Generic and fully automatic content-based image retrieval using color, shape, and texture

Image Retrieval (IR) problems are concerned with retrieving images that are relevant to a user's requests from an image database. A content-based IR (CBIR) system helps the user retrieve images relevant to their query based on the contents of the images. An attribute-based CBIR models image contents as a set of attributes extracted manually and managed within the framework of conventional DBMS. However, it lacks flexibility as it can not capture the rich contents of images and relationships between objects in an image completely. A feature-based CBIR approach uses computable and indexable low-level features, such as color, shape, and texture.; I have proposed and evaluated a fully automatic and generic feature-based CBIR. In most existing feature-based approaches, the "real" interimage distance function is non-linear and expensive for on-line retrieval from databases having a large number of images. In the proposed approach, the image retrieval is done in two phases of image population and retrieval. In the image population phase the real interimage distance is derived by using low-level features. The real interimage distance is defined according to which a user wants to perform the retrieval. They are used to compute a vector of composite features for each image and a training set. The training set is used to compute the feature vectors for images added on-line or a query image. This approach is very efficient as it has only to compute real distances between the query image and images in the training set, not all the images in the database. The process is fully automatic when low-level features are used for the real interimage distance. The method is generic since the processing steps remain largely the same for different choices of image content.; I have tested our approach using color, shape and texture separately as the low-level features. Experimental results indicate that our approach is very successful in reducing the feature space size without compromising the accuracy. The ranking of images using real interimage distances are preserved when estimated interimage distances are used. This enables us to perform nearest neighbor searches.