| Electronic Commerce has propelled the growth of buying and selling of a variety of products through the Internet. Often, the electronic catalogs contain pictures comprising of images of these products, either shown exclusively or combined with other products. Despite substantial research, facilitating search of an object based on the content of an image still remains as a major challenge. This problem of content-based image retrieval is even more difficult when multiple objects are present in an image or when they are partially occluded. While the main focus is on the recognition of partially occluded objects, in this dissertation, we make several contributions to the area of content-based image retrieval.; First, we have proposed feature extraction methods to support content-based image retrieval based on individual objects in images. We have employed rotational invariants using Run Length Code lines to extract shape parameters, and texture and shape features to retrieve parameterized image parameters such as roundness, form factor, aspect ratio, surface regularity, angle of second moment, entropy, contrast, and mean. RLC lines can be directly calculated for area and position measurements with less arithmetic than pixel array techniques. RLC lines can distinctly represent shape parameters, which are better discriminating features than parameters such as color and texture for recognizing image objects in electronic catalogs.; Second, we have proposed an object classification method, called Probability Interval Classification (PIC), using novel splitting rules, which are based on minimizing the sum of variance and maximizing the difference of probabilities of intervals. The splitting rules are given by considering the probabilities of pre-assigned intervals for covariates rather than exhaustively searching over all possible splitting values. It provides the user with control of the accuracy of the tree by adjusting the number of intervals. The results show that PIC is more stable than the exhaustive search method when the learning sample changes, produce better accuracy with reasonable tree size and depth, and demonstrates higher precision and recall scores.; Third, we have developed methods to reconstruct and estimate partially occluded shapes and regions of objects in images from overlapping and cutting. We have presented two robust methods for recognizing partially occluded objects based on symmetry properties. One is based on the contours of objects, and the other is based on individual parts of objects. Our methods have provided simple techniques to reconstruct occluded regions via a region copy using the symmetry axis within an object. Based on the estimated parameters for partially occluded objects, we have performed object recognition on the classification tree. Since our method relies on reconstruction of the object based on the symmetry rather than statistical estimates, it has proven to be remarkably robust in recognizing partially occluded objects in the presence of scale changes, rotation, and viewpoint changes. |
