RETRIEVAL OF IMAGE DATABASES

Abstract

With the spurt in computing and communication technology, more and more images are being captured, stored and are widely used in multimedia collections such as in medical imaging, geographic information systems, entertainment, education etc. Image Retrieval Systems are required to utilize such collections of images efficiently. The work presented in the thesis is an effort to propose different mechanisms and techniques for efficient retrieval of image databases using low level image features such as texture, color and shape. In the first part, texture feature extraction method based on Haar Wavelet Transform is presented. The image is divided into sub-images and clustering is performed to cluster sub-images with similar characteristics. As texture feature extraction is very difficult, it is shown in research that neural network can be a useful tool for feature extraction. Another technique, which employs multi-layer perceptron based neural network for texture feature extraction and classification is proposed. The proposed schemes are able to extract texture features and the results obtained on images from Brodatz texture album are found to be reasonable and acceptable. Color is an important image feature, which is used in most of the image retrieval techniques. As user normally submit query in natural language and is not aware of low level image features. It is very difficult for a user to associate numeric values with the image features. So a technique for color based image VII retrieval using fuzzy logic for defining the "imprecision" is proposed next. The algorithm has been tested on colored images from Kodak album. For accuracy in retrieval, we need to extract multiple features for querying the database. A number of studies have been carried out which combine the various features for efficient and effective querying by image features. A technique, which computes the integrated feature vector, which combines the texture and color features, is presented in the next part and is able to extract regions, based on texture and color. Next, we have attempted to combine the color feature with the shape feature. The technique indexes the images on the basis of dominant color and then the local shape feature turning angle is used to perform shape based retrieval. This scheme automatically filters out the images on the basis of dominant color, which helps in fast retrieval of images. We compare the performance of the proposed technique with that of the Fourier descriptor based method, andGrid basedmethod, and found that the precision Vs recall is improved. Next, we address the issue of indexing the image databases. The variable bin allocation method can store the color information in compact form as compared to global color histogram. The method presented employs a parallel approach for the frame slice signature file using the variable bin allocation technique for representation ofcolor image to speed up the retrieval ofimage databases. We have compared the performance ofthe proposed technique with that of the S-Tree parallel traversal implementation for color based image retrieval. The proposed approach has improved speedup performance. We have carried out a case study on the real MRI images obtained from PGI, Chandigarh, biomedical images from National Technical Information Services Springfield U.S.A., and some medical images downloaded from the Internet on which Vlll the technique, which computes the integrated feature vector for texture and color, is implemented. This technique can be useful to the medical community for better diagnostics of abnormalities present in medical images based on texture and color.