A concept of visual keys is proposed to provide efficient and useful content-based image retrieval systems to users. Visual keys are defined as representative sub-images which are extracted from an image database by using image feature clustering. The proposed system is implemented and is tested on 1,000 images, which are included in the COREL database. Although the system makes use of only 80 sub-images from 8,962 ones extracted from the image database, the performance is kept with 90%. The retrieval time is within 4ms on the proposed system, which has retrieval efficiency like that of text retrieval by being applied text retrieval techniques, and thus the system is expected to provide the services on the WWW.