Object detection for key components is an important means to ensure safe operation of transmission lines. However, the accuracy and robustness of object detection need to be further improved due to the influence of the weather and flight attitude of unmanned aerial vehicles. To solve these problems, this paper presents an object detection scheme based on an improved CornerNet deep network for transmission lines. Considering the characteristics of flight attitude, the background is divided into two types, and their own data sets are established for network training. Subsequently, the network structure is improved based on CornerNet deep networks to enhance feature propagation and improve network performance. In the test stage, the background classification is carried out first with a support vector machine, and then, object detection is implemented with the trained model for each background. The test results show the validity and feasibility of the scheme with an accuracy of 88.1% for the sky background and 92.3% for the ground background, which are 6.7% and 10.9% higher than those for the original network, respectively. In addition, through experiments simulating severe weather further verify the robustness of the proposed scheme.