CN-115829819-B - Image robust reversible information hiding method, device and medium based on neural network

CN115829819BCN 115829819 BCN115829819 BCN 115829819BCN-115829819-B

Abstract

The invention provides an image robust reversible information hiding method, device and medium based on a neural network, wherein the image robust reversible information hiding method of the neural network comprises two stages, namely, embedding a robust watermark sequence by using a convolutional neural network in the first stage, embedding auxiliary information for recovering an original carrier image by using the reversible neural network in the second stage, and completely recovering the original carrier image when no attack is received, wherein the watermark sequence can be correctly extracted, and the watermark sequence can still be correctly extracted due to the strong robustness of the first stage when the attack is received. The method of the invention can embed the watermark into the original carrier image to obtain a robust watermark image, and then embed auxiliary information for recovering the original carrier image to finally obtain the watermark image.

Inventors

QIAN YANG
LIU JIANRONG
ZHOU JIE

Assignees

江苏水印科技有限公司

Dates

Publication Date: 20260512
Application Date: 20221215

Claims (8)

1. The image robust reversible information hiding method based on neural network includes embedding watermark and extracting watermark, The embedding of the watermark comprises the following steps: S11, obtaining a characteristic image from an original carrier image through a CNN convolution block; S12, expanding the one-dimensional robust watermark sequence to obtain an expanded watermark sequence Is convenient for comprehensive embedding, and the characteristic image obtained in the step S11 is combined with the extended watermark sequence and the original carrier image Sequentially superposing, obtaining a coded image through a convolution block, and adding noise The sound layer is trained through a robust reversible model to resist various noises, so that a robust watermark sequence is extracted correctly; s13, subtracting the original carrier image from the coded image, carrying out Huffman lossless compression to reduce zero pixel values in the difference image, and embedding the difference image into the coded image by using a reversible neural network model of forward mapping between the obtained difference image and the coded image; s14, taking the forward output of the carrier image branch of the forward mapping reversible neural network model as a final embedded watermark image, and defining the forward output of the difference image branch of the forward mapping reversible neural network model as a constant matrix which does not contain effective information; the watermark extraction comprises the following steps: s21, the final embedded watermark image and the constant matrix are subjected to a reversible neural network model of reverse mapping; s22, utilizing reverse mapping of the reversible neural network, setting reverse output of a carrier image branch of the reverse-mapped reversible neural network model as a restored encoded image, setting reverse output of a difference value image branch of the reverse-mapped reversible neural network model as output information, and adding the restored encoded image and the output information to obtain a final restored carrier image; S23, obtaining a recovered watermark sequence from the recovered coded image through a CNN convolution block and the robust reversible model obtained through training ; S24 recovering the watermark sequence as obtained in step S23 The recovery coding image is subjected to malicious attack or information tampering, and the recovery watermark sequence is obtained again through a convolutional neural network, a pooling layer and a linear layer 。
2. The method for image robust and reversible information hiding based on neural network as claimed in claim 1, wherein said model in step S12 includes Encoder with a plurality of sensors , For training parameters, the encoder Receiving an original carrier image And watermark sequence The original carrier image has the size of CHW and the watermark sequence of binary character string Watermark sequence length is The encoder outputs an encoded image of size CHW ; Noise floor The noise layer Inputting coded images Noise attack is carried out on the coded image, and a noise image is output ; Decoder , For training parameters, the decoder From noisy images Extracting watermark sequence from a watermark For a pair of And Random gradient descent was performed.
3. The method for robust and reversible information hiding of image based on neural network as claimed in claim 2, wherein said encoded image in step S12 is indistinguishable by human eyes With the original carrier image By image distortion loss Representing the original carrier image And encoding an image The similarity between the two images is reduced, so that the size of the difference image is reduced, specifically: Wherein, the As an image of the original carrier body, In order to encode an image the image is encoded, Representing L2 norms, i.e. Euclidean distance, as the sum of squares of the individual elements of the x vector To the power, CHW is the size of the original carrier image.
4. The method for hiding robust and reversible information of image based on neural network as claimed in claim 2, wherein watermark sequence extracted from decoder of convolutional neural network in step S24 And the extended watermark sequence in the step S12 Similarity between them, by distortion loss of watermark information The specific method is as follows: Wherein, the In order to spread the watermark sequence, In order to recover the watermark sequence, Representing L2 norms, i.e. Euclidean distance, as the sum of squares of the individual elements of the x vector To the power of the method, For the length of the watermark sequence.
5. The method for hiding robust and reversible information of image based on neural network as claimed in claim 2, wherein the step S12 is performed by training through robust and reversible model by reducing sum of image distortion loss and watermark sequence distortion loss The method for gradient descent comprises the following steps: Wherein, the Representing the loss of distortion of the image, Representing the distortion loss of the watermark sequence, Representing the relative weights controlling the distortion loss of the image, To control the relative weight of the watermark sequence distortion loss.
6. The method for hiding robust reversible information of image based on neural network as claimed in claim 2, wherein said reversible neural network model in step S13 and said step S22 is formed by cascade connection of multiple reversible modules including neural network, and bidirectional mapping of said model shares all parameters, so that original carrier image can be recovered regardless of whether watermark image is attacked or not, and watermark sequence can be extracted correctly.
7. An electronic device, comprising: one or more processors; A memory for storing one or more programs; the instructions that when executed by the one or more processors cause the one or more processors to implement the method of any of claims 1-6.
8. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-6.

Description

Image robust reversible information hiding method, device and medium based on neural network Technical Field The invention belongs to the technical field of image analysis, and particularly relates to an image robust reversible information hiding method based on a neural network. Background In the fields of satellite images, medical images, miniature images and the like with high requirements on high fidelity and robustness of the images, the irreversible loss of the robust watermark and the vulnerability of the reversible watermark are intolerable, so that researchers develop a brand new watermarking algorithm, namely the robust reversible watermarking algorithm. The robust reversible watermarking algorithm can extract the watermark and recover the carrier image at the decoding end without attack. However, when the watermarked image is subjected to a malicious attack, although the watermark can be extracted correctly, the reversibility is destroyed, so that the carrier image cannot be recovered. De Vleeschouwer et al propose an early robust reversible watermarking algorithm, which is based on a histogram rotation technique of double mapping transformation, wherein each selected embedded block is randomly divided into two sets with the same number of pixels, pixel values are respectively mapped onto two circles, the circle centers and the centroids in the two circles are respectively connected to form two vectors, and information is embedded between the vectors through angles (technical document 1). Subsequently, ni et al improved the robust reversible watermarking algorithm proposed by De Vleeschouwer et al to avoid salt and pepper noise caused by rotation of the histogram, but this improvement inevitably caused erroneous bits of the overflow/underflow block (i.e., the block whose pixel value is less than 0 or greater than 255) at the decoding end, so Ni et al adopted error correction coding, which had the disadvantage that the embedding capacity of the method was significantly reduced (technical document 2). Coltuc et al propose a lossless robust watermarking algorithm based on a two-stage watermarking scheme (prior art document 3), wang et al improve the two-stage watermarking framework of Coltuc, divide the original carrier image into two independent embedded domains in the integer haar wavelet transform domain, wherein the watermark sequence is embedded in the low frequency coefficient region, the difference value of the carrier image and the robust watermark image is embedded in the high frequency coefficient region, but the robust feature utilized by the classical robust reversible watermarking algorithm is related to the position of the pixel, resulting in decoding failure for geometrical attack. Jiren Zhu et al applied neural networks for the first time to the field of digital image watermarking, but this method failed to reversibly recover the original carrier image (technical document 4). Therefore, the existing robust reversible watermarking algorithm has limitation in terms of robustness, and the image robust reversible information hiding method based on the neural network is provided by considering that the recovery of an original carrier image is completed by using the reversible neural network. [ Prior Art literature ] [ Technical document 1] De Vleeschouwer C, Macq B. Circular interpretation of bijective transformations in lossless watermarking for media asset management[J]. Multimedia, IEEE Transactions on, 2003, 5(1): 97-105. [ Technical document 2] Ni Z, Shi Y Q, Ansari N, et al. Robust lossless image data hiding designed for semi-fragile image authentication[J]. Circuits and Systems for Video Technology, IEEE Transactions on, 2008, 18(4): 497-509. [ Technical document 3] Coltuc D, Chassery J M. Distortion-free robust watermarking: a case study[C]//Electronic Imaging 2007. International Society for Optics and Photonics, 2007: 65051N-65051N-8. [ Technical document 4] Jiren Zhu, Russell Kaplan, Justin Johnson, Li Fei-Fei; Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 657-672. Disclosure of Invention The invention aims to embed a watermark into an original carrier image to obtain a robust watermark image, and then embed auxiliary information for recovering the original carrier image to finally obtain the watermark image. The method comprises two stages, wherein the first stage is to embed a robust watermark sequence by using a convolutional neural network, and the second stage is to embed auxiliary information for recovering an original carrier image by using a reversible neural network. The original carrier image can be completely restored without any attack, and the watermark sequence can be extracted correctly. When attacked, the watermark sequence can still be extracted correctly due to the strong robustness of the first stage. Other features and advantages of embodiments of the present disclosure will be apparent from the following detailed description, or