Research On Multi-target Detection Of Transmissio Line Based On Improved FCOS Algorithm

In recent years,following the rapid advancement of UHV approval and construction measures by State Grid Corporation of China,digital new infrastructure has adjusted UHV construction to the national strategic level.The difficulty of transmission line operation and maintenance is increasing year by year,and the common goals and defects along the transmission line need to be maintained in time.Due to the constraints of multiple factors such as the geography and environment where the tension tower is located,the traditional manual inspection method has missed inspections and low efficiency.And other issues.With the continuous development of artificial intelligence,the identification of foreign objects in transmission lines through aerial drones has gradually become the trend of power inspection.In view of the missed shooting of traditional hand-controlled aerial photography drones,this paper uses cattle farming method to plan the transmission line route and realizes the autonomous patrol of the drone.The images taken by drones are tested on multiple targets on the transmission line through four algorithms,and the generated data is analyzed.The main work is as follows:Research on the autonomous cruise mode of aerial photography drones,using the GPS positioning system of the drones to use the cattle farming method to take fixed-point shots of the transmission lines.The dark channel method and Wiener filter are used to pre-process the image to achieve image defogging and de-jitter.Labelme used Labelme to mark the normal targets of 9 types of transmission lines and the defects of type 1 insulators,and make corresponding data sets for training and verification.In order to solve the problems of the current mainstream deep learning algorithms in the object recognition of small targets and overlapping targets,the improved Res Net101+FPN network is used to improve the receptive field of the convolution kernel and enhance the learning ability of the neural network.Soft-The NMS method recognizes the heavily occluded anchor frames,and compares the accuracy of the three types of algorithms: Cascade R-CNN,Grid R-CNN and D2 DET through experiments.Aiming at the problem of poor image recognition ability taken by drones in complex backgrounds,an improved FCOS algorithm is proposed.Vo VNet(a new backbone network for real-time target detection)is used as the extraction network for multi-target features of transmission lines,and the SE module is used to enhance Feature representation capabilities,and use the SE module on the feature map for weight distribution,making the depth features more diversified.The accuracy rates of these four algorithms are compared and analyzed.The test results show that the method proposed in this paper improves the accuracy rate by 5% on the basis of the D2 DET algorithm,and the m AP is 95.6%.This method can deal with small targets and various targets in a complex background.Class defects have good accuracy and have certain practical value.