A Gaussian Blur Invariant SIFT Descriptor Based On Multi-Sampling And Optimization Searching

With the massive popularity of computer, a range of disciplines such as computer science, information science, electronics, material science, statistics and so on had been the rapid development, and had also laid a solid foundation for the digital image processing. As a hot field of digital image processing, image matching often involves the blurred image matching which include the motion blur, defocus blur and gaussian blur. These blur deformation seriously affected the result of image matching and target recognition., Therefore, efficient image matching algorithm has always been a hot research topic in the field of image processing. The Scale Invariant Feature Transform(SIFT), as a kind of classic local feature matching algorithm, has widely used in all kinds of image matching. But SIFT is not everything, many scholars has proposed some improved algorithms in different aspects in order to improve the performance of SIFT. ASIFT(Affine-SIFT), Which is representative of these algorithms has been almost provided the completely affine invariant, the calculation efficiency of SURF(Speed Up Robust Features) is very close to the real time calculation. But after consulting a large number of literatures, the author found that the current field of image matching did not exist any specifically algorithms for image matching in condition of gaussian blur. So improve the matching ability of SIFT algorithm in condition of gaussian blur is the research goal of this paper, it is also very difficult in the process of blur image matching. Therefore, the main work of this paper is discussed from the following aspects:(1) Firstly, this paper has proposed a new separate two-dimensional gaussian fuzzy algorithm according to the characteristics of the gaussian blur mathematics model, and then it has put forward a matching strategy which based on deformation space resampling. The first step of this strategy was sampled the gaussian blur parameters to construct the blur deformation space to approach the real blur space. And then the identified image would match with the all images in this space to find the best matching.(2) Secondly, this paper has proposed a optimization algorithm which based on hill-climbing method and down-sampling to traverse the deformation space. This algorithm could reduce to the calculation of one image to one-ninth calculation according to the regular of the matching curve between the identified images and deformation space in different resolution. And to take a further advantage of the characteristics of matching curve, hill-climbing method and 2D curve fitting was introduced to this algorithm. Theoretically, 2D curve fitting just need three points which around the peak point to directly calculate the real optimal matching.(3) A kind of target identification method which based on the local stable features of images has been introduced. This paper has achieved this algorithm and combined with GI-SIFT to matching and identification the target in gaussian blurred images;(4) Experiment and analysis was run on the platform which based on Open CV and GTK+. Four experiments, standard image matching effect, matching efficiency comparison, parameter identification, practical application effects, had fully proved the ability of gaussian blur image matching, parameter identification, calculation efficiency and target identification of GI-SIFT.