Abstract: Rotation invariant kernel was applied to face recognition and some other areas, but its antinoise ability was unsatisfied. Elide the specific form of distribution, the generalized rotation invariant kernel, which can convert the original nonlinear problem into linear one, was introduced. Meanwhile, parameter estimated difficulty was reduced. The index α of generalized Gaussian function played a decisive role on the peak. With reference to this property, index r was introduced into the algorithm. Through controlling the change of index r, recognition rate of the algorithm was observed. And the antinoise ability of the algorithm could be proved by adding different Gaussian white noise into the experimental data.Experimental results proved the superiority. Recognition rate was almost linearly changed while the index r changed and the best r always existed for the best recognition rate, which was better than that of rotation invariant kernels.Under the same experimental conditions,the antinoise ability was greatly improved.

Key words: kernel functions, rotation invariance, sphericalhomoscedastic, face recognition