Reversible Grayscale Method Based On Neural Network Coding And Image Steganograph

Color images are widely used in multimedia tasks at present,but compared with gray image,they are less cost-effective in terms of space occupancy and processing complexity.In some applications,a color image is converted to a gray image for transmission or presentation,but it is restored to a color version when needed in the future.To solve this problem,invertible grayscale method comes into being.The invertible grayscale method can convert the color image to gray image and restore the color image by combining the two processes of colorization and decolorization.However,the existing invertible grayscale method has some shortcomings in the mutual conversion of the two image forms,which leads to the existence of obvious coding texture in the generated grayscale result,or serious color offset in the reconstructed color image.Therefore,invertible grayscale method is studied in this paper.Its main purpose is to obtain high-quality invertible grayscale image and reconstruct color image.The main research contents include the following two points:(1)In response to the problems of poor visual quality of synthesized invertible grayscale images and low similarity between reconstructed color images and original color images in existing methods,this paper proposes an invertible grayscale method based on Variational Autoencoder(VAE)and Low Distortion Image Steganography(LDIS).The method first performs decorrelation processing on the original color components through color space conversion,decomposing them into luminance components and chrominance components.Then,the chrominance components are efficiently encoded by a variational autoencoder neural network.After that,the encoded color information is embedded into the luminance plane using a low distortion image steganography algorithm,achieving the grayscale image’s ability to carry color information and generating high-quality invertible grayscale images.The efficient image encoding effectively improves the problem of color information loss when reconstructing color images,while low-distortion image steganography makes the reconstructed color images of better quality.(2)In order to further improve the performance of the invertible grayscale method and achieve more reliable image results,this paper proposes an invertible grayscale method based on Invertible Neural Networks(INN)and Exploiting Modification Direction(EMD)algorithm.The new chrominance component encoder-decoder is implemented through an invertible neural network structure,which has stronger feature transformation ability and can reduce information loss in the color encoding and decoding process.In addition,dense convolution blocks and channel attention mechanisms are introduced to further improve the network’s performance.Furthermore,to enable the grayscale image to carry more encoding information and reduce image distortion caused by the embedding process,a modification direction-based image steganography algorithm is proposed in the method.This algorithm adaptively selects weight parameter values when embedding information into the carrier image,meeting different embedding capacities,and completes the loading of color information in a way that is close to optimal,reducing the overall modification of the luminance plane.The experimental results show that both invertible grayscale methods proposed in this paper can generate high-quality invertible grayscale images and reconstruct more realistic and reliable color images.In comparison with advanced invertible grayscale methods,they achieve better image similarity evaluation indices with reference images,while also having better visual effects.