Shape analysis for image retrieval

Abstract

Content-based image retrieval (CBIR) system is designed to help retrieve relevant images in an image database based on their image contents. This system will allow queries on large image databases based on example images, user-constructed sketches and drawings, and other graphical information. Different image features, or descriptions, may have different significance and effectiveness in the interpretation and representation of images in different applications. The Moving Picture Experts Group (MPEG) of the International Standards Organization (ISO) initiated the MPEG-7 standard, which provides standardized core technologies that allow for the description of audiovisual data content in multimedia environments. The most challenging technical issues for a CBIR system are the effectiveness and efficiency of feature extraction and recognition algorithms for content-based image retrieval. The objectives of this thesis are to investigate and develop efficient techniques for shape feature extraction, and to construct a content-based image retrieval system. An introduction to the general concept of image retrieval will be given in this thesis, and the recent development of the MPEG-7 standard will be described. Existing content-based image retrieval systems, and the feature extraction and recognition techniques based on color, texture, shape and motion will be reviewed. Furthermore, more efficient and effective features will be proposed so that a reliable and practical retrieval system becomes possible. Shape descriptors, which are high level descriptions, will be emphasized in this research work. In this research, the content-based image retrieval system developed consists of three major parts: boundary extraction, feature extraction and recognition. The first part is based on an active contour model for representing image contours. We have proposed an efficient active contour model which can represent highly irregular boundaries. The contour points can be used to form other shape descriptors such as chain code, curvature scale-space representation, skeleton, etc. After extracting the boundaries, the second part is skeletonization which is an important process that can provide a compact shape representation. We have proposed a fast, efficient and accurate skeletonization method for the extraction of a well-connected Euclidean skeleton based on the boundary information. The skeleton feature can be used as a shape descriptor, which can represent the shape more compactly, and consists of spatial and structural information. In the third part, we have proposed a robust and efficient histogram representation scheme for shape retrieval, which is based on the normalized maximal disks used to represent the shape of an object. The maximal disks are extracted by means of the fast skeletonization technique with a pruning algorithm. The logarithm of the radii of the normalized maximal disks is used to construct a histogram to represent the shape. The proposed representation scheme outperforms the other methods under affine transformation, different distortions and noise levels. Hence, these three major parts are integrated to form a complete system for content-based image retrieval. We have also devised a contour/region-based matching algorithm has been used for retrieving relevant images containing similar shapes from a database. In the algorithm, Hausdorff distance is used to measure the similarity of two point sets. We have devised a robust line-feature-based approach for model-based recognition based on this distance measure. The proposed algorithm can achieve a good performance level in matching, even in a noisy environment or with the existence of occlusion, and can be used as a similarity measure for image retrieval.