Abstract: In order to realize fast and accurate detection of components and abnormal targets in power lines, a target detection algorithm based on the improved CenterNet is proposed. Firstly, the lightweight MobileNetV2 was used as the feature extraction network for CenterNet, and the number of channels in the decoding network was reduced, so as to improve the detection speed. Secondly, a multi-channel feature enhancement structure was constructed and the low-level detailed information was introduced to solve the problem of low detection accuracy caused by CenterNet only utilizing a single feature. Thirdly, an equal-scale residual attention feature fusion module was designed to replace the fusion method of directly adding features during the upsampling process, in order to fit the same level features from different branches. Finally, the elliptical Gaussian scattering kernel was introduced to optimize label encoding and improve the quality of bounding box regression. Experiments were conducted on the improved CenterNet algorithm. The results show that the algorithm achieves an average accuracy of 96% on the constructed dataset, a forward inference speed of 13 ms/frame, and a model parameter size of approximately 5.9 MB. All indicators are superior to mainstream detection algorithms such as FCOS and YOLOX. The combination of this method with drones can provide reference for intelligent inspection of power grids.