Research On Automatic Image Annotation

The plethora of digital camera and other digital products with imaging portion enable image generation in a cheaper and easier manner. With the popularity of computer network, Twitter, and some image sharing websites, such as Picasa and Flickr, these technologies accelerate the spread of images. The explosive growth of image data has induced a paradigm shift in the research on the management and understanding of image data. And automatic image annotation is the core technique for image management and understanding. My works are as follows:(1) The improved CMRM algorithm is proposed, and the general formula of correlation model is presented. After that, the interpretations of general formula are given in the view of probability theory, information retrieval and multi-modal respectively. And based on that comparison and analysis, we point out core techniques of relevance model and research concentration of this area in the future. The experimental result shows that improved CMRM outperformes CMRM in terms of accuracy and efficiency.(2) A new annotation method based on positive and negative vectors is proposed. To improve accuracy, traditional classification-based annotation methods use more abstract model and complex algorithms, but this leads to a worse performance in terms of efficiency. Our proposed method is based on trans-media. By constructing an individual visual feature vector for every textual keyword or semantic concept, we can convert image annotation problem to the comparison of similarities for visual feature vectors. The visual feature vector of an image is determined by difference of the mean of positive examples and negative examples. Compared with traditional methods, our proposed method has①a simpler model,②lower time cost for training and testing process,③better annotation performance in terms of mean per-word. Our proposed (positive and negative vector based) method can be used as an independent tool for image annotation. Much more than this, it can be a reference method to determine whether a more complex model or feature descriptors necessary to be used.