Semantic-based Image Multiclass Annotation

With the popularization of the internet, multimedia data, especially the image data are growing at an unprecedented rate. It has been an important topic in the field of the information retrieval to effectively manage and retrieve images. Owing to the unresolved problem of the"semantic gap"between the low-level visual features and the high-level semantic concepts for the content based image retrieval, the semantic-based image retrieval is gradually becoming a hot research area. The key point of the semantic-based image retrieval is the semantic-based image annotation. This thesis firstly reviews the developing processes (including the text based, the content based and the semantic based) of the imaging annotation, the corresponding theories and their advantages and disadvantages. The current status of semantic based image annotation is summarized. The main contributions of this paper are as follows:Firstly, an algorithm based on the Hierarchical Gaussian Mixture Model with Denoising is proposed for image multiclass annotation. This algorithm includes the following processes: the low features extraction, the Gaussian Mixture Model representation of the images, the training of the concepts classifier based on the Hierarchical Gaussian Mixture Model, the updating of the classifier by the excluding of the noisy Gaussian components and the multiple concepts annotation. By excluding the noisy components(corresponding to the irrelative regions of the concepts) and re-training the classifier with the left components, the annotation performance is significantly improved.Secondly, an improved Probabilistic Latent Semantic Analysis method is proposed for image annotation. This method includes the following processes: the extraction of the low level features, the construction of the vocabulary, the image semantics extraction with a probabilistic latent semantic analysis method and the image annotation with a k-nearest neighbour method. In the process of vocabulary construction, by using the Hierarchical Gaussian Mixture Model instead of the K-means clustering method, the performance of image annotation has been improved.