Affine Invariant Feature Analysis Of Repetitive Patterns And Applications

With the rapid development of computer technology, simulating human eye vision system by computer vision system is more and more applicable, which can greatly reduce the workload of human beings. Image feature extraction and matching, as an important research topic in the field of image processing and pattern recognition, has obtained more and more attention in image registration, target classification and recognition, image mosaics, three-dimensional reconstruction, autonomous robot navigation and other fields.The affine invariant feature can reflect the important primitive properties of objects with better stability and adaptability. Moreover, the affine invariant feature is independent of the surroundings, the position and the attitude of the object and is invariant to the transformations of image translation, rotation, scale, lighting and perspective transformation to some extent. From the developing trend of the application, it can be seen that the affine invariant feature of images is one of the hot issues in the fields of computer vision.Based on the demand for the actual application, a deep exploration and research is made in the extracting and matching of local affine invariant features in this paper. The main research results are shown in three aspects as follows:(1)For the images containing the repetitive shapes, a novel affine invariant feature matching algorithm is presented in this paper based on the SIFT and ASIFT algorithm. For repetitive shapes, the bi-directional matching is used and the new matching threshold of descriptors is confirmed, which make the feature points that are independent of repetitive shapes are firstly matched with high stability and correctness. What is more, the correct matching pairs are beneficial to the estimation of homographic matrix between different views.(2)Secondly, an affine invariant geometric property (the area ratio of corresponding triangles is invariant) is used to further increase the feature pairs in repetitive shapes based on the first step. By considering the affine position relationship between feature points, this algorithm is not only invariant to affine transformation, but also can deal with the difficulty caused by the similarity measurement of feature descriptor when matching the repetitive shapes. Experiments demonstrate that the proposed method has some robustness and availability to the repetitive shapes and outperforms the ASIFT algorithm in both the matching accuracy and the number of correct feature pairs.(3)The locations of SIFT features are decided by the local extreme and their scales in the DOG Gaussian pyramid, which may easily lead to the position deviation of features and the missing of some real features meanwhile. In combination with Harris corner detection operator and the theory of scale space, in this paper, a new affine invariant feature based on the Harris corner is put forward which can adaptively determine the scale of corners by iteratively selecting scale from small to large. So the features not only have some invariance to affine transformation but also are located precisely.Summarizing the research achievements of the above three aspects, the proposed Harris affine invariant features and the new matching strategy can not only be applied to the general images, but also can substantially increase the stability of matching in repetitive shapes.