Bird Nest Identification On Transmission Lines Based On Improved SSD Algorithm

With the improvement of the national economic development level and the gradual increase of the scale of the city,the vegetation in many places has been destroyed,resulting in many birds nesting on the transmission line,which has caused great hidden dangers to the safe operation of the power grid.Detecting bird’s nests from overhead transmission lines is difficult to identify manually because the target of bird’s nests is too small.If the human eye is used for recognition,the recognition is prone to errors and the number of people required for human eye recognition increases,which leads to increased labor costs.Therefore,it is necessary to use a model based on a deep convolutional network to detect the target with the help of a drone,so that it can be Save labor costs and improve the efficiency of bird nest recognition.In this paper,an improved algorithm is proposed for the identification of bird’s nests in aerial images of transmission line inspections to achieve more automatic and high-precision automatic detection.And the position of the bird’s nest is located by the GPS information carried by the pictures in the bird’s nest data set.First,more than 6,384 high-definition bird’s nest pictures on transmission lines with GPS information were collected by UAV equipment,and then the pictures were screened to a certain extent,and 6,000 pictures suitable for training and verification were selected as the original data set.Secondly,in order to automatically identify the bird’s nest on the overhead transmission line to reduce labor costs,a deep learning method is chosen for object detection.Then,the commonly used target detection models are analyzed,and the SSD model is used as the basic model.However,the SSD model also has certain shortcomings in this experiment,because the target to be detected in this paper is the bird’s nest on the overhead wire,and the target features are not very obvious and relatively small,so the SSD model needs to be improved.The skeleton network of SSD is VGG16,with only 16 layers,and the convolution depth is not enough,so VGG16 is changed to VGG19 first,which is equivalent to adding more convolution layers on the original basis,so a better learning effect is obtained.After a lot of training on VGG19,it was found that the improvement effect was not obvious,so I tried to add residual blocks to further improve VGG19,and then used transfer learning to select Res Net50,which is similar in structure to VGG19 and composed of residual blocks.In order to make Res Net50 more suitable for the task of small target detection such as Bird’s Nest,the loss function of Res Net50 is further improved,and the calculation of center loss is increased,and the optimal model Res Net50(up)is obtained.Finally,Res Net50(up)is compared with the three SSD models based on different skeletons before improvement.The experimental results show that the improved network in this experiment has better effects on the comprehensive indicators F1 and AP.The F1 value of the improved model increased from87% to 89% before the improvement,an increase of 2%;the ap value increased from 86.23% to89.02%,an increase of 2.79%;the average detection accuracy(map)value increased to 89%,It can be seen that the improved SSD algorithm has better performance and can better identify the location of the bird’s nest in the picture with GPS information.The improved model can better identify the bird’s nest in the aerial image of the transmission line,and provide help for the identification of the transmission line’s bird’s nest.