Research On Power Line Facility Detection Using Virtual Scenes

With the development of intelligent inspection of power lines,the use of UAVs to collect images is increasingly replacing the artificial climbing tower survey,so the corresponding image data on the power equipment is also increasing.However,due to the Characteristic of the strong professional application,the narrow use range,therefore there has no open and relatively well-marked image data set of the power equipment.In this paper,a series of the iterative incremental learning methods based on combining the virtualized sample and small amount of artificial marker sample are proposed to detect and check the vibration hammer as the main target.Good results are obtained.And the method of this article has a good portability and can be easily applied to other power scenes and other power components.The main work of this paper includes the following contents:1)Based on the analysis of the characteristics of the power scene and the shooting mode of UAV,a set of methods are proposed with the characteristics of relatively easy to build virtual scene,rapid virtual samples generation and virtual samples labeling method.2)For the problem of the initial annotation generation,an annotation generation method is proposed with only using virtual data and based combined detector using geometrical constraint(GCAD).For the characteristics of virtual data and various recognition methods,the Faster R-CNN,DPM and Haar-like cascade classifier are integrated together.This method,combined with a small amount of manual assistance,can significantly reduce the workload of data marked.3)Combining the characteristics of virtual data and real data,a manual auxiliary annotation program based on GCAD has developed,this method makes full use of the rich terrain information of virtual data and real image data.Using iterative Faster R-CNN training method to joint train the virtual data and real data,the experiment shows that better results can be achieved.