SIFT Algorithm And Differential Homeomorphic Demons Algorithm Based On Fractional Differential

Image registration is a fundamental research topic in medical image analysis and computer vision.It has important applications in image fusion,image mosaic,and image reconstruction.SIFT(Scale Invariant Feature Transform)algorithm is a classic image feature-based registration method.The traditional SIFT-based image registration method has fewer matching pairs in the feature matching stage,and the registration accuracy is lower.Traditional Euclidean distances cannot measure the geometric space correlation of feature vectors.The differential homeomorphic Log-Demons algorithm is an important extension and development of the Demons algorithm.It is mainly used to solve the local small deformation problem of the image,and has achieved good results,but it has the same problem as the classic Demons algorithm: namely in the image The gray-scale uniform distribution area only uses the image gradient as the driving force,and there is a lack of obvious driving force,which makes it impossible to drive the floating image to deform to the reference image,and the algorithm is easy to fall into the local optimum value.In this paper,the above work is carried out as follows:(1)A SIFT image registration method based on adaptive fractional order is proposed.Firstly,the adaptive fractional differential is applied to the image feature-based SIFT image registration algorithm.When the fractional differential is solved for the optimal order,the image information entropy value and the gradient modulus are normalized,and then based on Two local features construct a fractional adaptive model,which can calculate the best fractional order of each pixel,and then construct an adaptive fractional mask template based on the optimal order.In the order selection,the order adaptive algorithm is adopted,which avoids the time and effort of manual selection.Then,the cosine similarity constraint is introduced for the spatial geometric constraint problem that the Euclidean distance cannot satisfy the feature vector in the matching process.The matching logarithm is reduced;the fine matching is performed by RANSAC,and the average value of the matching point to the transformation model distance is used as the threshold to further reduce the false matching rate.In the spatial transformation process,polynomial transformation isadopted to reduce the registration requirement.time.Through the analysis and discussion of the matching and registration experiment results,the improved algorithm is effective.(2)A non-rigid registration combining the SIFT features with the improved Log-Demons is proposed.Firstly,the reference image and the floating image are coarsely registered based on the SIFT feature.Then,for the problem of local micro-deformation in the registered image,the Log-Demons method of differential homeomorphism is used to achieve fine registration.Combined with feature-based and gray-based registration methods,the problem of image registration with larger deformation can be better solved.Enhance the image details by replacing the image gradient in the original differential homeomorphic Log-Demons algorithm with R-L(Riemann-Liouville)fractional step,improve the driving force of the registration,and avoid the algorithm falling into the local minimum,thus making the image The quality of the registration has been improved while reducing the time required for registration.Then an anisotropic Gaussian function based on image structure tensor is designed to regularize the deformation field,which makes better use of the structure information of the image neighborhood to avoid the missing gradient information in the calculation process.Finally,in order to reduce the registration time,the SIFT-registered image and the reference image are separately divided into three-layer images of different resolutions by down-sampling,and the low-to-high layer-by-layer registration is effectively improved.The quality and efficiency of registration,by comparing the algorithm and related algorithms,can be concluded that the quality and efficiency of the improved algorithm are effectively improved.