Research On The Recognition Methods Of Insulator In Aerial Images Based On Multi-deep Model

Aiming at the current situation that insulator recognit ion accuracy is low and lack of effective feature representation in aerial images under unconstrained environment,apply the deep learning theory,study the transmission of the insulator features in the deep network deeply,and propose a series of insulator recognition methods based on image features of deep convolut ional neural network.The main work of this thesis is as follows:Aiming at the problem that the recognit io n of insulators in infrared images is seriously affected by the change of angle and the recognit ion accuracy is not high under the condition of no constraint,propose a rotating insulator recognition method based on the parallel deep convolutional neural network feature extraction and feature dimension selection,and improve the accuracy of mult i-angle insulator recognition.Collected a large number of valuable scene inspection image data,and amplify the infrared insulator image recognition dataset.The dataset contains the change of angle,scale after amplification.Focusing on the feature representation methods of the infrared insulator images,propose an infrared insulator image recognition method based on deep convolut ional feature map aggregation,realize the robust feature representation by extracting the convolutional feature map for feature descriptors,and improve the discrimination of features and the accuracy of insulator recognit ion.The feature representation method is embedded into the detectio n framework,and it further proves the effectiveness of this method.On the basis of the feature representation methods of aggregating deep feature map,explore the response and distribution of neurons in convolutional layers,mine fully the powerful features of generalization and characterization,propose an insulator recognition method of feature aggregate layer selection based on the layer entropy and the relative layer entropy,for mining the most suitable feature representation layer for our insulator image recognition task.Study the influence of convolution layer width on recognit ion accuracy,use the selection method based on the important degree to select the number of the feature map within layers,reduce the computational complexity,and increase the accuracy of insulator recognit ion to 99% on the existing insulator dataset.