Affine Invariant Feature of Multi-Scale Maximally Stable Extremal Region

Affine invariant features based on local region are widely used for object recognition, scene classification and image retrieval in the field of computer vision. Among the proposed affine invariant region features, maximally stable extremal region (MSER) has been proven to have attractive properties in several aspects. But since MSER is extracted from one single scale image, MSER will become unstable when the image is blurred due to the change of scale. To solve this problem, an innovational affine invariant feature, which is maximally stable both in image space and scale space, is designed with a criterion to evaluate the stability of extremal regions in scale space. And a fast algorithm is also proposed for extracting extremal regions in the scale space. At the same time, according to the property that the extremal regions can describe the contour of feature fairly well, a new kind of feature descriptor is designed by combining the local gray grads and the shape information. The proposed affine invariant feature with its descriptor is called multi-scale maximally stable extremal region (MMSER) feature. The experimental results prove that MMSER is much more stable and discernable than MSER under different affine transform conditions, and that the computational time of the designed descriptor is about 45% of that of SIFT descriptor.