Visual inspection of transmission line defects by unmanned aerial vehicles based on convolution algorithm and deep forest network

The development of unmanned aerial vehicle technology provides a feasible solution for detecting transmission lines in complex terrain. How to effectively identify defects in power transmission lines in combination with unmanned aerial vehicle technology is the focus of research. The paper is optimized based on the traditional convolutional neural network, and a spatially deformable convolutional (SDC) algorithm is proposed to extract information from the images captured by unmanned aerial vehicles, and then features are fused based on the deep neural decision forest. At the same time, a more comprehensive hybrid loss function is proposed to further improve the model's recognition of multi-scale defect images. The instance verification is found that the model's recognition accuracy for the five types of normal, stains, cracks, corrosion and surface peeling is all above 92%. To determine the functions of each module, ablation experiments are conducted. When extracting image features using SDC, the mPA is increased to 90.48%. When the deep neural decision forest is adopted as the feature fusion method, the mPA is increased to 91.5%. When training with the loss function and then recognizing different images, the mPA is increased to 86.71%. By combining the three algorithms, in the detection of five different categories of transmission lines, the PA values are respectively increased by 7.86, 9.34, 8.95, 6.01 and 10.05. The mPA has been increased by 8.44%. The model is compared with several traditional recognition models. Compared with Faster R-CNN and YOLOv7, the detection speed is increased by 11.63 and 2.13 respectively, and the number of involved parameters is decreased by 54.57 and 37.37% respectively. Compared with YOLOv9, although the detection speed has decreased and the number of involved parameters has increased, the mPA has been increased by 2.34% through the increase in model complexity. The proposed model can effectively identify multi-scale transmission line defects.