The Research Of Image Matching Method Based On Feature Descriptor

At present, along with the fast growth of the ways and the number of obtained image, digital image processing has attracted more researchers'attention, some basic researches in digital image processing also became more and more important. In many application fields, digital image is applied as a kind of new digital information, so there are more and more images need to process.In practice , image matching is defined as to find the same scene between two or more different images , then make the transformation model for them, usually the images may obtained from different time, different angles or different sensors and so on. Image matching is a key problem in object tracking, artificial intelligence, image navigation, face recognition and computer vision fields, an effective method to this problem is to obtain the transform relationship with feature points which contain sigmination, scale, rotation, noise and perspective change. It can be seen that feature extraction and description is the basic problem in computer vision field, and the matching efficiency and accurnificant structure information for the images, there exist a difficulty that how to obtain the stable feature points and construct the feature descriptors that are robust to illuacy of image matching results are directly decided by the performance of feature test operator and feature description operator. In fact , because the different imaging limitations , the obtained image may be influenced by perspective, affine, illumination ,rotation and scale change, now how to choose reasonable image feature and feature descriptor to make these features not only have good performance, but also under the above change remain unchanged became a crucial problem.This paper analyzed and summarized the existed approaches about feature point extraction and feature descriptor, then proposed a fast rotation invariant image matching algorithm based on texture feature. In this new algorithm, firstly, use classic SIFT operator detect the feature points, and according to set up reasonable parameters to ensure the stability and quantity with feature points. Then use the local features in the surrounding neighborhood to construct a feature vector, in this paper, the local feature is defined as rotation invariant CS-LBP, this idea came from traditional local binary pattern (LBP). When got feature descriptor vector, use similarity measure method to determine the matching relation between feature point pairs, then realize two images matching through this way. At last, the algorithm proposed in this paper and SIFT have been experimented in the case of image scale and rotation changes, and compare the experiment results. The results show that the new algorithm has better performance than SIFT algorithm.