Deep Learning Feature Representation Method For Registration Of Infrared And Visible Images

Image registration as a key technology in image application research,which is a process of aligning two or more images of the same scene obtained by different time,different viewing angles or different sensors in geometric space.The accuracy of image registration directly affects the effectiveness of processing tasks such as image fusion,image stitching,visual navigation and target recognition.It is a very challenging task of different imaging modes of sensors and different sensing environments on the registration of infrared and visible images.Aiming at the problems of large geometric changes(such as scale,rotation,translation,distortion)and poor contrast consistency between infrared and visible light images,this topic mainly studies the deep learning feature representation method of infrared and visible light image registration.The contents and innovations are as follows:(1)A learning method based on feature extraction of affine covariance region is proposed for image geometric changes,illumination changes,and noise changes.First implements affine parameter estimation using Affnet model,and then affine covariance region are normalized to obtain stable affine covariance region features.The algorithm is robust to affine transformations such as image scaling,rotation and shearing.Experiments show that the features extracted by this method have stronger affine covariance and can better adapt to different affine transformations that may occur,alleviating the problem of matching errors due to inaccurate calculation of feature descriptors in traditional feature extraction methods,and effectively improving the accuracy of infrared and visible image alignment under geometric changes.(2)On account of the problems of large geometric changes and poor contrast consistency between infrared and visible images,the method focuses on infrared and visible image registration based on modal conversion.First reduces the multimodal registration problem to a single-modal registration problem using the COMIR model,to reduce the difference between infrared and visible images,and then performs the registration on similar modalities.Then,the Affnet model is used to extract features from the affine invariant region;finally,the kernel correlation filter KCF feature matching method is used to realize the robust matching of the extracted features,and thus realize the registration of infrared and visible images.Experimental results show that this method can effectively register infrared and visible light images with large geometric changes and contrast differences,which has strong adaptability to affine transformation of images.