Research On Feature-based Multimodal Automatic Image Registration

Image registration is a foundation problem in image processing. It is widely used in computer vision, pattern recognition, medical image analysis and remote sensing image analysis. The major purpose of registration is to establish geometric transformation between two images, and remove or suppress the geometric distortions between them. Multimodal image registration is one of the most challenges in image registration. When images have significant differences in translation, rotation or scale, most of multimodal registration methods either fail or became extremely time consuming, which cannot be applied to a task requiring fast automatic image registration. Feature-based methods perform well for the images having large misalignment, and their computational costs are low. In order to realize fast automatic registration among multimodal images, an image registration method using both corners and area-based matching is presented in thesis. The image obtained always has been blurred by noise. After comparing corner detection methods with their detection accuracy and computational efficiency, a high local-contrast corner, also call Significant Point, detection algorithm presented by Zitováin 1999 is found, which has much better performance to the blurred and noisy image data. Thus, it is suitable for multimodal image registration. Unfortunately, Zitová's method needs to search candidate corners in whole image, leading to the decrease of detection efficiency. To increase the efficiency of detection, an improved detection algorithm of corners with high local contrast is presented in the thesis. It has the same detection accuracy as the Zitová's algorithm, but with higher computational efficiency. Experiments demonstrated the feasibility of our algorithm. Based on the improved corner detection algorithm, an image automatic registration method combined corner detection and area-based matching was developed. In the corner matching procedure, the correspondence between the corner points in two images was established with three steps. First, a primary correspondence was initialized using normalized cross-correlation. Secondly, the wrong match points were deleted based on the Mahalanobis distance with affine invariance. Thirdly, the correct correspondence was validated further by deleting the pair points with large distance gap after the affine transformation. Finally, the affine transformation parameters were estimated from the correctly corresponding points. Experiments demonstrated the effectiveness of the method in both accuracy and computational speed. Our registration method works well for the multimodal images with significant translation and rotation. It is suitable for automatic registration between the multimodal images with less noisy and scale variance.