Combining regional and global features for automatic image annotation based on VQ

In this paper, a novel method of automatic image annotation based on the Vector Quantization (VQ) compression domain is presented. The Co-occurrence statistical model was the inspiration behind developing this method, in which the combination of the global and regional features is used for the annotation process. The labeled images are compressed using the VQ compression method. Subsequently, the regional features are extracted from the images and are weighted. The Seed K-means (SK-means) semi-supervised clustering method is employed to increase the accuracy of clustering the weights obtained. Since the global and regional features emphasize different aspects of images and complement each other, the combinational approach of the global and regional features is adopted for the testing stage, and the unlabeled images are annotated. The results of the test on 5000 images from the Corel collection revealed that the proposed method is more efficient than the other methods in the uncompressed domain.