Since the number of Japanese and Chinese characters is large, two-stage processing for recognition of the characters is useful to reduce the recognition time to have reasonable responses. The two-stage processing is a combination of rough and fine classification of characters. The rough classification is a selection of limited number of candidates by grasping whole image and is implemented as an image-type processing. The fine classification is a final decision among the candidates by logical-type processing. Based on this idea, we have constructed a fast and intelligent character recognition system, SEIUN. It carries out rough classification of 2,965 kinds of printed characters at a speed of about 200 characters per second. The rate the correct candidate is in the first three candidates is greater than 99.99% for characters with clear images. In this paper, parallel architecture of an image processor is presented, that is a significant subsystem of SEIUN. The architecture is based on systolic algorithms and it realizes useful functions for some image processings. Some functions of the processor is also applicable for fine classification and they are described.