Abstract: The paper proposes a Bayes-based relevant feedback approach by integrating visual features and semantics to effectively make use of semantics and visual features in content-based image retrieval systems. First,the data of the image database are divided into small clusters by semantic supervised clustering algorithm,so the data of each cluster are similar both in visual features and in semantics.Then,on the relevant feedback,users mark the positive and negative samples,in representation of clusters instead of images.At last,we use Bayes classifiers based on visual features and based on semantics respectively to adjust retrieval similarity distance.Experimental results on an image database and a video database show that a few cycles of the relevant feedback by the proposed approach can improve the retrieval precision significantly.