the model to converge on target regions. Experimental results demonstrate that VLF-DETR significantly outperforms baseline methods, offering a novel and effective solution for transmission line defect detection.