Structure features for content-based image retrieval and classification problems
Gerd Brunner · FreiDok plus (Universitätsbibliothek Freiburg) · 2006
During the past decades we have been observing a permanent increase in image data, leading to huge repositories. Content-based image retrieval methods have tried to alleviate the access to image data. To date, numerous feature extraction methods have been proposed in order to improve the quality of content-based image retrieval and image classification systems. Structure is one of the most important features for image analysis as shown by the fact that the human perception of objects and scenes is to a large extent based on particular spatial configurations and changes in intensity. In this thesis we introduce a structure-based feature extraction technique. The method is capable of representing the global structure of an image, as well as local perceptual groups and their connectivity. The advantage of the method is its broad range of applications and its invariance against changes in illumination and similarity transformations. We first discuss the creation of edge maps accompanied by an evaluation of various edge detectors. Therefore, we present a method that automatically computes the best set of parameters for the Canny edge detector. Secondly, we apply a line segment grouping method based on agglomerative hierarchical clustering, where the segments are extracted with an edge point tracking algorithm. The procedure automatically evaluates the best linkage method and prunes the dendrogram based on a subgraph distance ratio. Once the final clusters are obtained, an intra-class compactness measure is used to discard less significant segment groups. Thirdly, the structure-based features are computed on a global and local scale. The global scale ensures a holistic scene analysis of an image, whereas the local features account for perceptual groups and their connectivity. Finally, we apply the structure-based features to tasks as broad as binary, color, object class and texture image retrieval and/or classification. The first application is the classification and content-based image retrieval of ancient watermark images. The second application is a retrieval task of two color image databases from the Corel collection with 1.000 and 10.000 images. The results are accompanied by an invariance analysis, where our features have obtained a score of more than 96%. The third application is object class recognition and retrieval for the Caltech database, where we achieve a classification rate of 92.45% and 95.45% for the five and three class problem, respectively. The fourth and final application is the classification of textures obtained from the well known Brodatz collection. A support vector machine with an intersection kernel and a leave-one-out test obtained a classification rate of 98%.