Elsevier

Image and Vision Computing

Volume 28, Issue 8, August 2010, Pages 1293-1302
Image and Vision Computing

An adaptive data hiding scheme with high embedding capacity and visual image quality based on SMVQ prediction through classification codebooks

https://doi.org/10.1016/j.imavis.2010.01.006Get rights and content

Abstract

This study exploits the characteristics of image blocks to develop an adaptive data hiding scheme that is based on SMVQ prediction. Since human beings’ eyes are highly sensitive to smooth images, changes in smooth cause great distortion and attract the attention of interceptors. Hence, this study proposes a data embedding scheme for embedding secret data into edge blocks and non-sufficiently smooth blocks. The experimental results show that the proposed scheme improves the quality of the stego-image and the embedding capacity.

Introduction

The rapid development of the Internet and multimedia techniques has caused the hiding of data in digital media to attract increasing attention. Many researchers have studied watermarking [1], [7], [13], [17] and data embedding. Watermarking protects the copyright of multimedia products, while data embedding securely delivers invisible secret messages that are hidden in multimedia. The latter scheme is generally referred to as steganography.

Recently, some steganographical approaches have integrated data embedding with image compression approaches. They include vector quantization (VQ) [5], [6], [14], [15] and side-match vector quantization (SMVQ) [4], [8], [10], [18]. Such approaches are appealing because they reduce transmission costs and the awareness of hidden contents.

VQ compression is an effective compression method. Before VQ encoding, a codebook is trained with various images to get the representative blocks. These representatives in a codebook are called codewords. The LBG algorithm [13] typically is adopted to generate the codebook. In the encoding procedure, the cover image is divided into many non-overlapping blocks, each of which will be given an index value, which is determined from the minimal Euclidean distance between the input block and each codeword in the codebook. Each index value is associated with the all pixels of a block. Finally, the index table, instead of the cover image, is transmitted to the receiver. Obviously, VQ approaches can reduce the cost of transmission. To recover the image, the sender and receiver must have the same codebook. In the decoding procedure, the receiver looks up each index value in the codebook to find the corresponding codeword, which is used to recover an image block. Lin and Wang proposed a VQ-based image hiding scheme [12]. First, the codebook, represented by Y = (y1,y2,y3,,yn), is rearranged such that neighboring codewords are as similar as possible. Next, the codebook is partitioned into two sub-codebooks such as Y0 = (y1,y3,,y2×n/2+1) and Y1=(y2,y4,,y2×n/2). For each input block of a cover image, if the secret bit is ‘0’, then the nearest codeword of this block in Y0 is found and the corresponding index value instead of the block is transmitted to the receiver; otherwise, Y1 is used. By switching two sub-codebooks, the secret bit is carried by index value. Accordingly, embedding the secret information does not enlarge the compressed image. In extracting procedure, the embedded secret bit “0” or “1” can be extracted according to the index value that belongs to Y0 or Y1 while the image block can be recovered at the same time. Although the VQ compression approach has a high compression rate and can be adopted to carry secret data, this approach suffers from the block effect problem. To solve the block effect problem and improve embedding capacity, some extended schemes that combine the VQ technique and the SMVQ technique are presented.

Shie et al. [16] proposed an adaptive data hiding scheme for embedding secret data in a SMVQ compressed image. The side-match distortion among the currently processed block and the upper and the left reconstructed blocks is measured and compared with a pre-determined threshold, THsmd, to determine whether an image block is sufficiently smooth. Shie et al. only hid secret data in so-called “sufficiently smooth blocks” in the cover image. However, since human eyes are quite sensitive to the smoothness of image blocks, changes in these blocks attract the attention of interceptors. Moreover, the need for imperceptibility limits the embedding capacity. Therefore, the embedding capacity and the image quality achieved by Shie et al.’s scheme leave room for improvement. More details of this scheme can be found in [16].

This paper presents an adaptive data hiding scheme based on SMVQ prediction. The proposed approach improves on the image quality and embedding capacity achieved by Shie et al. Unlike Shie et al.’s scheme, in the proposed scheme, secret data are embedded into image blocks that are not necessarily sufficiently smooth. PSMVQ edge detection [2] is adopted to mask the edge direction of a block, and then each block is classified accordingly. The image block identified by PSMVQ edge detection can be referred to be as the “edge block.” Accordingly, four codebooks (denoted as CEW, CNS, CNW, and CES, respectively) are trained for those edge blocks. As in the scheme of Shie et al., a smooth codebook needs to be generated to encode the smooth blocks. Notably, instead of embedding secret data into the sufficiently smooth blocks, the proposed scheme selects edge blocks and non-sufficiently smooth blocks to hide secret data. Additionally, edge-directed prediction, also called EDP prediction for short, is adopted to improve image quality. Like SMVQ prediction, the EDP method exploits information about neighboring pixels to generate a set of state codebooks whose codewords contribute to the accuracy of the estimation of the predicted block.

This paper is organized as follows. Section 2 reviews several related works. Section 3 elucidates the proposed scheme. Section 4 presents the experimental results, and Section 5 draws conclusions.

Section snippets

Literature reviews

This section reviews the methods of SMVQ compression and PSMVQ edge detection that were proposed by Chang et al. [2] and the secret embedding based on the SMVQ [4], [8], [10], [18] compression method that was proposed by Shie et al.

The proposed scheme

In the scheme of Shie et al., the secret data are embedded into smooth blocks. However, the result is not natural to the human eye, which is quite sensitive to changes in sufficiently smooth blocks. Therefore, the modification of some pixels to embed secret data may attract the attention of interceptors. Additionally, most images have more smooth blocks than non-smooth blocks. Embedding secret data into smooth blocks yields PSNR values of stego-images of less than 30 dB, on average [16]. Instead

Experimental results

In the experiment, five gray images Lena, F16, Baboon, Pepper, and Barbara are tested. Each test image has a size of 512 × 512 pixels. Test images are divided into 16,384 non-overlapping blocks of 4 × 4 pixels for VQ and SMVQ encoding. Five codebooks of size 512 are generated using the LBG algorithm [11]. The image quality is evaluated using the peak signal-to-noise ratio (PSNR) and the embedded secret data are pure binary messages, which do not include extra indicator bits. To increase the data

Conclusions

Shie et al. [16] embed the secret data into the smooth image blocks. However, human visual system is more sensitive to the distortion while destroying the smooth image blocks. So, for avoiding the awareness of carrying secret messages, it is unlikely to embed much more secret bits into a cover image. Rather than embed secret into the smooth image block, the proposed scheme embeds the secret data only into the non-sufficiently smooth image blocks as well as the complicated image blocks. That is

References (19)

  • M.N. Wu et al.

    An embedding technique based upon block prediction

    Journal of Systems and Software

    (2008)
  • M. Barni et al.

    A DCT-domain system for robust image watermarking

    Signal Processing

    (1998)
  • C.C. Chang et al.

    Pattern-based side match vector quantization for image compression

    The Imaging Science Journal

    (2000)
  • C.C. Chang, Y.H. Chen, K.M. Li, Improved SMVQ scheme with edge-directed prediction, in: Proceedings of 2006...
  • C.C. Chang et al.

    A steganographic method for hiding secret data using side match vector quantization

    IEICE Transactions on Information and Systems

    (2005)
  • W.C. Du et al.

    Adaptive data hiding based on VQ compressed image

    IEE Proceedings-Vision, Image Signal Processing

    (2003)
  • R.M. Gray

    Vector quantization

    IEEE ASSP Magazine

    (1984)
  • H. Kii et al.

    The digital watermarking method by using both patchwork and DCT

    IEEE Multimedia Computing Systems

    (1999)
  • T. Kim

    Side match and overlap match vector quantizers for images

    IEEE Transactions on Image Processing

    (1992)
There are more references available in the full text version of this article.

Cited by (0)

View full text