EFFICIENT REVERSIBLE DATA HIDING BASED ON TWO-DIMENSIONAL PIXEL-INTENSITY-HISTOGRAM MODIFICATION Xinlu Gui, Xiaolong Li and Bin Yang Institute of Computer Science and Technology, Peking University, Beijing 100871, China ABSTRACT In this paper, referred to the general framework of histogram- shifting-based reversible data hiding, some valuable related works based on two-dimensional pixel-intensity-histogram modification are first introduced. In these schemes, all pixels are classified in- to two categories, which are embedded with data or shifted for reversibility, respectively. From this point of view, novel embed- ding schemes are proposed with more applicable pixel partition and more redundant information utilized. Moreover, pixel selection is conducted such that smooth pixels are priorly embedded to further exploit the image redundancy. The experimental results show the su- periority of the proposed methods over some state-of-the-art works in terms of capacity-distortion performance. Index Terms— Reversible data hiding (RDH), two-dimensional pixel-intensity-histogram, pixel selection (PS), capacity-distortion performance. 1. INTRODUCTION Reversible data hiding (RDH) is a technique to embed useful infor- mation into digital image in such a way that the intended recipient can extract the exact embedded data and recover the host medium as well [1]. This technique has been widely applied to some sensitive applications, such as law forensics, medical image processing and military image processing. In general, the performance of a RDH algorithm is evaluated in the following three aspects: embedding ca- pacity (EC), marked image quality and computational complexity. Specifically, for a given EC, one expects to minimize the embed- ding distortion (measured by PSNR in dB) and meanwhile keep the computational complexity as low as possible. A significant amount of research has been done on RDH over the past few years. Early RDH algorithms are mainly based on the lossless compression technique [2–5], in which certain features of host image are losslessly compressed to save space for embedding the payload. These methods usually provide low EC and may lead to severe degradation in image quality. Later on, more efficient al- gorithms based on difference expansion (DE) and histogram shifting (HS) have been devised. DE algorithm proposed by Tian [6] is preformed on pixel pairs, and 1 bit is embedded into each selected pixel pair by expanding its difference. This method can achieve a high EC with a low dis- tortion, in which the storage of the location map has a negative in- fluence. Later on, DE technique has been widely investigated and developed, mainly in the aspects of integer transform generaliza- tion [7–11], location map reduction [12–15] and prediction-error ex- pansion (PEE) [16–22] where the difference value is replaced by the prediction-error in expansion embedding. Corresponding author: Bin Yang, e-mail: yang bin@pku.edu.cn HS-based RDH is implemented by modifying host image’s his- togram of a certain dimension [23–26]. By HS, the maximum mod- ification to pixel values can be controlled and the location map used to record underflow/overflow locations is usually small in size. The HS-based method is firstly proposed by Ni et al. [23], in which the peak point of image histogram is utilized to embed data. In this method, each pixel value is modified by at most 1, and thus the marked image quality is well guaranteed with a PSNR larg- er than 48.13 dB. Afterwards, HS algorithm is improved by us- ing the histogram of difference image in Lee et al.’s method [25]. Compared with the ordinary one-dimensional histogram utilized in [23], the one-dimensional difference-histogram (equivalent to the two-dimensional pixel-intensity-histogram) in [25] is better for RD- H since it is regular in shape and has a much higher peak point. That is to say, more redundancy can be utilized by two-dimensional pixel- intensity-histogram, which is our focus in this paper. Conventional RDH schemes such as [16, 18] treat all image pix- els equally and sequentially embeds data into pixels one by one. However, it is widely recognized that embedding in noisy pixels may cause larger distortion than that in smooth ones with identi- cal embedded data size. To remedy this drawback, adaptive em- bedding based on the pixel selection (PS) strategy is proposed by Li et al. [21]. It guarantees that more data is embedded into a s- moother pixel according to a local complexity measurement. Based on this thought, for RDH based on two-dimensional pixel-intensity- histogram, we can embed data into the smooth pixel pairs with pri- ority over the noisy ones to obtain better marked image quality. In this paper, referred to the general framework of HS-based RDH proposed by Li et al. [27], we first introduce several related works. Then two novel RDH schemes based on two-dimensional pixel-intensity-histogram are proposed, which outperform some previous HS-based methods in the capacity-distortion performance. The first one adaptively embeds data into high frequency pixel pairs to improve the capacity without distortion increment. The other one guarantees that for a desired capacity, the payload is chosen to keep more pixel pairs unchanged after embedding. The proposed RDH schemes prove the superiority of designing efficient RDH based on high dimensional pixel-intensity-histogram and can be referred as valuable examples for further research. Moreover, PS strategy is adopted in our work such that smooth pixel pairs are priorly embedded to further exploit image redundancy. It should be admitted that though our methods outperform many previous works, some recent RDH schemes are better than ours. However, the idea of constructing better RDH schemes by designing more effective pixel pair mapping in the two-dimensional histogram is valuable as it is distinctive from existing ones and leaves much space for improvement. The rest of this paper is organized as follows. The general frame- work of HS-based RDH and several related works are introduced in Section 2. The proposed RDH schemes are described in detail in 2014 IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP) 978-1-4799-2893-4/14/$31.00 ©2014 IEEE 7470