Image retrieval based on index compressed vector quantization
Introduction
Image-based applications have significantly increased in recent years. A problem associated with image- based applications is the high storage capacity requirement. A solution is to compress the images. Most applications such as digital libraries, image search engines and medical images require effective and efficient image retrieval techniques to access the images based on their contents. Most of the compression schemes transform images from pixel into other domains. There is no image analysis algorithm to extract the important image features, such as edges, in the transformed domain. Therefore, it is desirable to present approaches to combine image indexing and image compression techniques.
The goal of image compression is to remove the image redundancy and represent the image in a compact form; the compressed image. The image indexing techniques extract a compact code (feature vector) from the image [1], [2]. Therefore, from the compact image representation viewpoint, image compression and image indexing are similar and, it is possible to develop image indexing techniques in the compressed domain. Incorporating these areas has several advantages. First, since there is less data in the compressed domain than the pixel domain, there is potential for reducing overall computation. Second, applying feature extraction techniques in the compressed domain avoids the overhead of unnecessary decompression operations. Third, many compression algorithms actually reveal some features of the image implicitly, which provide good foundations for image content description and improve the effectiveness of the image retrieval techniques [3], [4], [5]. Consequently, although the compressed domain imposes many constraints, it provides great potential for state-of-the-art image retrieval schemes [6].
Vector quantization (VQ) is an efficient, simple and attractive image compression scheme [7], [8]. VQ is inherently an indexing technique. Hence, it has become a promising approach for integrating image compression and image indexing techniques. We can describe VQ as a mapping of a k-dimensional vector space on a set with finite members, the codebook. Each codebook member (codevector) represents a vector from a k-dimensional space. Associated with each codevector is an index, which the decoder uses in indexing the codebook to reproduce the codevectors [7], [8]. The basic VQ scheme partitions the image into small blocks (vectors). Each vector is separately encoded and assigned an index. The compressed form of the image is a file of indices. These indices describe the image blocks characteristics; the correlation between the pixels within the block (the intra- block correlation) [7], [8]. Therefore, one can employ them in developing effective image indexing schemes [9]. If we consider the color information of a pixel as an index of a block of an image with 1×1 size, then the pixel domain color-based indexing is similar to the compressed domain VQ-based indexing. The color values of the image pixels ignore the intra-block correlation. Hence, the VQ-based indexing techniques provide descriptions that are more precise.
Idris and Panchanathan [10], [11], [12], [13] proposed two VQ-based image indexing schemes. They employed histogram of indices in one scheme and usage map in another one as feature vectors. These methods describe the entire image content by generating a proper histogram for the image, without considering the block's location. Panchanathan and Huang [14] proposed an image indexing technique based on vector quantization of color images. They improved the effectiveness of retrieval using color information, without including the relation between the image blocks. The image blocks’ relation, the inter-block rather than the intra-block correlation alone, provides most of the image information that satisfies human perception.
Jiang [15], [16] proposed a weighted codevector counting method, a VQ-based image indexing scheme, which indirectly reflects the image block location in the histogram of labels. Jiang considers the counting of each codevector with a weighting scheme by the number of consecutive usage of the same codevector. This scheme improves the effectiveness compared to the previous methods; however, it suffers from selecting a non-linear weighting function and high computational cost in feature vector generation.
An image compression technique with the capability of exploiting the inter-block dependency of the whole image can be a suitable candidate for developing an image retrieval scheme. Several VQ-based image compression techniques have considered the inter-block correlation [17], [18], [19], [20], [21]. A fast and simple group of these methods is index-compressed VQ (IC-VQ) proposed by Shanbehzadeh et al. [19], [20], [21]. This scheme, despite its simplicity, significantly improves the compression ratio of basic VQ by lossless coding of the indices generated by tree structured VQ (TSVQ); TSVQ is one of the VQ variations where the search is performed in a tree-shaped manner to reduce the search time. The lossless scheme employs the inter-block dependency in the entire image.
The indices generated by IC-VQ have two interesting properties. First, they show the information about the relation between two neighboring image blocks (inter-block correlation). Second, the indices show the characteristics of the neighboring pixels, such as edges or smooth image areas. This paper has selected the IC-VQ compression technique to exploit the intra- and inter-block correlation of image blocks and develop an effective image retrieval scheme.
The organization of the paper is as follows. Section 2 presents the feature vector generation based on the characteristics of indices of an IC-VQ compressed image. It also describes IC-VQ and its differences with VQ. Section 3 discusses the experimental results with examining the elements of an image retrieval test-bed and comparing the performance of the new scheme with some of the existing methods. The final section explains the concluding remarks.
Section snippets
IC-VQ-based image retrieval
Feature extraction is the basis of content-based image retrieval systems. The feature vector describes the content of an image and provides discrimination between two images, so that the distance between the feature vectors of two images resemble their dissimilarity. This section briefly explains the IC-VQ image compression scheme and then discusses the IC-VQ-based feature vector extraction method.
Experimental results
This section presents the simulation results and discussion of the new scheme in five subsections. The first subsection illustrates the most significant components of our image retrieval test-bed. The second subsection presents the results of the IC-VQ-based image retrieval. This subsection discusses the effect of correlation in image retrieval system. The third subsection compares the results of image retrieval system when applying two independently obtained query image sets. The next
Conclusions
Today, most of the image databases store images in the compressed form, and there exist compression algorithms that reveal image features implicitly. This characteristic of compression schemes provides suitable foundations for image content analysis and improves the effectiveness of the image retrieval techniques. IC-VQ is one of the compression algorithms with the capability of preserving image content in the compressed form.
This paper introduced a new image retrieval scheme based on the IC-VQ
About the Author—AMIR-MASUD EFTEKHARI-MOGHADAM was born in Tehran, Iran in 1964. He received post diploma in electronics from Amirkabir University of Technology, Tehran, Iran in 1986, B.E. and M.Sc. degrees in computer hardware engineering from Iran University of Science and Technology in 1992 and 1995, respectively, and Ph.D. degree in computer engineering from Islamic Azad University, Science and Research branch, Tehran, Iran, in 2002. In 1996, he was a lecturer of computer engineering at the
References (29)
- et al.
Image retrievalcurrent techniques, promising directions and open issues
J. Visual Commun. Image Representation
(1999) - et al.
Index compressed image adaptive vector quantization
Signal Process. Image Commun.
(1996) - et al.
Performance evaluation in content-based image retrievaloverview and proposals
Pattern Recognition Lett.
(2001) - et al.
Image retrievalpast, present and future
J. Visual Commun. Image Representation
(1999) Compressed domain techniques for image/video indexing and manipulation
Proc. IEEE Int. Conf. Image Process.
(1995)- W.B. Seales, C.J. Yuan, M. Brown, Efficient content extraction in compressed images, IEEE Workshop on Content-Based...
- et al.
Content-based image retrieval using block-constrained fractal coding and nona-tree decomposition
IEE Proc. Visual Image Signal Process.
(2000) - J. Shanbehzadeh, A.M. Eftekhari-Moghadam, F. Mahmoudi, Image indexing and retrieval techniques: past, present and next,...
Vector quantization
IEEE Acoust. Speech Signal Process. Mag.
(1984)- et al.
Vector quantization in speech coding
Proc. IEEE
(1985)
Storage and retrieval of compressed images
IEEE Trans. Comput. Electron.
Cited by (7)
Image indexing and retrieval in JPEG compressed domain based on vector quantization
2013, Mathematical and Computer ModellingCitation Excerpt :Therefore, from the compact image representation viewpoint, image compression and image indexing are similar and it is possible to develop image indexing techniques in the compressed domain. It has several advantages [10]. First, since there is less data in the compressed domain than the pixel domain, there is potential for reducing overall computation.
Combining regional and global features for automatic image annotation based on VQ
2012, AISP 2012 - 16th CSI International Symposium on Artificial Intelligence and Signal ProcessingLossless index coding for image vector quantization using huffman codes
2011, International Journal of Innovative Computing, Information and ControlResearch on anomaly detection of network traffic based on fractal technology and vector quantization
2010, 2nd International Workshop on Education Technology and Computer Science, ETCS 2010Improving VQ index compression with modified relative index table
2010, Proceedings - 2010 1st ACIS International Symposium on Cryptography, and Network Security, Data Mining and Knowledge Discovery, E-Commerce and Its Applications, and Embedded Systems, CDEE 2010Embedding image on vector-quantized index file
2006, WSEAS Transactions on Information Science and Applications
About the Author—AMIR-MASUD EFTEKHARI-MOGHADAM was born in Tehran, Iran in 1964. He received post diploma in electronics from Amirkabir University of Technology, Tehran, Iran in 1986, B.E. and M.Sc. degrees in computer hardware engineering from Iran University of Science and Technology in 1992 and 1995, respectively, and Ph.D. degree in computer engineering from Islamic Azad University, Science and Research branch, Tehran, Iran, in 2002. In 1996, he was a lecturer of computer engineering at the Islamic Azad University, Qazvin branch, Iran. Since 2002, he has been with the Iran Telecommunication Research Center, Tehran, Iran. His research interests include image compression, retrieval and indexing. Dr. Eftekhari-Moghadam has published about 20 papers in scientific journals and conference proceedings.
About the Author—JAMSHID SHANBEHZADEH was born in Abadan, Iran, in 1960. He received B.E. and M.Sc. in electrical engineering from the University of Teheran, Tehran, Iran and Ph.D. in electrical and computer engineering from Wollongong University, Australia. Since 1996, he has been with the Department of Computer Engineering at Teacher Training University, Tehran, Iran. Dr. Shanbehzadeh is interested in image processing, coding, retrieval, indexing and information technology. He has published about 30 Journal and conference papers on image coding and retrieval.
About the Author—FARIBORZ MAHMOUDI was born in Tehran, Iran, in 1966. He received post diploma in electronics from Amirkabir University of Technology, Tehran, Iran, in 1986, B.E. degree in computer software engineering from Shahid Beheshti University, Tehran, Iran, in 1991, M.Sc. degree in computer architecture engineering from Amirkabir University of Technology, Tehran, Iran, in 1994, and Ph.D. degree in computer engineering in 2002 from Islamic Azad University, Tehran, Iran. In 1996, he was a lecturer of computer engineering at the Islamic Azad University, Qazvin, Iran. Since 2002, he has been with the Iran Telecommunication Research Center, Tehran, Iran. His research interests include image retrieval and indexing and AI. Dr. Mahmoudi has published about 20 papers in scientific journals and conference proceedings.
About the Author—HAMID SOLTANIAN-ZADEH was born in Yazd, Iran in 1960. He received B.S. and M.S. degrees in electrical engineering: electronics from the University of Tehran, Tehran, Iran in 1986 and M.S.E. and Ph.D. degrees in electrical engineering: systems and bioelectrical science from the University of Michigan, Ann Arbor, MI, USA in 1990 and 1992, respectively. From 1985 to 1986, he was with the Iran Telecommunication Research Center, Tehran, Iran. In 1987, he was a lecturer of electrical engineering at the University of Tehran, Tehran, Iran. Since 1988, he has been with the Department of Diagnosis Radiology, Henry Ford Health System, Detroit, Michigan, USA where he is currently a Senior Staff scientist. Since 1994, he has been with the Department of Electrical and Computer Engineering, the University of Tehran, Tehran, Iran where he is currently an Associate Professor. His research interests include medical imaging, signal and image reconstruction, processing, and analysis, pattern recognition, and neural networks. Dr. Soltanian-Zadeh has contributed to about 200 papers published in scientific journals or conference proceedings.