Research On 2-channels-conv Network Algorithm Of Remote Sensing Image Matching

With the development of remote sensing technology,the types of sensors are becoming more and more diversified In the field of remote sensing.Images can be provided by different sensors,different times or different resolutions.These images can be complemented by other sensors.The information of surface monitoring,feature recognition,image fusion,3D reconstruction,quantitative analysis,etc is provided by images.The use of deep learning methods to match images has strong robustness and is less affected by changes in noise,deformation,and illumination.Therefore,using deep learning methods to match images has important practical significance.Deep learning requires a large number of data sets.Usually these data sets are produced manually.This article proposes a method to automatically generate image matching data sets and train the data sets in a convolutional neural network.The specific content is as follows:(1)In order to reduce the manpower and time spent in making the data set,this paper automatically constructs the remote sensing image matching data set.Using ZY-3 multi-spectral images and down-view images,as well as GF-7 front and rear view images and footprint images,use feature point matching algorithm to match the images and crop them into image blocks,and propose the use of hash algorithm and similar triangles The algorithm determines the similarity between image pairs,so as to obtain positive sample pairs and negative sample pairs suitable for image matching.The experiment quickly generated the image matching data set,which has practical significance for the image matching method based on deep learning.(2)The number of layers of the 2-channels network used for image matching is shallow,and the deep features of the image cannot be learned.In order to deepen the network and enable the network to learn deeper data features without increasing network training parameters,this paper proposes to add a 1*1 convolutional layer to the network structure.Improve and propose a 2-channels-conv network on the basis of the 2-channels network.The experimental results show that the accuracy and recall rate of the algorithm in this paper have been greatly improved compared with the original network,and are significantly better than other algorithms.