Abstract: The cable is a significant carrier of power transmission. As such, it is susceptible to surface erosion due to environmental impact in a high-altitude environment, resulting in cable damage, reduced transmission efficiency, and in serious cases, electric shock accidents. Thus, it is very important to detection of the cable in time. At present, the mainstream method for detecting cable defects is the use of unmanned aerial vehicles (UAVs) to conduct inspections. UAVs are capable of rapidly capturing images of cables in complex environments. These images are subsequently transmitted to neural network models, which output the corresponding detection results. Due to its efficient object detection performance, the YOLO algorithm has been widely employed in UAV inspection tasks. However, surface defects on cables are generally small in scale, and the images acquired under low-visibility weather conditions at high altitudes tend to suffer from poor quality, resulting in low detection accuracy for UAV-based inspection systems. This paper proposes a novel defect detection model called YOLOv9-USSD, which is based on an improved version of YOLOv9, to address the dual technical challenges of image quality degradation in low-visibility environments and insufficient detection accuracy of tiny defects in UAV power inspections. Specifically, a defogging network (Unfognet) is integrated into the original YOLOv9 architectureto enhance the visual quality of images captured in low-visibility conditions. attention mechanism (SEAM) and a specialized loss function (Shape–IoU) are introduced to improve the model's ability to extract fine-grained features of small-scale targets. The standard convolutional layers (Original) in the original model are replaced with newly designed convolutional layers (DualConv) to further improve the recognition accuracy of the enhanced algorithm. To evaluate the proposed method, high-definition cameras and sensors mounted on UAVs were deployed at cable monitoring sites to collect a total of 1834 images depicting various types of cable surface defects, including breakage, thunderbolt damage, wear, and dark surface conditions. Subsequently, eight data augmentation techniques were applied to expand the dataset, resulting in a total of 9150 effective images. These images were divided into training (80%), validation, and testing (10%) sets. Experimental results indicate that the improved YOLOv9–USSD model achieves effective improvements in multiple key performance indicators compared to the original YOLOv9 model. Specifically, it improves the mean (mAP) by 3.5%, enhances the recall rate (R) by 5.6%, reduces the model size by 13 MB, and lowers the Giga Floating Point Operations per Second (GFLOPS) by 16 units. Moreover, compared with other mainstream detection models, including YOLO–7, SSD, Fast R–CNN, and RT–DETR, the proposed model shows improvements in mAP by 8.2%, 13.67%, 5.5%, and 10.30%, and in R by 3.1%, 20.78%, 5.3%, and 11.40%, respectively. Ablation experiments further demonstrate the effectiveness of each individual module. When the DualConv, SEAM, and Unfognet are used separately, the mAP reached 88.60%, 88.10%, and 89.20%, respectively. When all three modules are integrated, the mAP increased to 88.90%. The above improvements enable the model to maintain a stable detection rate under low visibility conditions, providing a new visual inspection solution for UAV cable inspection that combines high precision, light weight, and strong environmental adaptability.