Matching RGB And Infrared Remote Sensing Images With Densely-connected Convolutional Neural Networks

Image matching is the foundation of other applications of remote sensing images,and RGB(red,green,blue)and near infrared remote sensing image matching is one of the important research hotspots among it.The complementarity between visible and near-infrared images enriches information sources and improves the ability of scene understanding,so the combination of RGB and near infrared images is widely used in computer vision and image processing tasks.However,there are great radiometric and geometric differences between visible and infrared images,as they collect spectral reflectance from different wavelengths with different imaging mechanisms.The visual differences prevent successful application of conventional matching methods that heavily rely on intensity and gradient,which makes image matching between RGB and near infrared images more challenging.This paper develops a deep learning-based matching method between a RGB and an infrared image that were captured from satellite sensors.The main work and contributions are summarized as follows:(1)In order to enhance the matching performance between visible and near infrared images,this paper proposes an innovative densely-connected CNN structure,which consists of a series of densely-connected convolutions.The dense connections in several previous convolutional layers ensure information of lower features being directly passed to the higher layers,which can make full use of low-level features and significantly improve the performance of image matching between different spectral bands.(2)In order to improve the learning ability and stability of the network,this paper proposes an enhanced loss function to optimize the training procedure of denselyconnected CNN method.The smoothing term in the proposed enhanced loss function ensures the effective learning of network models and significantly improves the generalization ability of method.(3)Contrast to the recent CNN structures that are only designed to compare visible and infrared images,a complete CNN-based template matching framework for optical and infrared images is introduced.Through replacing the feature-based matching scheme to a template matching scheme,a large number of correspondences can be found.In order to verify the transfer ability and versatility of proposed method,our method directly extended to other matching tasks such as heterogeneous image matching,multitemporal image matching and close-range image matching.We show that our method is effective and outperforms all the other conventional and CNN-based methods on various satellite images with different geometric distortions as well as on close-range images.