The Research Of Invariant Image Feature Matching

As the rapid development of modern technology, image information has been applied in various fields, and more and more images are to be processed. Image matching is the process of overlaying the same contents of two or more images taken at different conditions, which contain the effect of noise, climate. So, image matching technology is a difficult problem for current research, and it can deal with works like computer recognition and computer understanding.The images obtained from single sensor don't have enough information. The effective solution is to integrate different information by various of sensors in which image matching about same regions is the crucial step. But image acquisition devices are different. So the time, climate and viewpoint changes always exist. All of them cause the trouble to the matching. The method of image matching based features is emphasized, which extract part salient features from images and thus greatly compress the amount of information. Especially, it is robust to other interference like gray changes.The crucial step of image matching based features is to realize feature matching, which is to realize the corresponding between different features from images. Once the accuracy of feature matching is high, transform model can be estimated well, then image matching is success. In this paper, we apply point matching as feature matching, and propose a new descriptor and two matching algorithms about invariant features.A new descriptor method is robust to rotation and illumination. The algorithm firstly is to extract harris corner graph from the neighborhood region about feature point; secondly compute the LBP vectors about every corners in the graph and change to the decimal; a one-dismensional ordered list is lastly gained, which is the descriptor of the point.Using the thought of fusion, the two new matching algorithms combine original SIFT descriptor and other salient descriptors, then achieve feature matching by a suitable similarity measure. Experimental results demonstrate that the two new methods outperform original SIFT approach and can represent high matching accuracy, especially when a large viewpoint change are occurred.