Image Retrieval by Utilizing Structural Connections within an Image

Published Online January 2016 in MECS

Aiming to improve the efficiency of the image retrieval process, an image retrieval system proceeds toward a technique based on image contents known as contentbased image retrieval (CBIR).[1-7] In CBIR, the images are retrieved by their primitive visual contents such as color, texture, shape, and spatial layout. Multidimensional feature vectors, analogous to the visual contents of images, are extracted and stored in a feature database to build an image feature database. The example image is provided as a query by the user to retrieve relevant images. The system then converts these examples into its internal representations of feature vectors. The similarities between the feature vector of the query image and the database images are computed, and the retrieval is performed using an indexing scheme. This search is usually based on a similarity index rather than an exact match, and the retrieved results are then ranked according to the similarity index [1]. The general architecture of a CBIR system is shown in Fig. 1.

Fig.1. General Architecture of A CBIR System

The main concern in CBIR is the need for an effective and efficient feature extraction method for image representation, which conforms to the subjective human perception. This subjectivity transpires at all semantic levels while analyzing images because different users in the same situation or the same user in different circumstances may investigate or classify the same image differently. This inconsistency between image retrieval, by using low-level image features and high-level human semantics, is termed as the “semantic gap” [8-11].

In this study, we limited our goal to developing a general comprehensive algorithm for retrieving the images based on primitive visual features. Color is the most expansively used visual feature for image retrieval. Color features are relatively robust to the viewing angle, translation, and rotation of the regions of interest in an image. Although color is definitely not the most important visual quality of image data, it is a primitive image feature constituting a good launching point for developing a more sophisticated image retrieval system. In general, two types of image features are used to describe an image: (1) color features and (2) holistic structure features. The difference between these features is not always distinct. If spatial distribution is considered when extracting color features of an image, then the color features can be considered as holistic structure features. Thus, we used an approach to retrieve images based on color feature extraction by using two color descriptors: color histogram (CH) and color coherence vector (CCV). In addition, the local binary pattern (LBP), which amalgamates several orientations while keeping a low feature size, was used. The LBP transforms the relationship between two color levels into two levels (“0” or “1”) depending on a threshold represented by a binary code. This relationship of a pixel with eight neighbouring pixels is then transformed into a sequence of binary codes, resulting in an LBP level in the range of 0–255.

The CH [12, 13] represents the distribution of color contents effectively in an image when the color pattern is unique compared with the remaining the data set. The CH is easy and fast to compute and very robust for translation and rotation about the view axis. It is used for image retrieval by many commercial systems [14], such as QBIC, and academic systems such as NETRA, RETIN, KIWI, and Image Minor. Color moments (CMs) represent the color distribution of the images and are extensively used in QBIC, especially when an image comprises an object. Mostly, an image may not solely be described by its CH or the color distribution in it. One image may have a large number of scattered pixels with a particular color, whereas another may include the same amount of pixels with the same color but congregated in one area. Two images may have the same CH, but a different spatial color arrangement. This information about the spatial location of colors in an image is represented by another color descriptor known as CCV [15]. The results of this study demonstrate that the retrieval efficiency (73.8%) improves if we consider all the different aspects of an image by using the CH, CCV, and LBP.

The paper is organized as follows: Sec. 2 provides a rigorous literature survey. The methodology used in this paper is presented in Sec. 3. Other existing methods are mentioned briefly in this section as well. Sec. 3 also describes the proposed algorithm, whereas the experimentation is given in Sec. 4. The results of the proposed algorithm and their comparison with the existing methods are discussed in Sec. 5. Finally, Sec. 6 concludes the paper and throws light on the scope of future work.

• II.    Literature Survey