Anormal Target Detection for Power Transmission Cable Drone Images Employing Improved YOLO10 and ESRGAN

Abstract

The anomaly target detection accuracy and speed in transmission cable drone inspection batch images with low pixel size is not enough high to make operation decision. To address this problem by reconstruction of image enhancing pixel, a super-resolution transmission cable target detection method based on improved YOLO10 is proposed. First, the Best-Buddy loss function is used instead of the commonly used one-to-one MSE/MAE loss function, allowing low pixel images to dynamically find the most suitable supervised image, providing more reasonable image details and reducing training difficulty. Secondly, a region-aware adversarial learning strategy is introduced, focusing on training texture-rich image areas, enriching the texture structure of images, making them more realistic, reducing artifacts, and improving their visual effects. Finally, the feature-driven super-resolution algorithm is applied to the object detection of power transmission lines, using YOLO10's backbone as a feature extractor and fine-tuning the super-resolution model to generate more machine-readable feature maps to improve detection accuracy and speed. The experimental results show that the proposed model improves detection accuracy by 6.62% compared to using only the YOLO10 object detection model and 1.22% compared to using ESRGAN as image pre-processing before using YOLO7 for object detection, achieving a detection accuracy of 92.55%.