Machine learning techniques to build geometrical transformations for object matching a review
Image matching technology is the research foundation of many computer vision problems, and the matching algorithm based on partial features of images is a research focus in this field. In order to overcome the unstable performance of classic SURF algorithm on rotation invariance, an image matching algorithm combined with SURF feature-point and DAISY descriptor is proposed. Based on the feature point detection of SURF algorithm, a principal direction distribution method for DAISY descriptor is put forward, and a novel DAISY descriptor is obtained according to the rotation of the principal direction. In this paper, our proposed algorithm, on the basis of slight increase in running time, improves the image matching capability of the classic SURF algorithm on image rotation. The experimental results show that our proposed algorithm has stronger robustness in a variety of complex cases, such as image blurring, illumination variation, JPEG compression ratio variation, field of view variation, etc. Our proposed algorithm can not only keep the merits of the original SURF algorithm on computation speed, but also improve the matching accuracy on rotation invariance.