Region-based Multimedia Indexing and Retrieval Framework
Permanent address of the item is
Many systems have been proposed for automatic description and indexing of digital data, for posterior retrieval. One of such content-based indexing-and-retrieval systems, and the one used as a framework in this thesis, is the MUVIS system, which was developed at Tampere University of Technology, in Finland. Moreover, Content-based Image Retrieval (CBIR) utilising frame-based and region-based features has been a dynamic research area in the past years. Several systems have been developed using their specific segmentation, feature extraction, and retrieval methods. In this thesis, a framework to model a regionalised CBIR framework is presented. The framework does not specify or fix the segmentation and local feature extraction methods, which are instead considered as “black-boxes” so as to allow the application of any segmentation method and visual descriptor. The proposed framework adopts a grouping approach in order to correct possible over- segmentation faults and a spatial feature called region proximity is introduced to describe regions topology in a visual scene by a block-based approach. Using the MUVIS system, a prototype system of the proposed framework is implemented as a region-based feature extraction module, which integrates simple colour segmentation and region-based feature description based on colour and texture. The spatial region proximity feature represents regions and describes their topology by a novel metric proposed in this thesis based on the block-based approach and average distance calculation. After the region-based feature extraction step, a feature vector is formed which holds information about all image regions with their local low-level and spatial properties. During the retrieval process, those feature vectors are used for computing the (dis-)similarity distances between two images, taking into account each of their individual components. In this case a many-to-one matching scheme between regions characterised by a similarity maximisation approach is integrated into a query-by-example scheme. Retrieval performance is evaluated between frame-based feature combination and the proposed framework with two different grouping approaches. Experiments are carried out on synthetic and natural image databases and the results indicate that a promising retrieval performance can be obtained as long as a reasonable segmentation quality is obtained. The integration of the region proximity feature further improves the retrieval performance especially for divisible, object-based image content. Finally, frame-based and region-based texture extraction schemes are compared to evaluate the effect of a region on the texture description and retrieval performance utilising the proposed framework. Results show that significant degradations over the retrieval performance occur on region-based texture descriptors compared with the frame-based approaches.