Detection Of Power Lines On UAV Images Based On Convolutional Neural Network

With the rapid development of micro unmanned aerial vehicle(UAV)technology,the use of UAV equipped with optical camera for electricity corridor inspection is getting more and more attention in power industry,automatic UAV-based power line inspection system is becoming a hot research issue in this area.Power line detection plays an important role in UAV-based electricity inspection system,which is crucial for real-time motion planning and navigation along power lines and has significant value in research and broad prospects in industrial applications.In this thesis,power line detection methods based on the fusion of multiscale convolutional features are deeply investigated from the perspectives of boundary detection and semantic segmentation.The main contributions of this thesis can be organized as follows:Firstly,power line datasets for UAV platform was created.Data is a crucial element in supervised learning-based algorithms,but large and public power line datasets are scarce in the community.In the thesis,two power line datasets of urban and mountain scene are created and released,which have abundant samples and nice pixel-wise annotations.Besides,scientific evaluation indexes are set for the two datasets,which provides reliable data and evaluation strategy for deep learning-based power line detection methods.Secondly,a power line boundary detection method with convolutional features and structured constraints is proposed.The trimmed VGG16 is adopted as base network of the method,which captures multiscale convolutional features of power lines.The coarse-to-fine features are fully exploited to obtain power line boundaries,the whole network is trained with deep supervision to effectively fuse the features.Then structured priors of power lines are utilized to refine the fusion results,which improve the accuracy of power line detection significantly.In addition,the positives and negatives are biased in the datasets,a class-balanced cross-entropy loss function is proposed for the network.Finally,a power line semantic segmentation method with multi-level and structured features is proposed.The method adopts fine feature maps extracted from shallow convolutional layers to refine coarse feature maps from adjacent layer,which achieves structured fusion of low-level local features and high-level semantic features.Then improvements are made to residual convolution module,multi-resolution fusion module and residual pooling module,which significantly reduce parameters in the network.The modified network conducts power line semantic segmentation with both high precision and high efficiency.Two power line datasets of urban and mountain scene are used in network training and testing for evaluation of different methods.The experimental results on the two datasets demonstrate the effectiveness of the proposed methods in the thesis.