Image Matching Algorithm Based On Feature Extraction And Description

Image matching is an important technique in image processing and pattern recognition. It alignes two or more images,which may be taken under the different imaging equipment or different imaging environment, by finding their transformation model and establishing the mapping relationship. And it is widely used in many application areas, such as visual navigation, remote sensing, object tracking, image fusion, medical diagnostic,3D reconstruction and so on. Along with the development of these applications, the requirement for the efficiency and the accuracy of image matching increases. The basic idea of image matching based on feature is to extract the key points, then describe them making them stable to gray, rotation change, perspective and noise. At last, matching the feature descriptors between the images by the specific similarity measurement criteria.In this paper, we studied the previous achievements in image matching field. The research of this paper mainly focuses on the extraction and description construction of key points. A novel image matching algorithm is presented on this basis. The research involves three aspects:(1) The main related techniques of image matching based on feature points are introduced systemically. Including image preprocessing,the establishment of the scale space,purification of matching points,image matching performance evaluation standard.(2) Several classic feature points extraction algorithm are introduced detailedly, including:Moravec corner, Harris corner, SUSAN corner, Fast corner, SIFT and SURF, and the detection effect of the algorithm are given through experiments.In addition, two kinds of commonly used feature points describing methods are also introduced,including local descriptor which is based on gradient histogram or which is based on binary string. And then introduces several representative descriptor.(3) Puts forward a real-time robust feature point matching algorithm, RRM. Firstly, this algorithm determines the image edge area and edge direction by differential operation, then finds out the anchor which is likely to be the feature points in the edge area, and then uses the main curvature to weed out the unstable points, improving the stability of the matching algorithm. Use the improved BRIEF to describe the feature points.Finally, combined Hamming distance with bidirectional matching to match feature points, improving the accuracy of matching. The experiment proved that this algorithm has a good performance in scale change, rotation change, perspective change, light and noise etc. Besides, this algorithm is efficient enough in computing, and can be used for real-time applications.