Research On Multi-objective Detection Algorithm Of Transmission Lines Based On Improved Faster R-CNN

Patrol and maintenance of transmission line is related to the power grid can be safe and steady running,as target detection algorithm of gradually mature,to be used in electric power equipment for transmission lines to carry on the inspection,has become a key power inspection technology,using drones inspection not only saves the manpower,to ensure the safety of staff,but also greatly improve the inspection efficiency.Aiming at the difficulties and inaccurate accuracy of identifying small targets in aerial images of transmission lines by the traditional Faster R-CNN network model,a network architecture is designed in this paper,and the work is shown as follows:(1)In order to optimize Anchor parameters,reduce the Loss value of training network and accelerate the convergence speed of data set model.The k-means ++ clustering algorithm is used to class Anchor values of different scales for self-made data aggregation,so as to obtain the number and initial size of Anchor boxes with high fitness.The problems of NMS(non-maximum suppression)filtering anchor frame are studied and replaced by Softer-NMS.(2)In order to reasonably set network training parameters,the loss function is improved by introducing two weight factors M and N,WOA optimization algorithm was used to optimize the setting of learning rate and weight factor.The optimization process of learning rate is divided into two stages.Firstly,the optimization is conducted in the interval(0.0001,0.001)to quickly approach the global optimal solution.When the training iteration reaches a certain number of times,it enters the second stage and searches for optimization in the interval(0.00001,0.0001)to find the global optimal solution,which can speed up the training convergence and achieve a better effect.(3)In view of the difficulty of small target detection and low detection accuracy,a method of Faster R-CNN transmission line defect detection algorithm based on BAM(Bottleneck attention mechanism)is proposed.In feature extraction network,a certain channel attention computing module and a spatial attention computing module are added.Finally,the two modules are added together to form a BAM module.In BAM,the size of the input and output feature maps is consistent,which improves the significance of the target region of faults in the image.In order to further improve the recognition accuracy,FPN feature pyramid network is integrated to enrich the semantic information of the network.In order to verify whether the proposed algorithm has universality and application value,the algorithm is compared with other algorithms and good results are obtained.