Investigation Of Image Matching Algorithm Based On Feature Points

Image matching is aligning two or more images of the same scene in the space, which may be taken under the conditions such as by different sensors, from different aspects and in different time. It is a key problem to be solved existing in object tracking, image navigation, OCR, resources analysis and face recognition as well as in other fields such as computer vision. An effective solution to this problem is to obtain the transform relationship with feature points which is small enough in quantity and contains significant structure information representing the feature of the images. The difficulties to be conquered in the method is how to extract stable feature points and build feature descriptors that are adaptive to transform, distortion, occlusion, noise and some factors in other forms. On the basis of analyzing and summarizing the approaches to feature point extraction and descriptor generation, this dissertation presents a fast image matching algorithm based on DOG feature points. Firstly, the means to extract DOG feature points is simplified with the quantity and stability promised by selecting appropriate parameters. Following of this, a descriptor which is 25 dimensions and keeps invariant when rotated is proposed. Then the potential matches are determined using the similarities of descriptors, and false matches are removed according to the characteristic that the length ratio of corresponding segments maintains invariant. Therefore, two images can be matched through this way. At last, the algorithm proposed in this paper, SIFT and Imp-MOPs have been experimented in the case of image transform, noise disturbance, image blur, image compression, illumination changes, little changes of view as well as scale zoom, and the performance of them have been compared. The results show that new algorithm almost performs as well as SIFT, and better than Imp-MOPs, but the proceing speed is about as twice as the others.