| Local affine invariant feature can not only keep the invariant in the image transformation of translation, rotation, scale, illumination and visual angle, but also is more stable than global affine invariant feature in the influence of noise and partial envelop. So it is widely applied to many image processing fields, including image registration, image mosaic, image fusion, object recognition and tracking, image retrieval, 3D reconstruction and so on. As the main foundation of those applications, the precision and efficiency of local affine invariant feature extraction directly affects the research direction and implementation results of subsequent algorithms.Several current main affine invariant feature extraction methods are analysed in detail, and the multiple view imaging model and affine geometric transformation theory is studied in-depth. Through the advantages and disadvantages of affine invariant extraction methods, the novel method of affine invariant feature extraction is proposed and the new descriptor adapted for that method is also proposed. The new extraction and the new method are test and compared detailed. Main research contents and results of this paper are divided into several aspects as follow:(1) The concept of affine invariant feature area density is introduced to evaluate the extraction method, so extraction method can dynamically estimate how compatible it is with the image content. And the relative affine invariant feature area density is introduced to estimate repetition of the affine invariant feature extraction method, which can makes the expansion features or riddling features before registration. That can provide reliable guarantee for the evaluation of extraction method and the improving of efficiency and accuracy of extraction method.(2) For a large number of striped maximally stable extremal regions contained in texture image which are too difficult to match, the method of maximally stable extremal region(MSER) shape detection before matching is proposed. That methods can improve registration rate of affine invariant feature area by removing the striped MSER; For lots of tiny repetitive structure contained in the texture image makes the MSER has the high similarity problem, the B+MSER algorithm is proposed based on between maximally stable extremal regions by using the nested MSER whose difference sets have better gray texture. Experiments show that:The affine invariant feature regions extracted by the new algorithm have high distinguishability. The stability of the new feature repetition under strong noise conditions is higher than that of MSER algorithm.(3)In order to evaluate the stability of MSER under noise and blur conditions, the concept of relative stability of MSER is proposed. Since noise and blur has lower effect on nested MSER and MSER with the high stability, the C+MSER algorithm is proposed based on the combination of the most stable extremal region, with the help of high relative stability MSER located in the same parent node in the extremal tree. Experiments show that:The affine invariant feature regions extracted by new algorithm have strong resistance to noise and image blur, and the repetition and registration radio of C+MSER are better than the MSER algorithm.(4) For improve the efficiency and accuracy of descriptor direction, the two directional calculation methods of descriptors based on brightness centroid invariant moment are proposed, using the method of axial intensive interpolation and area integral respectively. Adopting the coefficient matrix, the complex calculation is designed into coefficient matrix before matching, which greatly optimized the calculation speed. The experimental results show that:the calculation speed of the two methods are 4 times faster than SURF; the accuracy of direction calculation method using area integral is more balanced than SURF algorithm.The high accuracy in certain directions of calculation method using axial intensive interpolation suggested that the coefficient matrix still can be improved. |
PDF Full Text Request