| The whole-pixel registrationof non-rigid image is to align all the pixels in spatial position by matching two images with non-rigid deformations.This thesis introduces the traditional image registrationalgorithms,as well as those based on deep learning,analyzes their advantages and disadvantages,and proposes theimproved algorithms on their basis.In order to achieve the whole-pixel registration of non-rigid images,this thesis proposes a Correspondence Vector Field interpolation method based on the sparse matching of feature seeds.First,according to the requirement of dense CVF interpolation,two types of feature seedswere detected and matched to realize the sparse feature matching.One is to ensure the accuracy of the estimated CVF on the motion boundary,while the other is to achieve a uniform distribution of seeds.And then,a CVFinterpolation method was proposed to achieve the dense CVF on the basis of the detected sparse seeds.Experimental results demonstrate that compared to the traditional CVF interpolation methods based on optical flow field,our method achieves higher accuracy due to implementing on the basis of sparse seeds.This thesis proposes an image registration framework based on deep learning.The proposed framework adds a self-attention mechanism and an adaptive resolution strategy to achieve the dense alignment of non-rigid images.Our method extracts feature using the VGG16 network,obtains the feature map with the lowest resolution of feature pyramid,and estimates the correlation of features through the self-attention layer.By combining the global and local features together,the large displacement deformations can be estimated.Finally,the optical flow fields obtained in each layer are optimized using the hole convolution,and then the fine optical flow vector fields can be estimated to achieve thewhole-pixel registration of non-rigid images.The experimental results on the general datasets demonstrate the effectiveness and robustness of the proposed algorithm. |
PDF Full Text Request