Intelligent system for content-based image retrieval segmentation.

摘要

Large amounts of digital images are created and accessed daily by the public, academia, and corporations. Keyword indexing is useful but limited in describing image content. Intelligent content-based retrieval is a key technology to address this problem and to facilitate efficient image-based knowledge. This dissertation presents an attempt to improve image segmentation and region-based image retrieval utilizing artificial intelligence methods of probabilistic perspective to achieve this goal. Two novel systems are proposed: fuzzy-logic expert system for objects labeling OLFES and cluster-based retrieval system CoIRS. The two systems are based on probabilistic learning framework called EMIS and are integrated for image segmentation and retrieval. The EMIS is based on Expectation-Maximization (EM) algorithm that estimates Bayesian Maximum Likelihood parameters to fit data into Gaussian Mixture Model. The color and texture features of the image's small patches are fed to EM. The first employment of EMIS is to segment domain-dependent images in two phases. One phase initiates an EMIS application to find K coherent regions, construct K-Leveled image, and obtain K one-leveled subimages of the K-Leveled image. The other phase applies a new one-Leveled segmentation method called EDM (Edge Detection and Morphology) for refining the K one-Leveled subimages segmentation. The OLFES system identifies, labels, and explains the decision-making of segmenting the domain-dependent images into categories using the K-Leveled image histogram's maximum bins. The second employment of the EMIS is to cluster general-purpose images for CoIRS design. For each image, a privileged component called Cluster Signatures (CS) is constructed and utilized to search for similar images. Six features of CS are color and texture features of the K clusters' centroids produced by EMIS, and another seven CS shape features are the Hu's invariant moments of the K one-Leveled subimages. Another distinctive aspect is CS Feature ranking (FR) which brings the closest features in the same rank and put more emphasis on the ranked feature. CoIRS is evaluated and compared with another system by a database of 2000 images using precision estimation. The database is composed of different categories of images and provides successful retrieval results estimated by precision calculation. This dissertation presented efficient intelligent-directed methods based on probabilistic-model learning for improving segmentation, building a rule-based decision-making system for labeling objects, and developing a content-based retrieval system.