Non-rigid Image Registration Method

Image registration techniques are an important research in image processing, which are widely used in military, remote sensing, medicine and other fields. According to the deformation nature of the image, image registration techniques can be divided into rigid image registration and non-rigid image registration. Besides, rigid registration techniques had matured after years of research. And non-rigid registration techniques still need further development, have more wide application foreground, are becoming a hot topic in the field of image registration research in recent years. Based on those theories, this paper has done a thorough study of the non-rigid registration methods based on the optical flow constraint equation, and have obtained the following results:1,Have studied the Horn-Schunck and the Lucas-Kanade optical flow models, and have applied them to register image. Then, compared and analyzed their registration results to obtain the two models'advantages and disadvantages.2,A systematic study has been made on Demons and Active Demons registration model based on diffusion theory, then, respectively applied them to register large deformation and small deformation images. Experiments proved that Demons model which gained the displacement field only by the moving image gradient is easy to produce registration errors, and the improved Active Demons model which overcame Demons model's deficiency can solve the problem of large deformation image registration.3,This paper proposed a curvature driven force to improve Demons and Active Demons models, established a new G&C model for small deformation image registration and Active G&C model for large deformation image registration. Experiments show that these two improved models have achieved higher precision results in image registration.4,This paper combined the six gray evaluation methods----mean square error, peak signal to noise ratio, correlation coefficient, normalized mutual information, structural similarity and offset curves to evaluate the quality of image registration with Demons, Active Demons, G&C and Active G&C models.This paper mainly researches on the non-rigid image registration methods based on diffusion theory and the quality evaluation methods of registered images. Experiments have proved that:as the combination of gradient and curvature driven, whether the G&C or the Active G&C model are better than other non-rigid registration methods, and have achieved higher precision results in image registration.