On An Image Local Affine Invariant Feature

Among most of images, there commonly exist some deformations, i.e., rotation, scale variation, partial occlusion, and affine transformations. Therefore, extraction and matching of local features that are invariant to such above deformations are important steps in many computer vision fields, such as three-dimensional image reconstruction, digital watermarking, robot vision, image registration, and so on.First, based on the scale invariant feature transform (SIFT), the fully affine invariant feature extraction and matching framework——Affine-SIFT is thoroughly studied in this thesis. Experimental results show that ASIFT significantly outperforms the classic SIFT in various kinds of image deformation situations, such as JPEG compression, image blur, illumination changes, and especially for large view angle changes.Second, an improved ASIFT algorithm is proposed for feature extraction and matching in multimodality images. By combining the fully affine invariant framework and modifying the gradient orientations in the local neighborhoods, the main orientation, and the descriptor of each ASIFT feature, the improved ASIFT works well for multimodality image feature extraction and matching. Experiential results show that the improved ASIFT performs better than the SIFT and the ASIFT methods, especially in feature matching accuracy criteria.Finally, an image processing software package, Multi-Feature Genius (MFG), is developed with C/C++languages under Visual Studio2008MFC programming environment. Some local feature extraction and matching algorithms, including SIFT, Harris-SIFT, ASIFT and the improved ASIFT, are integrated into MFG. Moreover, some related image processing methods are also realized, such as image registration, geometric transformation functions, image fusion, etc. The MFG software package is with user friendly interface as well as practical application value.