Abstract:

A new method for multi-scale affine geometric invariant extraction is proposed. The method begins with the self defined  multi-scale convolution transformation, combines with gray scale normalization, and builds a series of affine covariant  forms of the object image. After that, a series of extended centroids of each covariant form are calculated through a set  of designed nonlinear functions, and the new multiscale affine geometric invariants are obtained. Compared with the  classical extended centroid features and multiscale autoconvolution, the introduced invariant only needs cutting once  and can construct any number area ratio invariant features. More invariant features can be extracted from a single affine  covariant form. All of these can reduce feature errors effectively and improve the efficiency of the feature attainment.  A typical “fish” test database is adopted to validate the efficiency of the proposed method from the perspective of  computational complexity, noise immunity, anti-blocking and image expansion.