Article Preview
Top1. Introduction
Rapid increase of image based internet services, online/offline applications using image records have put forth a need for a robust and efficient system for image retrieval from large-scale image databases (Balamurugan, 2008). Using manual text annotation to assist retrieval of images is time consuming and impractical considering the large image data size. In general, content based image retrieval (CBIR) systems estimate the visual similarity that exist between certain user defined query image/images and images resident in databases. CBIR considers a set of user defined features to explore and retrieve visually similar images from huge database contents (Jadhav &Aehmad, 2012). CBIR is used in various applications such as the internet or online image retrieval, multimedia mining and allied utilities, biomedical image information retrieval, entertainment, digital libraries and crime avoidance etc. (Mumtaz, 2006). The prime function of CBIR systems is the extraction of significant low-level visual characteristics of images (also called image features). Using image features, similarity estimation with images (also represented as image features) in the database is achieved. Normally, the images are characterized in terms of various low-level visual features, such as shape (Ferrari, 2008), colour (Ott, 2009), texture (Hadid, 2004), background, orientation (Dalal, 2005), etc. Arbitrary environments, the non-rigidness nature of objects, higher degree of articulation, varied postures, cluttered background, lighting conditions and viewpoints are a few problems that exist in developing CBIR systems. Numerous works have been proposed in recent years for efficient image information retrieval to overcome these problems. Based on existing literature reviewed, limited work is carried out considering shape characteristics of images in CBIR systems. Drawing motivation from this a shape descriptor based CBIR system is proposed in this paper.
An image in CBIR is represented as a set of features. Colour and texture features are critical, but shape features are the most significant to understand content (Sokic, 2014). Shape descriptors are broadly classified as region and contour based (Sikora, 2001). Contour based shape descriptors are quick and compact when compared to the region based descriptors. The contour based methods developed are found to be sensitive to distortions, noise and variations in scale, translation and rotation are not accurately represented (Sokic, 2014). To overcome these drawbacks the use of Fourier transforms (Sikora, 2001), Wavelet Transforms (Sokic, 2016), Radon transforms (Zhang, 2015) and their combinations (Yadav, 2007) is used. The major drawback of these shape descriptors is that they are predominantly contour based and region based features / local features are not comprehensively represented (Sokic, 2014). The evaluation of shape descriptors discussed above is predominantly carried out using binary images. Limited evaluations are carried out using colour images.