Research On Multi-source Image Registration Techniques

In the field of computer vision, image registration is a fundamental and important one, which is the base of image stitching, image fusion, image detection, recognition and tracking, and is widely used in the fields such as medical image processing, remote sensing and military task etc. As the broad application of multi-wavelength imaging, the need for multi source image registration and fusion is growing. This paper mainly focus on the research upon image registration algorithms based on SIFT, and discuss the difficulties it faces when applied to the registration of multi-wavelength images and propose possible solution to get better registration performance.This paper first elaborates the framework of image registration and its components, which is feature detection, feature description, image transformation model and methods for parameters estimation. For each component in the registration we do some analysis and research on the popular algorithms, and then we make some detail research on image registration based on SIFT algorithm. In order to better understand the key issues in the whole image registration process, we first apply it on registering visible images. During the experiments on registering visible images, we find SIFT experiencing some defects like too much computational burden and low discriminative of similar features in different regions of the image. To tackle these defects, we propose to preprocess images for extracting ROI (Region of Interest) before registering, and adopt simplified Gaussian pyramid to detect feature for lowering down the complexity of computational burden. And then confine the search for corresponding matches to the roughly estimated by strong match points, which can eliminate most of the mismatches caused by similar features in different regions to some extent. The experiment results shows that the improved algorithm can get better performance in lower down the computational burden and eliminating wrong matches than the original one.We then expand SIFT to register multi wavelength image registration, and propose our improved methods which is based on its defects occurred in the experiments. First, in feature detection we take SURF as our reference, and make the use of integral image to speed up SIFT feature detection. While we adopt simplified one octave Gaussian image pyramid to detect features, we enlarged the factor k between each layer to get a more sound coverage of scale distribution, thus we maintain the property of scale invariance and simplify the computation of feature detection at the same time. In the process of getting feature descriptors, since SIFT descriptor is of128dimensions which makes it cost a lot of resources to compute and search for matches, and it appears that SIFT descriptors cannot discriminate different features in different wavelength images. We introduce BRIEF descriptors to multi wavelength image registration. The experiment result shows that BRIEF descriptor can not only enjoys a good performance in the efficiency of speed and the space needed for storage, but it also obtains a good results of describing multi wavelength image features. For methods of searching corresponding matching points, we combine the statistical data with the actual matching images to set a search window and employ the bilateral search for match points. These strategies together provide an enhancement in the matching precision and lower down the computational burden. Finally, for estimation of transformation parameters, we do some research on the popular algorithm for estimation of best inliers set RANSAC, and we find its potential shortcomings. And we replace the generation of random initial data set from only one data set to multiple data sets, and then compute the mean square distance within points in the data sets, choose the maximum distance data set as the initial data to iterate for the best inliers sets. According to the parameters we estimated, apply bilinear interpolating or cubic interpolation to resample and interpolate the images to accomplish the registration process.This paper made detail discussion on the key steps in image registration, and analysis the advantages and disadvantages of the popular algorithm in the state of art, for its shortcomings, we proposed available solutions to improve the efficiency and the robustness of the image registration process.