CN115049541B - Reversible gray scale method, system and device based on neural network and image steganography - Google Patents

Abstract

The invention discloses a reversible gray scale method, a system and a device based on neural network and image steganography, which comprise the steps of carrying out reversible conversion on an original color image to obtain a gray scale component Y and color components U and V; performing neural network coding and arithmetic coding on the color components to obtain a characteristic code stream and a super prior code stream; according to the image steganography, the characteristic code stream and the super prior code stream are steganographically written into the gray component Y, and a reversible gray image A is generated; reading a characteristic code stream and a super prior code stream in the reversible gray scale image A, and taking the read gray scale image A as a gray scale component Y _R of the color image to be reconstructed; performing neural network decoding and arithmetic decoding on the characteristic code stream and the super prior code stream to convert the characteristic code stream and the super prior code stream into color components U _R and V _R of a color image to be reconstructed; and combining the gray component of the color image to be reconstructed and the color component of the color image to be reconstructed, and performing reversible conversion to obtain a reconstructed color image I _R. The invention can be realized in the reversible gray scale of the neural network and the image steganography.

Description

Reversible grayscale method, system and device based on neural network and image steganography

Technical Field

The present invention relates to the field of reversible grayscale, and in particular to a reversible grayscale method, system and device based on neural network and image steganography.

Background technique

The method of generating grayscale images from color images has important applications in many fields, such as printing, engraving, monochrome display, image processing and other scenarios. Among them, the conventional type of grayscale image generation method focuses on perceptual factors, such as contrast, texture features and so on. There is also a type of grayscale generation method called reversible grayscale, whose main purpose is to generate a grayscale image while encoding the color information of the color image in the generated grayscale, and to restore the original color image as perfectly as possible when needed.

In 2018, Xia et al. proposed a reversible grayscale method in ACM Transactions on Graphics, which modeled the image decolorization and colorization process as a closed loop through an encoding-decoding network. This method can embed color information into the generated grayscale image, so that the decoded image can accurately reconstruct the color consistent with the original.

In 2020, Ye et al. proposed a dual feature set network in IEEE Access, which used dense residual representation, integrated local residual learning and local feature fusion capabilities, and suppressed the redundant characteristics generated by the dual-path module through the attention mechanism, thereby obtaining more consistent grayscale images and reconstructed color images.

In 2021, Liu et al. proposed a JPEG robust reversible grayscale system in IEEE Transactions on Visualization and Computer Graphics. This method introduced adversarial training and JPEG simulator based on the codec network, making the generated grayscale image JPEG robust and reducing the coding texture of the generated image. Zhao et al. proposed a new reversible grayscale method in IEEE Transactions on Image Processing. This method uses a reversible neural network to forward map a color image into a grayscale image and latent variables, and then uses reverse mapping to convert the grayscale image and a set of random variables that conform to the Gaussian distribution into a color image that is close to the original.

For the reversible grayscale method, the most critical performance is the similarity between the generated grayscale image and the reconstructed color image and the target grayscale image and the original color image respectively. The existing technology adopts an end-to-end structural framework, which is not perfect in the above performance and has obvious room for improvement. This is mainly attributed to the two main deficiencies in these technologies: ① Failure to truly and efficiently eliminate the redundancy of color information, resulting in a large amount of information to be encoded. ② In the process of encoding color information, the grayscale image loses a lot of information.

Summary of the invention

The purpose of the present invention is to provide a reversible grayscale method, system and device based on neural network and image steganography, aiming to solve the problem of reversible grayscale of images.

The present invention provides a reversible grayscale method based on neural network and image steganography, comprising:

S1, performing a reversible RGB2YUV conversion on the original color image to obtain the grayscale component Y and the color components U and V;

S2, performing neural network coding and arithmetic coding on the color components U and V to obtain a feature code stream and a super priori code stream;

S3, according to the image steganography, the characteristic code stream and the super prior code stream are steganographically written into the grayscale component Y to generate a reversible grayscale image A;

S4, reading the characteristic code stream and the super-prior code stream in the reversible grayscale image A, and using the read grayscale image A as the grayscale component Y _R of the color image to be reconstructed;

S5, performing neural network decoding and arithmetic decoding on the feature code stream and the super priori code stream to convert them into color components _UR and _VR of the color image to be reconstructed;

S6. Combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed, and performing a reversible YUV2RGB conversion to obtain a reconstructed color image _IR .

The present invention also provides a reversible grayscale system based on neural network and image steganography, comprising:

Conversion module: used to perform reversible RGB2YUV conversion on the original color image to obtain grayscale component Y and color components U and V;

Coding module: used to perform neural network coding and arithmetic coding on color components U and V to obtain feature code stream and super priori code stream;

Steganography module: used to steganographically write the characteristic code stream and the super prior code stream into the grayscale component Y according to the image steganography to generate a reversible grayscale image A;

Reading module: used to read the characteristic code stream and the super-prior code stream in the reversible grayscale image A, and use the read grayscale image A as the grayscale component Y _R of the color image to be reconstructed;

Decoding module: used to convert the feature code stream and the super-prior code stream into the color components _UR and _VR of the color image to be reconstructed by performing neural network decoding and arithmetic decoding;

Reconstruction module: used for combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing reversible YUV2RGB conversion to obtain a reconstructed color image _IR .

An embodiment of the present invention also provides a reversible grayscale device based on neural network and image steganography, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the computer program is executed by the processor.

An embodiment of the present invention further provides a computer-readable storage medium, on which a program for implementing information transmission is stored, and when the program is executed by a processor, the steps of the above method are implemented.

In the first aspect, the present invention designs a neural network to encode the color components of a color image, extracts the key features of the color components and models their probability models, and then encodes the features into a binary code stream through arithmetic coding. Compared with the prior art, such a strategy can more efficiently eliminate the redundancy of color information and reduce the loss of color information, solving the problem that too much information needs to be hidden in the intermediate process, resulting in poor quality of the generated grayscale image.

Secondly, the present invention writes (or reads) color information into the grayscale component through image steganography technology, embeds the color information into the grayscale component with minimal changes, and thus generates a target grayscale image. Compared with the prior art, the grayscale image generated by the present invention has a more ideal visual effect and less grayscale information loss.

Thirdly, the present invention utilizes a reversible component conversion method to decompose a color image into grayscale components and color components for orthogonal processing, and then combines neural networks and image steganography to significantly improve the comprehensive performance indicators of both generating grayscale images and reconstructing color images.

The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it is implemented in accordance with the contents of the specification, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a reversible grayscale method based on neural network and image steganography according to an embodiment of the present invention;

FIG2 is a schematic diagram of a framework of a reversible grayscale method based on a neural network and image steganography according to an embodiment of the present invention;

FIG3 is a schematic diagram of a neural network of a reversible grayscale method based on a neural network and image steganography according to an embodiment of the present invention;

FIG4 is a schematic diagram of modifying pixels in a reversible grayscale method based on a neural network and image steganography according to an embodiment of the present invention;

FIG5 is a schematic diagram of a reversible grayscale system based on neural network and image steganography according to an embodiment of the present invention;

FIG6 is a schematic diagram of a reversible grayscale device based on neural network and image steganography according to an embodiment of the present invention.

Detailed ways

The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Method Embodiment

According to an embodiment of the present invention, a reversible grayscale method based on neural network and image steganography is provided. FIG1 is a flow chart of a reversible grayscale method based on neural network and image steganography according to an embodiment of the present invention. As shown in FIG1 , the method specifically includes:

S1, performing a reversible RGB2YUV conversion on the original color image to obtain the grayscale component Y and the color components U and V;

S2, performing neural network coding and arithmetic coding on the color components U and V to obtain a feature code stream and a super priori code stream;

S3, according to image steganography, the characteristic code stream and the super prior code stream are steganographically written into the grayscale component Y to generate a reversible grayscale image A;

S4, reading the characteristic code stream and the super-prior code stream in the reversible grayscale image A, and using the read grayscale image A as the grayscale component Y _R of the color image to be reconstructed;

S5, performing neural network decoding and arithmetic decoding on the feature code stream and the super priori code stream to convert them into color components _UR and _VR of the color image to be reconstructed;

S6. Combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed, and performing a reversible YUV2RGB conversion to obtain a reconstructed color image _IR .

FIG. 2 is a schematic diagram of a framework of a reversible grayscale method based on a neural network and image steganography according to an embodiment of the present invention. As shown in FIG. 2 ,

The process of generating grayscale images from color images:

Step a: Perform a reversible RGB2YUV conversion on the original color image I and decompose the image to obtain the grayscale component Y and color components U, V;

Step b: The color components U and V are analyzed and encoded through a neural network to obtain features and super prior/> And create features/> The prior probability model of and super prior/> Independent probability model of/> Then, the features are combined with arithmetic coding Super Prior/> Convert to feature code streams respectively/> Super a priori code stream/>

Step c: Use image steganography to convert the feature code stream and super a priori code stream/> Steganographically write into the grayscale component Y to generate a reversible grayscale image A;

Grayscale image reconstruction color image process:

Step d: Read the feature code stream from the grayscale image A and super a priori code stream/> The read grayscale image A is used as the grayscale component Y _R of the color image to be reconstructed;

Step e: According to the independent probability model The super-a priori code stream/> Arithmetic decoding as a super prior /> Then the features are obtained by neural network decoding/> The probability model parameters, and then the prior probability model is obtained/> And the feature code stream/> Arithmetic decoding as a feature/> Then the neural network analyzes the features Perform synthesis to obtain color components _UR , _VR of the color image to be reconstructed;

Step f: combine the grayscale component Y _R and the color components _UR and _VR to perform reversible YUV2RGB conversion to obtain a reconstructed color image _IR ;

Preferably, the reversible RGB2YUV conversion formula of step a is:

Among them, Y represents the gray component value, U and V represent the color component values; R, G, and B represent the pixel values of the original color image; Indicates rounding down.

Preferably, the content of the neural network is implemented as follows:

FIG3 is a schematic diagram of a neural network of a reversible grayscale method based on a neural network and image steganography according to an embodiment of the present invention. As shown in FIG3 ,

The neural network consists of four parts: feature analysis network, feature synthesis network, super prior encoding network, and super prior decoding network.

The feature analysis network extracts the main feature x of the color component, rounds it to an integer and quantizes it into features Features/> When the information entropy is as small as possible, it is beneficial for the feature synthesis network to effectively reconstruct the color components; the feature analysis network is composed of a convolutional network layer and a GDN (generalized divisive normalization) nonlinear layer; assuming that the input color component dimension is H×W×2, the output feature dimension of the feature analysis network is/>

The super-prior coding network further encodes and quantizes the feature x and calculates the super-prior with the smallest possible information entropy. And input it into the super prior decoding network to accurately model the probability model of feature x; the super prior encoding network is composed of a convolutional network layer and a RELU nonlinear activation layer, and the corresponding output super prior dimension is/>

The super prior decoding network realizes the super prior Decode, the decoded variable is the normal distribution probability model parameter of feature x/> Further construct the feature/> The prior probability model of The super priori decoding network consists of a transposed convolutional network layer and a RELU nonlinear activation layer, and its output probability model parameters are /> Among them, the prior probability model/> The form is:

The feature synthesis network is based on the input features Perform synthesis to reconstruct color components that are as close to the original as possible; the feature synthesis network consists of a transposed convolutional network layer and an IGDN (inversed generalized divisive normalization) nonlinear layer, and the dimension of the reconstructed component output is H×W×2;

Preferably, the image steganography method is implemented as follows:

The pixels of the carrier grayscale image are scanned by columns and represented as P ₁ , P ₂ , P ₃ ,…, P _m ; the feature code stream and super a priori code stream Merge into a binary code stream O and represent it as b ₁ , b ₂ , b ₃ , … , b _n ; sequentially take every 3 pixels and every 3-bit binary code as a group, the 3 pixels of the i-th group are represented as P _{i_1} , P _{i_2} , P _{i_3} , and the 3-bit binary code is represented as b _{i_1} , b _{i_2} , b _{i_3} ;

In step c, the steganographic process embeds 3 bits of binary code information into each group of 3 pixels, which is specifically implemented as follows:

First, a 3-bit prediction code is calculated based on the values of the 3 pixels:

Where _{Bi_1} , _{Bi_2} , _{Bi_3} represent the 3-bit prediction code of the i-th group, Indicates the lowest j-th bit of the binary value of the n-th pixel _{pi_n} of the i-th group,/> Represents XOR operation;

If the predicted code _{Bi_1} , _{Bi_2} , _{Bi_3} is equal to the binary code _{Bi_1} , _{Bi_2} , _{Bi_3} , then there is no need to modify the pixel; otherwise, the predicted code and the binary code are made equal by modifying the pixel _{Pi_1} , _{Pi_2} , _{Pi_3} , thus completing the steganography of the binary code;

FIG4 is a schematic diagram of pixel modification of a reversible grayscale method based on a neural network and image steganography according to an embodiment of the present invention. As shown in FIG4 ,

_{Bi_1} ≠ _{bi_1} , _{Bi_2} = _{bi_2} , _{Bi_3} = _{bi_3} ; if _{Pi_2} mod 2 = 0, then _{Pi_2} = _{Pi_2} -1; if _{Pi_2} mod 2 = 0, then _{Pi_2} = _{Pi_2} +1;

In addition, if the last number of bits does not meet 3, the lowest binary bit of the corresponding sequence of pixels is directly replaced to achieve steganography.

In step d, the reading process is:

According to the above prediction method, the prediction codes of every 3 pixels _{Pi_1} , _{Pi_2} , _{Pi_3} from _P1 to _Pm are calculated in sequence. For pixels that do not meet the 3-bit number, the lowest binary bit of the pixel is directly read to replace the corresponding prediction code, thereby reading the binary code stream O hidden in the grayscale image. After decomposition, the feature code stream can be obtained. and super a priori code stream/>

Preferably, the reversible YUV2RGB conversion formula in step f above is:

Among them, R, G, B represent the pixel values of the color image; Y represents the gray component value, and U, V represent the color component value; Indicates rounding down.

Specifically, this embodiment is implemented by Python, the neural network is constructed by using the Pytorch deep learning framework and optimized and trained by the Adam optimizer; the training set is 20,000 images randomly selected from the Pascal VOC2012 public data set and the cropping resolution is 512×512, and the test set is the Kodak Photo CD image data set; the initial learning rate of the training is set to 1× ^10-4 , and the learning rate decays to 1× ^10-5 after 2× ¹⁰⁶ iterations, and continues to iterate 5× ¹⁰⁵ times; the training process constrains the optimization direction of the neural network through the loss function, comprehensively reducing the conversion loss of the color component and the feature Super Prior/> The loss function used in training is:

The first term of the equation is the mean square error between the color components U, V before conversion and the color components _UR , _VR after conversion; the second and third terms are the feature Super Prior/> Information entropy of

In summary, this embodiment ultimately achieves: the original color image generates a corresponding reversible grayscale image, and then the grayscale image can be used to reconstruct a color image that is substantially consistent with the original one.

The technical solution provided by the present invention has at least the following advantages compared with the prior art:

In the first aspect, the present invention designs a neural network to encode the color components of a color image, extracts the key features of the color components and models their probability models, and then encodes the features into a binary code stream through arithmetic coding. Compared with the prior art, such a strategy can more efficiently eliminate the redundancy of color information and reduce the loss of color information, solving the problem that too much information needs to be hidden in the intermediate process, resulting in poor quality of the generated grayscale image.

Secondly, the present invention writes (or reads) color information into the grayscale component through image steganography technology, embeds the color information into the grayscale component with minimal changes, and thus generates a target grayscale image. Compared with the prior art, the grayscale image generated by the present invention has a more ideal visual effect and less grayscale information loss.

Thirdly, the present invention utilizes a reversible component conversion method to decompose a color image into grayscale components and color components for orthogonal processing, and then combines neural networks and image steganography to significantly improve the comprehensive performance indicators of both generating grayscale images and reconstructing color images.

System Example

According to an embodiment of the present invention, a reversible grayscale system based on a neural network and image steganography is provided. FIG5 is a schematic diagram of a reversible grayscale system based on a neural network and image steganography according to an embodiment of the present invention. As shown in FIG5, the system specifically includes:

Conversion module: used to perform reversible RGB2YUV conversion on the original color image to obtain grayscale component Y and color components U and V;

Coding module: used to perform neural network coding and arithmetic coding on color components U and V to obtain feature code stream and super priori code stream;

Steganography module: used to steganographically write the characteristic code stream and the super prior code stream into the grayscale component Y according to the image steganography to generate a reversible grayscale image A;

Reading module: used to read the characteristic code stream and the super-prior code stream in the reversible grayscale image A, and use the read grayscale image A as the grayscale component Y _R of the color image to be reconstructed;

Decoding module: used to convert the feature code stream and the super-prior code stream into the color components _UR and _VR of the color image to be reconstructed by performing neural network decoding and arithmetic decoding;

Reconstruction module: used for combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing reversible YUV2RGB conversion to obtain a reconstructed color image _IR .

The conversion module is specifically used for:

The original color image is reversibly converted to RGB2YUV to obtain the grayscale component Y and the color components U and V. The reversible RGB2YUV conversion formula is as follows:

Among them, Y represents the gray component value, U and V represent the color component values; R, G, B represent the pixel values of the original color image; Indicates rounding down;

The encoding module is specifically used for:

The color components U, V are analyzed and encoded through neural networks to obtain features and super priors, and the feature prior probability model and super prior independent probability model are established. Then, the features and super priors are converted into feature code streams and super prior code streams respectively by combining arithmetic coding;

The steganography module is specifically used for:

The pixels of the gray component Y are scanned by columns and represented as P ₁ , P ₂ , P ₃ , ..., P _m . The characteristic code stream and the super-prior code stream are combined into a binary code stream and represented as b ₁ , b ₂ , b ₃ , ..., b _n .

The 3-bit prediction code is predicted based on the 3 pixel values. The formula is as follows:

Where _{Bi_1} , _{Bi_2} , _{Bi_3} represent the 3-bit prediction code of the i-th group, Indicates the lowest j-th bit of the binary value of the n-th pixel _{pi_n} of the i-th group,/> Represents the exclusive OR operation;

If the binary code _{bi_1} , _{bi_2} , _{bi_3} is equal to the predicted code _{Bi_1} , _{Bi_2} , _{Bi_3} , then there is no need to modify the pixel. Otherwise, modify the pixel _{Pi_1} , _{Pi_2} , _{Pi_3} until the binary code and the predicted code are equal, thereby completing the steganography of the binary code. If the number of code streams does not meet the 3-bit requirement, directly replace the lowest binary bit of the corresponding sequence pixel to achieve steganography, and complete the hiding of the 3-bit binary code _{bi_1} , _{bi_2} , _{bi_3} by the 3 pixels _{Pi_1} , _{Pi_2} , _{Pi_3} , and generate a reversible grayscale image A.

The decoding module is specifically used for:

According to the independent probability model, the super-prior code stream is arithmetically decoded into the super-prior, and then the probability model parameters of the feature are obtained by decoding through the neural network, so as to obtain the a priori probability model and arithmetically decode the feature code stream into features, and then the neural network synthesizes the features to obtain the color components _UR , _VR of the color image to be reconstructed;

The reconstruction module is specifically used to:

After combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing reversible YUV2RGB conversion, a reconstructed color image _IR is obtained.

The reversible YUV2RGB conversion formula is:

Among them, R, G, B represent the pixel values of the color image; Y represents the gray component value, and U, V represent the color component value; Indicates rounding down.

The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.

Device Example 1

An embodiment of the present invention provides a reversible grayscale device based on neural network and image steganography, as shown in Figure 5, including: a memory 50, a processor 52, and a computer program stored in the memory 50 and executable on the processor 52. When the computer program is executed by the processor, the steps in the above method embodiment are implemented.

Device Example 2

An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by the processor 52, the steps in the above method embodiment are implemented.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements of the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of this solution.

Claims (8)

1. A reversible grayscale method based on neural network and image steganography, characterized by comprising: S1, performing a reversible RGB2YUV conversion on the original color image to obtain the grayscale component Y and the color components U and V; S2, performing neural network coding and arithmetic coding on the color components U and V to obtain a feature code stream and a super priori code stream; S3, according to the image steganography, the characteristic code stream and the super prior code stream are steganographically written into the grayscale component Y to generate a reversible grayscale image A; S4, reading the characteristic code stream and the super-prior code stream in the reversible grayscale image A, and using the read grayscale image A as the grayscale component Y _R of the color image to be reconstructed; S5, performing neural network decoding and arithmetic decoding on the feature code stream and the super priori code stream to convert them into color components _UR and _VR of the color image to be reconstructed; S6, combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing a reversible YUV2RGB conversion to obtain a reconstructed color image _IR ; The S2 specifically includes: The color components U, V are analyzed and encoded through neural networks to obtain features and super priors, and the feature prior probability model and super prior independent probability model are established. Then, the features and super priors are converted into feature code streams and super prior code streams respectively by combining arithmetic coding; The S3 specifically includes: The pixels of the gray component Y are scanned by columns and represented as P ₁ , P ₂ , P ₃ , …, P _m . The characteristic code stream and the super-prior code stream are combined into a binary code stream and represented as b ₁ , b ₂ , b ₃ , …, b _n . Take every 3 pixels and every 3-bit binary code as a group in order, the 3 pixels in the i-th group are represented as _{Pi_1} , _{Pi_2} , _{Pi_3} , and the 3-bit binary code is represented as _{bi_1} , _{bi_2} , _{bi_3} ; In step c, the steganographic process embeds 3 bits of binary code information into each group of 3 pixels, which is specifically implemented as follows: First, a 3-bit prediction code is calculated based on the values of the 3 pixels: Where _{Bi_1} , _{Bi_2} , _{Bi_3} represent the 3-bit prediction code of the i-th group, Indicates the lowest j-th bit of the binary value of the n-th pixel _{pi_n} of the i-th group,/> Represents XOR operation; If the predicted code _{Bi_1} , _{Bi_2} , _{Bi_3} is equal to the binary code _{Bi_1} , _{Bi_2} , _{Bi_3} , then there is no need to modify the pixel; otherwise, the predicted code and the binary code are made equal by modifying the pixel _{Pi_1} , _{Pi_2} , _{Pi_3} , thus completing the steganography of the binary code; The code streams whose number does not meet 3 bits directly replace the lowest binary bit of the corresponding sequence pixels to achieve steganography, and complete the hiding of 3 pixels Pi ₊₁ , Pi ₊₂ , Pi ₊₃ into 3-bit binary codes bi ₊₁ , bi ₊₂ , bi ₊₃ to generate a reversible grayscale image A. 2. The method according to claim 1, characterized in that the S1 specifically comprises: The original color image is reversibly converted to RGB2YUV to obtain the grayscale component Y and the color components U and V. The reversible RGB2YUV conversion formula is as follows: Among them, Y represents the gray component value, U and V represent the color component values; R, G, B represent the pixel values of the original color image; Indicates rounding down. 3. The method according to claim 1, characterized in that the S5 specifically comprises: The super-prior code stream is arithmetically decoded into a super-prior according to an independent probability model, and then the probability model parameters of the feature are obtained by decoding through a neural network, thereby obtaining a priori probability model and arithmetically decoding the feature code stream into features, and then the neural network synthesizes the features to obtain the color components _UR , _VR of the color image to be reconstructed. 4. The method according to claim 3, characterized in that said S6 specifically comprises: After combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing reversible YUV2RGB conversion, a reconstructed color image _IR is obtained. The reversible YUV2RGB conversion formula is: Among them, R, G, B represent the pixel values of the color image; Y represents the gray component value, and U, V represent the color component value; Indicates rounding down. 5. A reversible grayscale system based on neural network and image steganography, characterized by comprising: Conversion module: used to perform reversible RGB2YUV conversion on the original color image to obtain grayscale component Y and color components U and V; Coding module: used to perform neural network coding and arithmetic coding on color components U and V to obtain feature code stream and super priori code stream; Steganography module: used to steganographically write the characteristic code stream and the super prior code stream into the grayscale component Y according to the image steganography to generate a reversible grayscale image A; Reading module: used to read the characteristic code stream and the super-prior code stream in the reversible grayscale image A, and use the read grayscale image A as the grayscale component Y _R of the color image to be reconstructed; Decoding module: used to convert the feature code stream and the super-prior code stream into the color components _UR and _VR of the color image to be reconstructed by performing neural network decoding and arithmetic decoding; Reconstruction module: used for combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing reversible YUV2RGB conversion to obtain a reconstructed color image _IR ; The encoding module is specifically used for: The color components U, V are analyzed and encoded through neural networks to obtain features and super priors, and the feature prior probability model and super prior independent probability model are established. Then, the features and super priors are converted into feature code streams and super prior code streams respectively by combining arithmetic coding; The steganography module is specifically used for: The pixels of the gray component Y are scanned by columns and represented as P ₁ , P ₂ , P ₃ , …, P _m . The characteristic code stream and the super-prior code stream are combined into a binary code stream and represented as b ₁ , b ₂ , b ₃ , …, b _n . Take every 3 pixels and every 3-bit binary code as a group in order, the 3 pixels in the i-th group are represented as _{Pi_1} , _{Pi_2} , _{Pi_3} , and the 3-bit binary code is represented as _{bi_1} , _{bi_2} , _{bi_3} ; In step c, the steganographic process embeds 3 bits of binary code information into each group of 3 pixels, which is specifically implemented as follows: First, a 3-bit prediction code is calculated based on the values of the 3 pixels: Where _{Bi_1} , _{Bi_2} , _{Bi_3} represent the 3-bit prediction code of the i-th group, Indicates the lowest j-th bit of the binary value of the n-th pixel _{pi_n} of the i-th group,/> Represents XOR operation; If the predicted code _{Bi_1} , _{Bi_2} , _{Bi_3} is equal to the binary code _{Bi_1} , _{Bi_2} , _{Bi_3} , then there is no need to modify the pixel; otherwise, the predicted code and the binary code are made equal by modifying the pixel _{Pi_1} , _{Pi_2} , _{Pi_3} , thus completing the steganography of the binary code; The code streams whose number does not meet 3 bits directly replace the lowest binary bit of the corresponding sequence pixels to achieve steganography, and complete the hiding of 3 pixels Pi ₊₁ , Pi ₊₂ , Pi ₊₃ into 3-bit binary codes bi ₊₁ , bi ₊₂ , bi ₊₃ to generate a reversible grayscale image A. 6. The system according to claim 5, characterized in that the conversion module is specifically used for: The original color image is reversibly converted to RGB2YUV to obtain the grayscale component Y and the color components U and V. The reversible RGB2YUV conversion formula is as follows: Among them, Y represents the gray component value, U and V represent the color component values; R, G, B represent the pixel values of the original color image; Indicates rounding down; The decoding module is specifically used for: According to the independent probability model, the super-prior code stream is arithmetically decoded into the super-prior, and then the probability model parameters of the feature are obtained by decoding through the neural network, so as to obtain the a priori probability model and arithmetically decode the feature code stream into features, and then the neural network synthesizes the features to obtain the color components _UR , _VR of the color image to be reconstructed; The reconstruction module is specifically used to: After combining the grayscale component of the color image to be reconstructed and the color component of the color image to be reconstructed and performing reversible YUV2RGB conversion, a reconstructed color image _IR is obtained. The reversible YUV2RGB conversion formula is: Among them, R, G, B represent the pixel values of the color image; Y represents the gray component value, and U, V represent the color component value; Indicates rounding down. 7. A reversible grayscale device based on neural network and image steganography, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the reversible grayscale method based on neural network and image steganography as described in any one of claims 1 to 4 are implemented. 8. A computer-readable storage medium, characterized in that an implementation program for information transmission is stored on the computer-readable storage medium, and when the program is executed by a processor, the steps of the reversible grayscale method based on neural network and image steganography as described in any one of claims 1 to 4 are implemented.