Although methods employing chain codes or Fourier descriptors are known for calculating an outline similarity between a model image and an object image, these methods are difficult to achieve detection of both an approximate similarity and local similarity. In view of this, according to the present invention, wavelet transformation is performed on outline points of an object image, and similarity calculation is performed on a plurality of model images with the use of a low frequency component of the transformation result. Only the model image, having a close agreement in matching processing, is subjected to similarity calculation using a high frequency component of the transformation result. By virtue of this processing, similarity calculation can be performed at high speed with high precision.