Automatic image annotation has been becoming an attractive research subject. Most current image annotation methods are based on training techniques. The major weaknesses of such solutions include limited annotation vocabulary and labor-intensive involvement. However, Web images possess a lot of texts, and rich annotation of samples is provided. Therefore, this report provides a novel image annotation method by mining the Web that term-image correlation is obtained from the Web not by learning. Without question, there are many noises in that relation, and some cleaning works are necessary. In the system, entropy weighting and image clustering technique are employed. Our experiment results show that our solution can achieve a satisfactory performance.