Article Preview
TopIntroduction
As a technique that embeds messages into cover signals, information hiding has been widely applied in areas such as convert communication, copyright protection and media annotation. Reversible data hiding (RDH), as a special type of information hiding technique, has received much attention from the information community (Shi, Ni, Zou, Liang, & Xuan, 2004; Shi, 2004; Caldelli, Filippini, & Becarelli, 2010) in the last decade. Specifically, RDH ensures not only the embedded messages shall be extracted precisely, but also the cover itself should be restored losslessly. This property is important in some special scenarios such as medical imagery (Bao et al., 2005), military imagery and law forensics. In these applications, the cover is too precious or too important to be damaged (Feng et al., 2006). Moreover, it has been found recently that reversible data hiding can be quite helpful in video error-concealment coding (Chung et al., 2010).
A plenty of reversible data hiding algorithms have been proposed in the past decade. Classical RDH methods roughly fall into three categories. The first class of algorithms follows the idea of compression-embedding framework, which was first introduced by Fridrich, Goljan, and Du (2002). In these algorithms, a two-value feature is calculated for each pixel group, the sequence is compressible and messages can be embedded in the extra space left by lossless compression. The send class of techniques is based on difference expansion (DE) (Tian, 2003; Thodi & Rodriguez, 2007), in which the differences of each pixel groups are expanded, e.g., multiplied by 2, and thus the least significant bits (LSBs) of the differences are all-zeros and can be used for embedding messages. The last RDH schemes are based on histogram shift (HS) (Ni, Shi, Ansari, & Wei, 2006). The histogram of one special feature (for example, gray-scale value) of the nature image is quite uneven, which implies that the histogram can be modified for embedding data. For instance some space can be saved for watermarks by shifting the bins of histogram.
In fact, by applying DE of HS to the residual part of nature images instead, e.g., the prediction errors (PE) (Tsai, Hu, & Yeh, 2009; Luo et al., 2010; Peng, Li, & Yang, 2012; Li, Yang, & Zeng, 2011), better performance can be achieved. This extended method is called prediction-error expansion (PEE), which is currently a research hotspot and the most powerful technique of RDH. Unlike in DE where only the correlation of two adjacent pixels is considered, the local correlation of larger neighborhood is exploited in PEE. Most recently proposed RDH works are based on PEE by incorporating some strategies such as better prediction algorithm utilization (Sachnev, Kim, Nam, Suresh, & Shi, 2009; Fallahpour, 2008; Yang, Chung, Liao, & Yu, 2013; Ou, Li, Zhao, & Ni, 2013), double-layered embedding (Luo et al., 2010; Sachnev et al., 2009), embedding position selection (Li et al., 2011), context modification (Coltuc, 2011), optimal bins selection (Wang, Li, & Yang, 2010; Wu & Huang, 2012), etc.