Power towers in unmanned aerial vehicle (UAV) inspection images have the characteristics of multi-attitude and large-aspect ratios, which are difficult to accurately locate and distinguish towers with different degrees of inclination according to unique prior knowledge. To improve the incline detection accuracy, this article proposes a method for processing UAV images based on the improved refined rotation RetinaNet (Multi-Head-KF-R 3 det). Firstly, the incline detection head is added to the original R 3 det model to achieve the classification of power towers and their inclination degrees, as well as the accurate location. Then, the Kalman-filter intersection over union loss is introduced into the regression loss function to effectively improve the overall detection accuracy and recall rates of incline detection without additional hyperparameters. Finally, the improved model is reasonably compressed based on the design principles of ghost lightweight network, which lays a foundation for the deployment of the model in embedded devices. The experimental results show that the mAP and recall rates of Multi-HeadKF-R 3 det on multi-scale and multi-attitude power tower datasets can reach 94. 5% and 94. 9%, respectively.