Multi-size Target Detection In Transmission Line Images Using Deep Learning

As a key technology of smart grid construction,intelligent inspection of transmission lines is essential to ensure the stable operation of the power system.With the development of machine learning,it has become a mainstream trend to replace the traditional manual inspection with intelligent inspection based on deep learning and UAV aerial photography.However,in the actual inspection scene,the large-scale changes of inspection targets seriously restrict the detection performance of deep learning methods such as SSD and Faster R-CNN.To this end,this thesis proposes a research on multi-size target detection algorithm for transmission lines based on deep learning,and improves the method from the following four aspects,including colleting a transmission line inspection image dataset,constraint method of anchors,multi-scale feature fusion and Ro I Pooling.It is of great significance to improve the precision and adaptability of multi-size target detection.The main work is as follows:Firstly,we collect and annotate a target detection dataset for the task of intelligent inspection.The dataset contains 6 types of typical targets,a total of 33155 images,covers different target scales,different shooting angles and various scenes.It also includes common interference factors such as complex background,lighting and viewpoint changes.Secondly,to solve the problem that the preset anchor of Faster R-CNN doesn’t match the target size of inspection images which leads to the decrease in accuracy,we use K-Means clustering algorithm to learn the morphological distribution of different targets and initial the anchor boxes according to the results.Experiments show that our method can improve the adaptability of RPN to multi-size targets.Thirdly,aiming to enhance the poor presentation of multi-size targets,we propose a contextual feature enhanced pyramid method.This method consists of four steps: feature rescaling,contextual feature integrating,feature strengthening and feature remapping.The global scale feature with contextual information is obtained by weighted summation of features from different levels of pyramid.Then,the global feature is enhanced by multidimensional feature enhanced module and fused with high-level and low-level features of pyramid.Experiments show that,compared with the FPN structure,the feature map extracted by our method has more abundant information which significantly improves the accuracy of multi-size target detection.Finally,to solve the problem of high false detection rate caused by inaccurate positioning of multi-size targets,an improved object detection network is designed.In order to improve the quality of region proposals,we use Cascade R-CNN to gradually optimize the distribution of region proposals.In order to solve the problem of feature misalignment,we use a bicubic interpolation-based Ro I Pooling method.This method cancels the two quantization operations in Ro I Pooling and uses 16 integer feature points around the sampling point to calculate feature value of the sampling point by bicubic interpolation.The experiments show that,our results show better localization performance.Our method improves the accuracy and adaptability of multi-size target detection and breaks through the limitation of low accuracy of multi-size target detection,which provides a new technical approach for the development of intelligent inspection system for transmission lines.