Research On Infrared And Visible Image Registration Based On Hypercolumns Of Substation Equipment

With the expansion of the scale of the power grid,more and more substation equipment is needed.In order to get the operation status accurately,a lot of infrared and visible images will be acquired during the maintenance process.Infrared and visible images can collect different information of the image.The registration between them can get richer complementary information of the image,which is conducive to the fault detection of substation equipment.Because the infrared and visible images come from different time and angle,and there is a great difference of grayscale,the traditional image features can not finish the task well on time.In this thesis,the image of the infrared and visible images of the substation equipment is taken as the object,and the registration method based on the depth feature is studied.The main work of this thesis is as follows:First,the local invariant feature theory in registration and the typical algorithms SIFT and SURF are studied,and the experimental analysis is carried out.In view of the problem that traditional features can not effectively realize multimodal image registration,a new method of infrared and visible image registration based on hypercolumns is designed.In order to solve the big gray difference,it is difficult to obtain consistent characteristics,in the feature extraction of key points to describe when using convolutional neural network,can get more comprehensive expression characteristics.At the same time,a similarity measure is designed to reduce the range of registration and improve the accuracy of registration.Compared with the traditional method,the method is more than 50%,and the mosaic result of the image is better than the traditional method.The experiment shows that the convolution neural network can better express the characteristics of infrared and visible images.Aiming at the problem that the VGG16 model based on initial weights can not learn the similarity between infrared and visible images,this thesis proposes a training method based on Triplet loss based on the super column.Network training is carried out by restricting the similarity between infrared and visible images to better express the conformance characteristics of the two images.In contrast experiments,the proposed algorithm achieves an average increase of 26% in accuracy of matching points compared with the method based on initial weights,and achieves better results in image mosaics.The experimental results show that this method can accurately register multi-modal images.In this thesis,a new method of image registration based on end to end learning is adopted in this thesis.In order to solve the model can only enter a target training,not for infrared and visible images at the same time training problem,this thesis designs a kind of training method.This method can complete the training of the two kinds of modes images,consistent their characteristics so as to better learning,and get more precise parameters of the transformation.The experimental results show that the method has good effect on multi-modal image registration.