Combining Visual Features and Contextual Information for Image Retrieval and Annotation.

摘要

This thesis is primarily focused on the information combination at different levels of a statistical pattern classification framework for image annotation and retrieval. Based on the previous study within the fields of image annotation and retrieval, it has been well-recognized that the low-level visual features, such as color and texture, and high-level features, such as textual description and context, are distinct yet complementary in terms of their distributions and the corresponding discriminative powers for dealing with machine-based recognition and retrieval tasks. Therefore, effective feature combination for image annotation and retrieval has become a desirable and promising perspective from which the semantic gap can be further bridged. Motivated by this fact, the combination of the visual and context modalities and that of different features in the visual domain are tackled by developing two statistical pattern classification approaches considering that the features of the visual modality and those across different modalities exhibit different degrees of heterogeneities, and thus, should be treated differently. Regarding the cross-modality feature combination, a Bayesian framework is proposed to integrate visual content and context, which has been applied to various image annotation and retrieval frameworks. In terms of the combination of different low-level features in the visual domain, the problem is tackled with a novel method that combines texture and color features via a mixture model of their joint distribution. To evaluate the proposed frameworks, many different datasets are employed in the experiments, including the COREL database for image retrieval and the MSRC, LabelMe, PASCAL VOC2009, and an animal image database collected by ourselves for image annotation. Using various evaluation criteria, the first framework is shown to be more effective than the methods purely based on the low-level features or high-level context. As for the second, the experimental results demonstrate not only its superior performance to other feature combination methods but also its ability to discover visual clusters using texture and color simultaneously. Moreover, a demo search engine based on the Bayesian framework is implemented and available online.