Method Research Of Multimodal Remote Sensing Images Matching Based On Convolutional Feature Representation

With the development of remote sensing technology,obtaining the required information from remote sensing images has become a very important means of information acquisition.Different satellite sensors can provide multi-spectral,multi-temporal,multi-resolution and multi-modal remote sensing images for the same regions.Multimodal remote sensing images(optical,infrared,SAR,etc.)reflect the different attribute information and characteristics of the ground objects.They can make up for the lack of a single image source,and improve the image information,and provide a variety of information that complements the surface monitoring,law enforcement inspectors,etc.The comprehensive use of multi-modal remote sensing images could enable a more comprehensive,more thorough and understandable analysis of the target area.Image matching is a key pre-processing link for many remote sensing applications and analysis such as image fusion,image mosaic,feature recognition,change detection,3D reconstruction,and remote sensing quantitative information analysis.The matching accuracy will have a significant impact on many subsequent remote sensing applications.Therefore,it is necessary to match images strictly before using multimodal remote sensing images for data analysis and processing.However,Due to significant non-linear radiometric differences between multimodal remote sensing images,the same region often presents completely different gray-scale information,which makes the matching of the same-named points very difficult.Therefore,high-precision matching of multimodal remote sensing images is still very challenging.Remote sensing image matching methods can be divided into region-based matching methods and feature-based matching methods.The former is sensitive to grayscale characteristics and is susceptible to noise.The latter is often subject to feature repetition rates and affects matching efficiency.Currently,deep learning technology has been widely used in various fields of remote sensing image processing.Based on the technical framework of remote sensing image template matching,this paper focuses on the use of convolutional neural network to extract features for image matching.The paper is extracted the common features from the image information itself and automatically trained the optimal convolution kernel based on the image structure information for exploring the multimodal remote sensing image matching methods with strong adaptability and high matching precision.The main research contents and results of this paper are as follows:(1)Focusing on remote sensing image matching,this paper systematically discusses the relevant research progress at home and abroad,analyzes the current research status and existing problems of multimodal remote sensing image matching,and summarizes the current remote sensing image matching strategy based on convolutional neural network.The feature detection repetition rate of multimodal remote sensing images is experimentally demonstrated.Experiments show that due to the significant nonlinear radiation difference,micro geometric distortion and image noise between multi-modal remote sensing images,the feature repetition rate between multimodal images is lower,which has a greater impact on image matching.(2)The Siamese network model in convolutional neural networks has great advantages in image similarity detection research.The network model has deep feature extraction ability for image pairs,and is considered to be an efficient deep network.It can learn the common features between images through deep network learning,improve the similarity detection performance of images under different modal conditions,and effectively resist the significant nonlinear radiometric differences between multimodal remote sensing images.To address that,the Siamese network is applied to multi-modal remote sensing image matching.The network is first optimized to effectively extract the common features between images.Then,the template matching strategy is adopted to achieve high-precision matching of multimodal images.The experiments show that the proposed method outperforms traditional template-matching methods in both the matching correct ratio and matching accuracy.(3)The structure features of the image have good anti-interference to the slight geometric distortion,speckle noise and radiometric differences between multimodal images.Structure information is robust between multimodal images.This paper studies image matching methods based on image structure features.Firstly,the convolutional neural network is used to extract features from multi-orientated gradients,and the convolution orientated gradient features are constructed.Then,the convolution orientated gradient feature based on image structure information is extracted to achieve high-precision matching of multimodal remote sensing images.The experimental show that the proposed method can effectively extract image features based on multi-orientated gradients,and the convolutional orientated gradient features can improve the matching accuracy of multimodal images based on image structure information.