CNN Based Local Complexity Estimation for Reversible Data Hiding
In the reversible data hiding(RDH)community,how to select smaller prediction er-rors to reduce the embedding distortion is a crucial issue.A logical way is to use the neighboring pixels to estimate the local complexity of a target pixel for selection.Previous local complexity calculation methods often use the correlation between neighboring pixels to calculate the local complexity,in which only a small number of neighboring pixels are considered for calculation.In addition,these calculation methods are only related to the embedding methods,which are inde-pendent of the prediction methods.Due to the small calculation range and the separation from the prediction methods,the correlation between prediction errors and local complexities is limited and should be further considered.In this paper,we estimate that the local complexity is the predic-tion of the prediction errors.To better explain the estimation,the classical calculation method named local absolute error(LAE)is adopted,which greatly reveals that the local complexity is highly correlated with the prediction error.Based on the discovery,we propose a new CNN-based method titled LCCN to better calculate the local complexity of pixels.To our knowledge,we are the first to calculate the local complexity by using the deep learning-based method.Compared with previous local complexity calculation methods,the proposed LCCN can use more neighbor-ing pixels as the context by exploiting the multi-receptive fields of CNN and the global optimiza-tion capacities of deep learning.As a bonus,satisfactory performance in selecting those smaller prediction errors for RDH can be achieved,e.g.,the proposed LCCN can achieve lower mean and variance of the sorted prediction errors than the classical local complexity calculation meth-ods,including the local variance(LV),forward variance(FV),error energy(EE),and LAE.In addition,the correlation coefficient between prediction error and local complexity is higher and the value range of the sorted prediction is more concentrated.Furthermore,we put forward a LCCN based two-stage embedding scheme,which is suitable for various embedding techniques.With several standard benchmark images and the Kodak dataset,experimental results have shown that the proposed LCCN can be applied for existing dominant embedding techniques to im-prove the embedding performance with higher PSNR values,e.g.,the PSNR of the image Lena in the LPVO-based embedding technique can reach as high as 62.09 dB after embedding 10000 bits,which is higher than 1.05 dB,1.02 dB,0.78 dB,and 0.90 dB compared to LV,FV,EE,and LAE algorithms,respectively.Compared with several state-of-the-art works in eight bench-mark images,the proposed method can achieve better results on average and in most cases.To further reveal the improvement of the proposed LCCN,the Kodak dataset is adopted to calculate the average PSNR with different embedding capacities,in which the proposed LCCN can achieve a higher average PSNR.For future works,we focus on improving the theory of local complexity and enhancing the similarity between local complexity and prediction error by using deep learn-ing.In addition,the consistency relationship between prediction techniques and local complexity calculation methods should be further investigated.
reversible data hidinglocal complexitydeep learningconvolution neural net-worksprediction errorprivate computing
NETL
NSTL
万方数据
