Research On Aerial Image Classification Based On Deep Learning Models

In recent years,computer vision,especially deep learning,has been widely used in various fields.And with the rapid development of the electric power industry in China,the safety and reliability of transmission line is very important.In order to avoid the major accidents,regular inspection of transmission lines must be carried out.Traditional manual inspection has been gradually replaced by unmanned aerial vehicle inspection.A large number of aerial images are followed.The manual sorting method is no longer able to meet the needs of smart grid.The classification of aerial images by deep learning is imperative.This thesis studies the classification of aerial images based on the deep learning models.Firstly,the related algorithms of image denoising,image segmentation used in image classification process and several common deep learning models are studied,which provides the theoretical basis for the following work.Then,aiming at the problem that traditional manual design features have low classification accuracy,we extract deep features through deep residual network,and combine them with different classifiers.Experiments show that the combination of deep features and support vector machine is the best.A single classification model is not ideal,we use 50 parameter layers and 152 parameter layers deep residual network to extract features,respectively,training support vector machine(SVM).Then the weighted fusion is carried out according to the expressiveness of the two classification models,and achieve the initial classification of aerial images.Finally,Sub-classification of the most vulnerable insulators in the transmission line is carried out.In order to solve the problems that the traditional algorithms are easily affected by the environment,it is proposed that insulators are firstly detected by Faster-RCNN,eliminating the influence of the environment,then divide them into a single disc,feature extraction using the deep residual network,and finally sent into the sparse representation classifier for the insulator status classification.The experimental results show that the method in this thesis has a higher classification accuracy and extensibility.