In recent years, tagging methods have been widely used on the Web for multimedia search and recommendation systems. Social tags are the keywords annotated by users to the multimedia objects. These tags contain the information which allows multimedia objects to be located and classified. Unfortunately, irrelevant information and noise are frequently included in such tags, especially when we use social tagging systems. Imperfect as they may be, social tags are still a significant source of human-generated contextual knowledge. We present the Waking And Sleeping (WAS) algorithm to overcome the obstacle of noisy tags. WAS refines social tags of each image and assigns each tag into either the waking state or the sleeping state. With WAS, the recommendation system not only enhances the accuracy of the recommendation results based on tags with higher quality but also reduces the consumption of computational resources. The experimental results demonstrate that the proposed approach is superior to compared methods.