Reasearch On Image Registration Method Based On Deep Learning

Image registration is to establish the corresponding relationship between images of the same scene,and has a wide range of applications in the fields of computer vision,medical image processing,material mechanics,and remote sensing.Homography estimation is a key issue in image registration tasks.Due to the geometric distortion of the actual imaging system,the linear affine transformation model is not accurate,and the matching pair of point coordinates constitutes a contradiction equation,so the traditional method is not reliable for homography estimation.Deep learning extracts the inherent laws and multi-scale high-dimensional features of large samples,and fits more reliable estimation models through data-driven methods.In deep image homography estimation tasks,illumination change,lack of labels in actual data,and the complexity of deep network models all are the challenges.Aiming at the illumination images,a new cascaded feature module is constructed on the basis of asymmetric convolution pairs.This network unit increases the feature dimension and increases the ability to extract features of illumination images.Aiming at the problem of high parameter complexity of the previous network fully connected layer and slow training convergence,it is proposed to use a joint attention mechanism network to replace the fully connected layer,use the attention mechanism to automatically learn the coordinate offset equation,and reduce the parameters by removing the fully connected layer which improved the fitting accuracy of the deep registration network.The real images have problems such as illumination,multiple perspectives,lack of labels,etc.Facing these problems,a weakly-supervised network model is proposed based on the above network.Using traditional unsupervised algorithms to estimate the homography of unlabeled image pairs to train the network.It uses a composite loss function constructed based on the coordinate offset of the image pairs and the reprojection pixel difference.Through the above method,the weakly supervised learning based on the unlabeled real images is realized.Experiment uses self-generated datasets which generated from MSCOCO and ICDAR 2019-LSVT datasets,and illumination and Hpatches datasets.The robustness of the above models are verified.Compared with mainstream depth image matching algorithms,the methods are less complex and more accurate.