The Research And Application Of Image Matching Algorithm Based On Feature Extraction And Feature Point Descriptor

Image matching is aligning two or more images by finding their transformationmodel and establishing the mapping relationship, which may be taken under thedifferent imaging environment or different imaging equipment. More and morecomputer vision and artificial intelligence applications rely on matching keypointsacross images, such as military navigation, remote sensing, object tracking, OCR,Medical diagnostic,3D reconstruction, and so on. The basic way of image matchingbased feature is to extract the feature, then describe them, which can contain significantinformation of the images and the process should be stable to the perspective, gray,rotation change, noise and other factors. Finally, using the specific similaritymeasurement criteria to match the feature descriptors between the images.Established leaders in the field of local feature are the SIFT and SURF algorithmswhich exhibit great performance under a variety of image transformations, and manyimproved algorithms based on SIFT are proposed. However, these algorithm impose alarge computational burden, especially for real-time systems such as visual odometry, orfor low-power devices such as cellphones, which encourage the development ofdescriptor with low computational cost and small storage, and several descriptors basedon binary are proposed in the last two years, such as BRIEF, ORB, BRISK.In this paper, we deeply studied on the previous achievements in image matchingfield. The research of this paper mainly focuses on feature points extraction and featuredescription construction. Meanwhile, a novel image matching algorithm is presented.Additionally, we provide an application of real-time electronic image stabilization basedon image matching technology. The main contribution of this paper are:1. Since the FAST corners always distribute densely, and some of them have strongedge response, a new feature point extraction algorithm is proposed. It can eliminate thedistribution clustered of the corners and remove the edge response corners which isbased on FAST corner detect.2. Propose a new descriptor based on binary bits. A new sampling mode is used inthis feature description algorithm. We also use integral image to make the compute more efficient. The experiment in this paper show that our descriptor has a goodperformance on noise, rotation or gray changes, etc. What’s more, we only pay lowcomputational cost and small storage to construct the descriptor, so it could be used inreal-time application.3. We provide an application of real-time electronic image stabilization based onimage matching technology. Firstly, the correct matching point pairs between adjacentframes using our image matching method and RANSAC algorithm are obtained, andthey can be utilized to compute the global motion vector. Secondly, use kalmanalgorithm to filter the motion vectors, and the true movement of the image is preserved.Lastly, bilinear interpolation motion compensation is used to output steady videosequence. We have also done some simulation experiments to prove the effectivenessand real-time property of the electronic image stabilization algorithm.