The FourierMellin descriptor has been widely applied in digital image matching. However, because of the noise of spectrum aliasing and errors, the Fourier Mellin descriptor can not be sure that the form of the changed relative value distribution is the standard twodimensional pulse function. Based on the basis of introducing the distance formula with a kernel function, first using wavelet transform in the frequency domain extract stable feature points to form a  feature vector；then the kernel function is used to amend the feature vector, thus making the formation of a standard twodimensional pulse function, to correctly compute the matching parameters, so as to achieve the matching result of two images. The experimental results show that the new algorithm is efficient and reliable.