University of Crete (UOC)
Perceptually relevant mechanisms for the description and retrieval of visual information
Abstract
dc:descriptionThe large volume and variety of digital images, currently acquired and used in different application domains, has given rise to the requirement for content-based image management and retrieval techniques. In particular, there is an increasing need for the development of automated image content analysis and description techniques in order to retrieve images efficiently from large collections, based on their visual content. In this dissertation, mechanisms for the perceptually relevant description and retrieval of visual information are presented and discussed, motivated by the need to provide a better match between content-based image retrieval results and end user expectations. The proposed mechanisms concern the description of primitive visual features and spatial arrangements of such features, and emphasize the representation of this information with respect to scale of observation. This scale dependent representation is subsequently used to extract image regions that exhibit a characteristic spatial arrangement of primitive features and identify gradient-derived dominant structural elements, which are both known to be significant descriptive components of visual content. The organization of gradient-derived dominant structural elements into perceptual groups yields an additional component of visual content. Attributes of such perceptual groups are then integrated with information about the spatial arrangement of primitive visual features and used in the description and content-based retrieval of images. Initially, the role of primitive visual features in the formation of image content is considered and a physiology-inspired method is presented for their representation, based on the scale-summarization of visual content. The proposed representation utilizes the scale-normalization of feature detection response functions to summarize visual feature information from a range of scales into a single image. This scale-summarized representation facilitates the introduction of a method for the description of primitive features at the range of image scales at which they occur, the extraction of more than one meaningful ranges of scales from an image, and the classification of primitive visual features with respect to the range of scales at which they occur. The additional information thus generated is demonstrated to be useful in the description of image content, as well as in a number of image processing tasks. Furthermore, the scale-summarized representation can be computed in parallel and exhibits computational and descriptive properties that extend the standard representation of visual information with respect to scale. The study of primitive visual features is concluded with a discussion about their role in content-based image retrieval. Subsequently, the investigation for the perceptually relevant description and retrieval of image content is focused on the ability to extract and compare, with respect to their visual similarity, image regions that exhibit a constant spatial arrangement of primitive visual features. Using local descriptors with varying image sampling aperture, a multiscale representation of the spatial arrangements of primitive features is derived. The extension of the scale-summarized representation for local descriptors makes their scale-normalization possible. This scale-normalization can then be utilized for the constant description of scale-varying spatial arrangements of primitive features. The clustering of scale-normalized local descriptors facilitates the extraction of image regions that exhibit a constant spatial arrangement of primitive features, even when these features vary at scale. In addition, the scale-summarized representation of spatial arrangements of primitive features results in reduced memory capacity requirements. Furthermore, attributes of local descriptors are proposed for the refinement of the description of spatial arrangements of primitive features and are used as predicates in the formulation of visual queries. Such attributes are mapped onto visual properties of images to provide visual queries which are comprehensible by end users. Finally, the acquired representation of spatial arrangements of primitive features is utilized for the browsing and retrieval of visually similar images. In order to further enrich the derived description of image content with perceptually relevant attributes, the component of visual content resulting from the process of perceptual organization is considered and two approaches are presented for the extraction and description of two classes of perceptual groups. These are the class of linearly perspective perceptual groups and the class of silhouette boundaries. Regarding the first class, perceptual groups that consist of converging line segments are extracted from images, based on hypotheses about their convergence to a vanishing point. Such hypotheses are initially formulated, based on the contrast and the size of line segments in an image and subsequently tested, as to their validity, utilizing supporting or contradicting image evidence. Line segments for which the hypothesis is verified are assigned to the same group along with appearance-related attributes of that group. The same algorithm is generalized for subjective line segments that is line segments that are composed of collinear local image features, such as corners or dots. The detected perceptual groups as well as their attributes are demonstrated to be of use in the content-based retrieval and classification of images. With respect to the second class of perceptual groups, an approach is presented for the description and retrieval of silhouette boundaries, which utilizes a curvature-based method to detect perceptually significant and computationally stable anchor points. In particular, curvature extrema are tracked along the scale-space of the contour and their scale-normalized curvature across scale is utilized to formulate a salience metric. Using the most salient contour points, a piecewise decomposition of contours is achieved and further utilized in their alignment. The ability to align contours is finally utilized in matching similar contours, based on a similarity metric which captures information about the relative displacement of aligned (corresponding) contour segments (pieces). The integrated use of information about spatial arrangements of primitive visual features and perceptual groups in content-based visual information retrieval is made possible by the introduction of a high specificity user interface, which facilitates the selection of individual visual content components and their attributes and their use as predicates in visual queries. In addition, the iterative introduction of image classification criteria is proposed as a method for integrating information about image content and interactively classifying image collections.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2002
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Ζαμπούλης, Ξενοφών
- Zabulis, Xenophon
Subjects
dc:subject × 14Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/13369