Research On Automatic Image Annotation

The popularity of high resolution digital camera and the development of broadband communication networks have resulted in a huge amount of image data being generated and made accessible in digital form. It has been an important topic in the field of information retrieval to effectively manage and retrieve such huge collections of images. Due to the problem that "Semantic Gap" brings, traditional content-based image retrieval techniques could not provide sufficient semantic-based retrieval, which is more user-friendly by building indexes for images.The key point for the semantic-based image retrieval is to automatically annotate the images. Automatic image annotation is an important but highly challenging task in image retrieval.Providing automatic image annotation requires an accurate mapping of images' low-level perceptual features (e.g.,color and texture) to some high-level semantic labels (e.g.,landscape, architecture).From the training data, a discriminative approach (such as SVMs) can be adopted to learn a classifier, which is used to predict semantics for unlabeled images. In practice, automatic image annotation involves multiple classes.A multi-class classification problem is commonly solved by decomposition to a series of binary problems such that the standard SVM can be directly applied. The time complexity of this approach increases rapidly due to the growing number of training data and classes.Aiming at this problem, this dissertation proposes two methods to tackle it.First of all, we focus on the multi-class classification and present a Simplified Multi-class Support Vector Machine(SimMSVM).SimMSVM gives a direct solution for simultaneously training multi-class predictors;it also introduces a relaxed multi-class loss function so as to reduce the overall size of the resulting optimization problem. We then apply SimMSVM in the automatic image annotation framework, and show its capability to speed-up the training and testing process.Secondly, we propose using Support Vector Data Description(SVDD) for improving the efficiency and incremental learning on automatic image annotation. For each class, a minimum bounding hyper-sphere containing one-class data is obtained. An unlabeled image is classified and annotated accordingly with these boundaries. Keywords:Automatic Image Annotation,Content-based Image Retrieval,Support Vector Machines,Support Vector Data Description,MPEG-7...