Abstract: An improved image region classification method is presented in this paper, which can works with inaccurate segmentation of image regions.First, each coarsely segmented region was divided into several sub-blocks by using bag-of-words method, and the semantic features of the region were obtained by clustering the low-level image features extracted from these sub-blocks.According to the confidence of the semantic features, these semantic features were further refined by using multi-instance learning method.Finally, these refined semantic features were learned by a SVM classifier for region classification.The experimental results show that the proposed method achieves good region classification with the coarse region segmentation of the image.