Researches On Registration Of Multimodal Image Based On Structure Information

In the fields of Military terminal guidance, remote sensing image fusion, medical imaging diagnosis etc., multi-sensor technology embodies the important application value. With the rapid development of sensor imaging technology, only one sensor has been unable to meet the needs of practical application. As the foundation of multi-sensor application, the multimodality registration technique has important research significance. Multimodality registration is to match and align images captured by different sensors at different times, different viewpoints or different imaging conditions. Due to significant difference of imaging mechanisms between multimodal images, their registration methods are difficult to achieve the ideal effect in terms of time consumption and matching precision. Therefore, this paper, based on stable structure information, carried out the studies of multimodality registration.First of all, through the experiment we found the keypoints useful for matching results mostly distribute on or near the edge. Aiming at the problem of a large number of redundant keypoints affecting matching performance, this dissertation proposes an accelerating method of multimodality registration based on the structure of edge. According to the distance to the edge this method categorizes keypoints. The keypoints which are far from the edge are deleted. The remaining ones are then matched separately. Finally, the threshold obtained by the Bayesian formula is used to filter the initial matching results. The experimental results show that the method can effectively improve the performance of Symmetric-SIFT and SIFT, especially in the aspect of reducing time consumption.Secondly, by the experiments we verified that in the locations which contain a large amount of structural information there are a lot of gradient reversals. Base on this phenomenon, MM-SURF algorithm is proposed. The algorithm uses SURF detector to obtain keypoints, utilizes neighborhood gradient magnitude to compute the dominant orientation of each keypoint and finally creates MM-SURF descriptors for the keypoints. Experimental results show this algorithm can reach good effect both in the terms of time and precision. In addition, the algorithm also maintains robustness and stability in the presence of image blurring, rotation, noise and luminance variations.Finally, this paper proposes a Structure-LBP operator which can describe the structure information of multimodal images, and based on the operator puts forward a fast and robust multimodality registration method. The method uses SURF algorithm framework to detect keypoints, removes low significant keypoints, then utilizes the structure degree of the neighborhood to determine the dominant orientations of the remaining keypoints, next calculate the Structure-LBP descriptor of each keypoints and finally deletes the keypoints with high self-similar factor. Experimental results show the method has good performance in the time and precision, especially that in the presence of image blurring and noise the proposed algorithm can show very good stability.