Convolutional neural networks-based stereo image reversible data hiding method

For copyright and integrity protection of stereo images, a stereo image reversible data hiding (SIRDH) method based on the convolutional neural network (CNN) is presented. To increase the pixel prediction, a CNN-based predictor is designed. Firstly, pixels are divided into two types. For predicting one type of pixels in one view, the other type of pixels in the same view is considered as the low-resolution (LR) image. Secondly, a LR difference view between the LR image and the other view is computed, since the difference view consists of the texture and depth information. Then, CNN is trained to recover the high-precision difference view from the LR difference view, and the high-resolution image is computed to predict pixels accurately. Moreover, the prediction error expansion (PEE) is employed to embed data. Experimental results show that the proposed method is superior to existing SIRDH methods. (C) 2019 Elsevier Inc. All rights reserved.