| Abstrakt: |
Numerous approaches exist for unmanned aerial vehicles (UAVs) to perceive the surrounding environment. CMOS and charge coupled device (CCD) image sensors are commonly used in machine perception due to their cost-effectiveness and ease of acquisition. The use of image sensors for UAV power line (PL) inspection presents challenges due to the disorganized nature of the PL aerial images and the abundance of redundant background information. To address these challenges, we propose a novel approach that combines wavelet transform theory and deep learning methods. The image acquired by the image sensor is decomposed into low-frequency coefficients and high-frequency coefficients using wavelet transform, and the high-frequency coefficients are used to guide the low-frequency coefficients to generate a feature map of weak semantic information. This approach helps suppress complex and redundant background feature information, enhancing the detection of slender PLs. Furthermore, we incorporate an asymmetric dilated convolution global attention mechanism, guided by wavelet decomposition, to further enhance the features of slender PLs. This attention mechanism focuses on key regions and details, improving the PL detection process. Finally, we present a lightweight PL detection algorithm, WaveGNet, which combines deep learning and wavelet transform. Through extensive experiments on pinhole and fish-eye aerial PL datasets, WaveGNet achieves a remarkable trade-off of speed and accuracy, surpassing current lightweight segmentation models. This algorithm offers a novel approach for PL detection using image sensors and provides valuable insights for future deployment on UAVs. |
