In this paper, we propose a shift and rotation invariant pattern recognition architecture, using morphology pattern spectrum as image feature representation and 2-D vector quantization network as vector classifier. Morphology pattern spectrum is invariant not to image shift, but to image rotation when the structure element is rotationally symmetric. We use it constructing a feature vector of an image to train a 2-D vector quantization network. After training, it can implement the classification of feature vector. The optical implementation of the 2-D vector quantization network is dementrated and the experiment results are given.