Image Registration And Denoising Of Multi-Frame Ultrasonic TOFD Images Based On Homography Estimation

Image registration is a very important part in the field of image processing and pattern recognition.Image registration is widely used in many fields,such as target detection,model reconstruction,motion estimation,feature matching,tumor detection,lesion location,angiography,geological exploration,aerial reconnaissance and so on.Each registration methods are usually based on a specific design problem,in many ways,the only common is that each registration will eventually find a some of the most in transform space,this transform can achieve in a sense matching between two images,but for different applications,different the type of image,we need to analyze specific issues.According to the relationship between the images to be registered,the image registration can be divided into four categories:multi-source image registration,template based registration,multi angle image registration,time series image registration.In this thesis,we mainly discuss the image registration algorithm based on mutual information,which is based on the mutual information.The main research contents are as follows:In this thesis,we discuss and Research on improving the SNR of the image.Introduces and analyses the relevant technology,wavelet image denoising algorithm includes:the image data by wavelet multi-resolution analysis,this thesis adopts Daubechies32 wavelet for each signal in the image by layer n decomposition in scale space(n?2),the formation of high-frequency wavelet coefficients and low frequency approximation of wavelet coefficients.Since the noise is mainly in the high frequency coefficient,the high frequency coefficients are extracted to maintain the low frequency coefficient.Then,a new threshold function is proposed to compress the wavelet coefficients.The curve l2-risk shows that the threshold function has better effect.Then,the high frequency coefficients and low frequency coefficients are processed by inverse wavelet transform.Finally,the image registration technology based on gray value is used to eliminate the deviation in the process of data acquisition.Compared to the classic method,SNR is improved with 12.2 dB,about 99%improvement.This thesis uses the method of homography estimation for image registration.In this thesis,the calculation method of homography matrix is discussed and analyzed.The SIFT algorithm and SURF algorithm are adopted to extract the keypoints and feature descriptors to generate the corresponding matching relationship between the template image and the floating image.Then the DLT and Bundle Adjustment are used to compute the homography matrix,respectively.In order to improve the quality of the TOFD images,we adopt image registration based on features.Firstly,DoG operator is used to extract the feature points in the image.The matching relationship between the template image(fixed image)and the image(float image)is calculated based on the principle of minimum Euclidean distance,and then use RANSAC algorithm to optimize the correspondence model,which remove some of the false matching points between two images,thereafter,the homography matrix is obtained.Finally,the homography matrix is used to complete the image registration.The image registration based on homography matrix can be used for image expansion,rotation,and so on,and it has higher registration accuracy compared with the image registration based on the maximum correlation method.Cmpared to image registration based on maximum correlation,SNR is improved with 2dB,about 10%improvement.This thesis discusses an image registration methods via multi local homography transform.We propose a new method of image registration based on AGAST-DAISY feature point extraction and red black tree matching algorithm.Firsly,the AGAST operator is used to extract the feature points in the image.Due to the AGAST operator can extract a large number of feature points,the DAISY descriptor is used to select the feature points,because the DAISY descriptor has good performance for the dense distribution of feature points.In this thesis,the principal component extraction algorithm based on the three-order statistics is used to reduce the dimension,and then the minimum Euclidean distance principle is used to match the feature points between the template image and the image to be registered.In order to improve the randomness of the traditional RANSAC algorithm,this thesis adopts the method of cluster consistent principle to improve it.Firstly,the number of clusters is selected by the principle of Jeffrey divergence and Gap statistics,and the computer simulation of a variety of experimental data shows that the two methods have the same number of clusters.After determining the number of clusters,the C-means fuzzy clustering method is used to form the C clusters in the template image and the image to be registered.Then,the maximum entropy model is used to select 4 points for each cluster,and the red black tree data structure is used to match.By the homography transformation of the corresponding cluster.Thus,the local registration of the image can be done,and effectively eliminate the influence of the randomness of the traditional RANSAC algorithm.The registration accuracy is improved.To the TOFD images,with the imcresing of the number of blended images,SNR is improved with 1dB,RMSE is reduced about 0.01,about 85%of the value obtained via the classic method.To Saturn images,SNR is increased with 10dB,and CNR is increased with 10,while RMSE is reduced with 0.01,about 50%.To Mars images,SNR is increased with 5dB,CNR is increased with 45,and RMSE is reduced with 0.02,about 60%reduction.