Research On Detection Algorithm Of Transmission Tower In Optical Image Based On Deep Transfer Learning

High-voltage transmission tower is one of the most important infrastructures in power transmission system.Its operation status determines the stability and safety of the whole power grid.Its detection and identification is the basis of high-voltage transmission tower operation status monitoring,and also an important part of highvoltage transmission lines monitoring.On the other hand,with the rapid development of optical remote sensing technology,the resolution and quality of satellite images are getting higher and higher.Wide-area monitoring of power transmission line operation using optical remote sensing images has become the development direction of remote sensing application research.However,the traditional detection and recognition methods are complex in the process of artificial feature design.Although the end-to-end detection and recognition model can be constructed by convolution network and deep learning method,it is usually difficult to meet the requirements of deep learning for quantity and labeling quality in practical applications.Aiming at this problem,this paper carries out the design of target domain small sample database and network framework,the research of detection and identification method of high voltage transmission tower based on deep transfer learning,and carries out experimental verification and analysis.The main research work is as follows:(1)Design of target domain database and network framework.Aiming at the problem that it is difficult to meet the requirement of deep learning for a large number of high-quality labeled samples in practical application,small real remote sensing data enhancement in target domain is realized by rotating and flipping the selected data sets,changing brightness and saturation,twisting and deforming,scaling etc.,and the requirement of data integrity coverage of the model is reduced.It provides necessary data support for subsequent detection and recognition model migration.On the other hand,combined with the basic theory of convolution neural network and the research and analysis of classical convolution network,the Inception feature extraction module and RPN module are selected to construct a target detection and recognition neural network based on multi-scale feature fusion to realize end-to-end detection and recognition of remote sensing images for small samples in target domain.In addition,the deep convolution neural network framework design provides basic support for the subsequent design of convolution layer parameters of high voltage transmission tower detection and recognition network based on deep transfer learning.(2)Research on detection method of High-voltage transmission tower based on deep transfer learning.Based on the design of network framework,the functional modules of deep network are designed,the parameters of each layer of network are given,and the loss function is determined.The hyper-parameters are set on the basis of summarizing the experience of classical network.Considering the efficiency of derivative calculation,convergence of model parameters and stability of network updating,SGD(batch gradient descent)and Momentum methods are combined to update parameters.In order to satisfy the requirement of similarity between target domain data and source domain data at the bottom of the model,DOTA dataset of remote sensing algorithm is selected as source domain data.(3)Experimental verification of detection algorithm for high voltage transmission tower.In order to realize end-to-end detection and identification and further reduce the requirement of model for data samples,a deep migration learning method based on Multi-task learning and fine-tuning is developed at the level of model parameters,and the method flow of migration learning is given.Based on the realization of detection and recognition method for high voltage transmission towers based on deep transfer learning,enhanced acquisition of test data from existing real remote sensing data,the performance indicators of detection and recognition accuracy,precision and recall under different conditions such as tower collapse,truncation and normal targets,different background types,different illumination,different size of targets and whether to transfer learning or not are analyzed by using confusion matrix.The experimental results show that the proposed detection and recognition algorithm for High-voltage transmission towers still has a high probability of detection and recognition under the condition of small samples in the target domain,which can provide scientific basis and technical support for the actual operation condition monitoring and alarming of High-voltage transmission lines.The network training is completed to realize end-to-end detection and identification of High-voltage transmission towers based on small samples in the target domain.