الفهرس | Only 14 pages are availabe for public view |
Abstract The ever increasing amount of multimedia data creates a need for new sophisticated methods to retrieve the information one is looking for. Especially for the visual content this is still an unsolved problem. The classical approach <U+2013> manual annotation <U+2013> alone cannot keep up with the rapid growth of available data anymore. Thus contentbased image retrieval attracted many researchers of various fields in an effort to automate data analysis and indexing. There exist many systems for image retrieval meanwhile. However, many lack a clear definition of similarity or rely on noisesensitive preprocessing steps like segmentation or edgedetection. Instead we first define, what shall be considered similar, and then develop robust, compact features that do not require any preprocessing steps. Additionally these can be calculated fast. Generally speaking, the invariant features can be calculated by averaging local image features. However, averaging is too sensitive to occlusion and is not suitable for partial matching. We therefore propose to keep more information about the local image features by counting their occurrence within histograms thus obtaining invariant feature histograms. Color histograms are computationally efficient, and generally insensitive to small changes in camera position. Color histograms also have some limitations. A color histogram provides no spatial information; it merely describes which colors are present in the image, and in what quantities. Color Coherence Vector (CCV) describes a colorbased method for comparing images which is similar to color histograms, but which also takes spatial information into account. CCV classifies each pixel in a given color bucket as either coherent or incoherent, based on whether it is part of a large similarlycolored region. A CCV stores the number of coherent versus incoherent pixels with each color. By separating coherent pixels from incoherent pixels, CCV provides finer distinctions than color histogram. In most applications, image analysis must be performed with as few computational resources as possible. Especially in visual inspection, the speed of feature extraction may play an enormous role. The size of the calculated descriptions must also be kept as small as possible to facilitate classification. Gabor filters have been shown to give good results in comparison to other texture decompositions. There is strong evidence that simple cells in the primary visual cortex can be modeled by Gabor functions tuned to detect different orientations and scales on a logpolar grid. We therefore chose to employ such a set of Gabor functions to characterize texture. Edges in images constitute an important feature to represent their content. Also, human eyes are sensitive to edge features for image perception. One way of representing such an important edge feature is to use a histogram. An edge histogram in the image space represents the frequency and the directionality of the brightness changes in the image. It is a unique feature for images, which cannot be duplicated by a color histogram or the homogeneous texture features. To represent this unique feature, there is a descriptor for edge distribution in the image. This Edge Histogram (EH) expresses only the local edge distribution in the image. Using the features like colors, edges, shapes and textures individually may not be sufficient to represent efficiently the contents of the image. Thus, to improve the retrieval performance, we need to combine various features. |