An Improved Image Camouflage Technique Using Color Difference Channel Transformation and Optimal Prediction-Error Expansion

• Main Authors: Heng Yao, Xiaokai Liu, Zhenjun Tang, Yu-Chen Hu, Chuan Qin
• Format: Article
• Language: English
• Published: IEEE 2018-01-01
• Series: IEEE Access
• Subjects: Image camouflage; reversible image transformation; color difference channel; image encryption; reversible data hiding
• Online Access: https://ieeexplore.ieee.org/document/8418376/

With the development of the ability to crack encrypted images, conventional image encryption techniques are no longer safe enough. Disguising a to-be-encrypted image into another visually-different image that is similar to the prepared target image is a solution that can be used to transmit an image securely. Inspired by the existing reversible image transformation technique, in this paper, we propose an improved method for camouflaging images that further decrease the distortion between the final camouflaged image and the target image. First, the G channel of the secret image is transformed and further refined to create a tentative camouflaged G channel with the reference of the G channel of the target image. Then, the color difference channels of the secret image are transformed with the reference of subtraction of the R channel (or B channel) and the tentatively camouflaged G channel. After shifting the color difference channels back to the R channel (or B channel), the sub-blocks are refined by a 16-candidate-pattern optimization strategy to generate the tentative camouflaged R channel (or B channel). After the combination of the RGB channels to generate the tentative camouflaged image, the final camouflaged version is generated by embedding all of the auxiliary information, which is collected for lossless recovery of the secret image, into the tentative image in a reversible manner. The experimental results demonstrated the efficacy of the proposed method, and our average gain in the color image peak signal-to-noise ratio (CPSNR) was more than 0.35 dB, whereas the state-of-the-art method is around 32.28 dB.