Research On Transmission Line Birdhouse Detection Based On Deep Learning

Transmission lines are an important part of the power system.With the gradual improvement of the ecological environment,birds are often active in some places,which leads to the occurrence of many bird damage failures,and bird damage failures are one of the main reasons for transmission line failures.Therefore,in order to ensure the safe and stable operation of the transmission line,it is necessary to conduct regular inspections of the line bird nests.The traditional manual inspection method wastes a lot of people and things.With the continuous development of technology,drone inspection methods have emerged.The method of drone patrol is flexible and maneuverable,which improves the efficiency of line patrol,but for line bird nest detection,it still needs to manually judge and mark one by one.In order to efficiently perform line bird nest detection,deep learning methods can be used.In view of the above problems,this paper has done the following work:(1)Aiming at the lack of bird’s nest data set of transmission lines,the collected bird’s nest data set is expanded.The datasets collected in this paper cover a complete range of lines and are well representative,but the number of each type is small.In this regard,this paper uses the traditional method to expand the data set,by transforming left and right,flipping up and down,and adjusting the brightness for data expansion.In order to improve the detection efficiency,under the premise of not affecting the detection performance,the compression processing of the data set is studied.At the same time,since the data set is self-built,this paper uses the Label Img software tool to manually label the bird’s nest pictures.(2)In order to improve the detection efficiency,this paper compares the bird’s nest detection performance of the two target detection networks based on YOLOv5 s and YOLOv5 m and using different data sets to train the networks.Firstly,the network structure and training process of YOLOv5 s and YOLOv5 m are studied in detail,and the detection models are trained by using the two networks respectively,and the experimental data are compared and analyzed using the evaluation indicators.It is verified that both YOLOv5 s and YOLOv5 m networks can achieve high-performance line bird nest detection.Secondly,the model performance of the compressed image is compared with that of the uncompressed image.The experiment shows that the training time of the compressed image training model is reduced by 97% compared with that of the uncompressed image training model,and the detection time is also reduced,while the detection performance remains basically unchanged.Finally,the network is trained with the enhanced data set,and the performance is compared.The experimental results show that the m AP value of the model trained with the enhanced data set is increased by about 4.7%.