Research On Dental DR Image Registration And Stitching Technology

In oral imaging,the use of a small field of view flat-panel detector combined with a linear scanning strategy can solve the problem of repositioning when performing lateral head imaging in cone beam CT.Due to the narrow visual field and wide dynamic range of grayscale,the accuracy of classical image registration algorithm is low,which directly affects the stitching quality of lateral cephalic panoramic images.In this thesis,from the perspective of accurate prediction and seamless fusion of registration parameters,a local normalized cross-correlation registration algorithm based on the multi-resolution hybrid optimization strategy is proposed.Meanwhile,the fade in and fade out fusion algorithm is applied to the sequence image Mosaic processing,finally achieving high quality cephalic lateral panoramic image Mosaic.The main work is as follows:1.The characteristics of the sequence image are analyzed,and the Block-Matching and 3D Filtering(BM3D)is used to deal with sequence images with strong noise.Through experimental comparison and analysis,it is verified that BM3 D algorithm can remove quantum noise and impulse noise to the maximum extent while retaining image information.2.In order to remove the influence of irrelevant regions on registration,a similarity measurement function based on improved normalized cross-correlation is proposed.This method combines the Gaussian mixture model and the expectation maximization algorithm to realize the region segmentation of the sequence image,and the normalized crosscorrelation of the region of interest is used as the standard to measure the degree of registration.The experimental results show that the measurement curve peak of the similarity measurement function proposed in this paper is sharper and the registration accuracy is higher.3.Aiming at the problem that the particle swarm algorithm has low optimization accuracy in registration,a hybrid optimization strategy based on multi-resolution is proposed.The wavelet decomposition is used to form a multi-resolution pyramid of the image,combined with a hybrid optimization strategy to achieve accurate registration of the image from low resolution to high resolution.The experimental results show that the error fluctuation range of the registration algorithm proposed in this paper is within 1 pixel,and the registration accuracy reaches the sub-pixel level.4.Designed a fully automatic stitching algorithm for sequence images.The overlapping area of the registered images is merged gradually in and out,and a panoramic image is formed after multiple rounds of stitching.Comparative experimental analysis shows that the panoramic image obtained by the stitching algorithm in this paper performs best in the parameter evaluation of information entropy,standard deviation,and average gradient,and there is no stitching error.In this thesis,a local normalized cross-correlation registration algorithm based on multi-resolution hybrid optimization strategy is proposed,the experimental results show that the proposed method can achieve sub-pixel registration for the oral skull sequence images.Combined with the fade in and fade out fusion algorithm,the obtained lateral cephalic panoramic image has complete information and no stitching gap,which is of great significance for oral imaging.