Energy Reports (Sep 2023)
Semi-supervised learning-based satellite remote sensing object detection method for power transmission towers
Abstract
It is well known that as the power grid becomes more and more complex, the traditional manual survey of transmission towers is inefficient and cannot meet the requirements of safe and stable operation. Although the development of the satellite remote sensing technology provides a new prospect for the efficient and stable survey of transmission towers, there are still many problems that need to be solved. Due to the harsh climate and limitations of the imaging equipment, some of the transmission tower objects in the remote sensing images are blurred, which makes it extremely difficult to generate datasets and to achieve high precision transmission tower object detection. To further develop the detection precision of transmission towers, the image enhancement algorithm based on the dark channel priori is firstly applied for remote sensing images, which can boost the image interpretability. Then, considering that there are still some transmission towers in the enhanced images that cannot be manually labeled, a pseudo-label-based semi-supervised learning method is employed to maximize the use of the existing data. Based on this high-quality dataset, a satellite remote sensing object detection model for transmission towers is constructed using mobile inverted bottleneck convolution and deformable convolution. Finally, the ablation and comparative experiments are conducted according to the satellite remote sensing image dataset of a certain area in China. The experimental results indicate that both the image enhancement and the semi-supervised learning methods can improve the detection precision, and compared with the existing mainstream model, the proposed method performs better.
