Similarity Evaluation in Image Retrieval Using the Hough Transform

This work investigates the use of orientation features, computed using the Hough transform, as a criterion for image similarity evaluation in content based picture retrieval. The context of this work is the management of thematic catalogues, in which the coherence in the meaning of the image contents can be relied upon to a certain degree. The vector space model, a well-assessed technique in textual information retrieval, is utilized for the retrieval model. Introduction Current technology allows the acquisition, transmission, storing, and manipulation of large collections of images. Yet systems for their classification and retrieval still rely heavily on textual descriptions associated to images. Recently a number of methodologies, techniques and tools have been studied for identification and comparison of images features in order to develop classification and retrieval systems based on (almost) automatic interpretation of image contents [1-4]. This work focuses on the use of orientation features as a criterion for similarity based retrieval; the Hough Transform [5] algorithm is used for features extraction. We do not claim that the analysis of the image directional features can fulfill the requirements of a content based image retrieval system, but we argue that this feature is well suited for evaluating the visual coherence of a set of non trivial, semantically homogeneous images. Motivations and goals The approach we consider in this paper is oriented to retrieving images from a thematic database, where the semantic content of the images is limited to a specific domain. Most image collections available in the public domain or through the commercial and professional distribution channels are organized in sub-collections (directories), each covering a separate theme. The case study we discuss is a small database of eighty public domain images picturing planes. Most of them are true photos, someone is a (good) painting. Size, resolution and number of colors are very different. Also image quality is not constant, and in some cases quite low. For space reasons, in this paper a set of only 16 images from the test set is used as an example, see figure 1. While very small, the set is representative of the different images contained in the whole collection. The application of the retrieval model is illustrated through the similarity measures found among the images, and with sample sketches standing for user graphical queries. Database population and image analysis Various models have been proposed for similarity analysis in image retrieval systems. In this work we investigated the application of the vector space model [6], that has been widely used in conventional document retrieval systems. The basic model considers the same N elements term set to identify both stored data and queries, that can be represented as vectors in a N-dimensional space. Our approach assumes as index terms the + Current address: Università Ca’ Foscari di Venezia, Dipartimento di Matematica Applicata e Informatica, Venezia, Italia Work supported in part by the Italian MURST in the framework of the project “Basi di dati evolute: Modelli, Metodi e Strumenti”. Published in Journal of Computing and Information Technology, 4 (3), 1996