Texture Representations for Image Retrieval Open Access
Downloadable ContentDownload PDF
Image databases are used in medical research, hospitals, in scientific research, in museums to catalogue content, in security and biometrics and many other fields. The entire World Wide Web can be considered an image database from the point of view of a search engine. Image retrieval involves searching and retrieving specific or similar images from a database given a query image or textual descriptor. Content-based image retrieval uses low-level visual features, such as color, shape and texture, in digital images to search and retrieve content from an image database.Texture plays a major role in the human visual system for image understanding and object recognition and is found, to some degree, in all natural images. The thesis explores texture representation in general and texture feature extraction methods for image retrieval. Three texture representations in the spatial, spatial-frequency and sparse domains are discussed and implemented to extract feature vectors from texture images. Sparse representations are invaluable to many signal processing and computer vision algorithms. Sparse coding and dictionary learning algorithms are discussed and implemented with the goal of evaluating the merit of sparse representation for image retrieval.To compare the performance of these three texture feature methods a small texture image database is implemented using images from USC's SIPI and MIT's VisTex texture databases. Three measures are used to compare retrieval performance of the extraction methods. The results show that the sparse image retrieval method performed favorably but has room for improvement in both performance and complexity.