Research On Image Registration Algorithm Based On SIFT Algorithm

Image registration problem, a classical problem in the field of image registration, has been widely used in the aspects of pattern recognition, computer vision, medical image processing and remote sensing information processing, etc. So far, as for the method of image registration, there has been a considerable amount of research work in the image registration research field both at home and abroad to bring about a lot of image registration algorithm. Among many algorithms, scale invariant feature transform (SIFT) is characterized by rotation invariance, scale invariance, significance and great quantity as well as the stability of the perspective changes and noise. These features, after being put forward, make the SIFT features become a kind of local character of classic image. Now, this method has been widely used in the fields such as target recognition,3D reconstruction, video retrieval, etc.SIFT (Scale Invariant Feature Transform) algorithm is a kind of extraction algorithm based on local invariant feature which was put forward by D.G Lowe in1999and perfected and concluded in2004. The image registration is a critical step of image processing, and along with the gradually wide use, there are higher and higher requirements for the adaptability, matching precision and instantaneity of the image matching. This topic, based on the principle of image registration and in view of the analysis and study of SIFT features detection algorithm, Euclidean distance calculation and other aspects, presents how to improve the SIFT algorithm based on image Euclidean distance and the distance of absolute value.Firstly, it studies the realization process of the SIFT algorithm and analyzes the advantages and disadvantages of the algorithm. When the feature vector of the two images is generated, the Euclidean distance of the key feature vector is adopted as similarity judgment measure of the key points in the two images. As the SIFT feature amount of each image is large and every SIFT feature is a128-dimensional vector, so the computation efficiency matching the SIFT feature is very low; and, the traditional Euclidean distance can only calculate out the variance of the two images to the corresponding pixel, so when a slight shift or distortion occurs in the image, it may produce a large deviation, and the algorithm in this paper avoids the error generated in the measurement of the image similarity and obtains the correct matching point.Secondly, it studies the method to measure the Euclidean distance by the general similarity, but this method does not take the space relation of the pixel into account, and it is sensitive to image deformation. Based on the above issues, Wang LiWei et al proposed an improved algorithm of Euclidean Distance-the IMage Euclidean Distance, IMED for short. It takes the space relation of the pixel into account, which has good robustness for small deformation. IMED thinks there is gray correlation between any two pixels in the image, the related degree is inversely proportional to the distance of the pixel, larger the distance is, lower the correlation degree will be, and the distance between the two images is composed by all the pixel distances. On the basis of this thought, this paper puts forward the improved method of similarity measurement.Finally, the replacement of the traditional Euclidean distance in the original algorithm by the absolute distance simplifies the operation process of SIFT algorithm and improves the efficiency of the algorithm. Then the image processed by IMED algorithm is taken as the input image and inset to the improved SIFT algorithm. As the IMED algorithm fully considers the correlation of the pixel space position, after it is inset to the SIFT algorithm, it fully considers the correlation of the pixel of each SIFT feature descriptor to effectively improve the accuracy of the algorithm.Experiments show that, under conditions such as noise, rotation, light and shade and so on, the computation efficiency to improve algorithm is higher than that of the original algorithm, and compared with the original SIFT algorithm, the matching accuracy is improved to some extent, and the experimental results are consistent with theoretical results.