Region-Based Supervised Semantic Image Annotation

With the growth of images and image data at an unprecedented rate, image retrieval, aiming to retrieve data effectively from mass data of images that satisfy the user request, has become an important research topic in the field of information retrieval. In traditional content based image retrieval, the queries are provided in the form of example images or low-level features and the retrieval performance is impaired by the problem of"semantic gap"between the low-level visual features and the high-level semantic concepts. Therefore, how to precisely represent the image semantic content has become a hot research area, and the key of semantic-based image retrieval is semantic-based image annotation. In this dissertation the development of imaging annotation is presented, and a few new region-based supervised semantic image annotation methods are proposed. The main contributions of the dissertation are as follows:Firstly, we propose an improved algorithm for supervised semantic image annotation. This algorithm includes the following processes: the images segmentation with JSEG algorithm, the color and texture features extraction, the Gaussian Mixture Model representation of the images using the EM algorithm, the updating of the classifier by excluding of the noisy Gaussian components and the concepts annotation. Color and texture features form two separate vectors, for which two independent Gaussian mixture models are estimated from the training set as class densities combined with a denoising technique. Two posterior probabilities are calculated, and both their ranks among different concepts are used to determine the labels for the image to be annotated. Better annotation performance is achieved as compared with methods that treat color and texture as one feature vector on the Trecvid2005 dataset.Secondly, an algorithm for supervised semantic image annotation using region relevance is proposed. The image-level posterior probabilities are obtained by combining the regional posterior probabilities which are modified using relevance with the other regions in the same image, since region features contribute to image concepts and concepts of neighbor regions are correlated. The proposed algorithm achieves good annotation performance on the Trecvid2005 and Corel5K benchmark dataset. Using a segmentation method which splits the image into small blocks of regular size instead of JSEG algorithm, the performance of image annotation has been further improved.