Generative image watermark information hiding method based on multi-channel neural network

Abstract

The invention belongs to the technical field of image processing and information security, and discloses a multi-channel neural network-based generative image watermark information hiding method, which comprises the following steps of: generating a target carrier image through a self-developed generative image model; performing code conversion on the original watermark information to generate a watermark code; and constructing a multi-channel neural network embedder based on a Swi n Transform backbone, and fusing the watermark code to the carrier image to generate a watermark-containing image. According to the method, the remote dependence modeling capability of the Swi n Transformer and a multi-channel feature fusion mechanism are combined, and a composite loss function (including absolute difference loss, mean square error loss, gradient loss and connectivity loss) and anti-attack training are utilized, so that high-concealment embedding and strong-robustness extraction of the watermark are realized; the resistance to image processing such as cutting, compression and filtering is remarkably improved, meanwhile, original image recovery after watermark verification is supported, and a light-weight and high-reliability copyright protection and traceability tracking solution is provided for e-commerce commodity images.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing and information security, and in particular relates to a generative image watermark information hiding method based on a multi-channel neural network. Background Art

[0002] With the rapid development of e-commerce, product images, as a key medium for product display, play an irreplaceable role in attracting user attention and increasing conversion rates. High-definition, sophisticated, and uniformly styled images have become a crucial tool for brand communication. However, high-quality product images carry the risk of unauthorized downloading, tampering, and misappropriation, seriously infringing merchants' intellectual property rights and posing significant challenges to platform compliance supervision and brand protection.

[0003] Traditional image watermarking technologies, such as plaintext watermarking, steganographic coding, and watermark layer superposition, have achieved information embedding to a certain extent, but they have the following shortcomings:

[0004] High visibility and disturbance: Traditional watermarks can easily affect the aesthetics of images and reduce the effectiveness of product display;

[0005] Easy to be removed or tampered with: Watermark information can be easily damaged or removed by simple image editing operations (such as cropping, compression, and filtering);

[0006] Lack of intelligent adaptability: When faced with complex situations such as diverse product styles, shooting backgrounds, and texture details in e-commerce images, traditional watermarking methods find it difficult to dynamically adapt to image content and achieve the unity of "seamless" embedding and stable extraction.

[0007] In recent years, generative image technologies such as generative adversarial networks (GANs) have made great breakthroughs in image synthesis and editing, providing a new approach for deep embedding of image watermarks. At the same time, multi-channel neural networks have performed well in processing complex data fusion and high-dimensional feature modeling, providing strong support for improving the embedding capacity, concealment, and robustness of watermarks. However, existing research mostly stays at the technical verification level, lacking a systematic design that is deeply integrated with practical e-commerce scenarios, especially a lack of lightweight and highly stable solutions for the high-fidelity requirements of product display images. Therefore, there is an urgent need for a generative watermark information hiding method based on a multi-channel neural network for e-commerce detail page images, which can achieve "imperceptible" encrypted information embedding and reversible extraction of product images, helping e-commerce platforms build product content copyright protection and tracking and traceability systems. Summary of the Invention

[0008] The purpose of the present invention is to provide a generative image watermark information hiding method based on a multi-channel neural network to solve the problems of strong visibility, easy tampering, lack of robustness and poor adaptability in existing image watermarking technology. It is particularly suitable for the copyright protection and information tracking needs of e-commerce product images.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a generative image watermark information hiding method based on a multi-channel neural network, comprising the following steps:

[0010] S1: Image generation preparation. Generate a high-definition target image using a self-developed generative image model, which serves as the carrier image for watermark embedding, ensuring that the image quality meets the display requirements of e-commerce platforms;

[0011] S2: Watermark information preprocessing. The original watermark information is encoded and converted to obtain watermarked encoded data to meet the requirements of neural network embedding and extraction;

[0012] S3: Build a watermark embedder based on a multi-channel neural network. This embeds the pre-encoded watermark information into the target image, generating a watermarked image containing the steganographic watermark. This embedder uses the Swin Transformer as its backbone network, leveraging its strengths in modeling long-range image dependencies and expressing multi-scale features to effectively improve the watermark's embedding stealth and capacity.

[0013] S4: Watermark Extraction Network Construction and Joint Training. Design a corresponding watermark extraction neural network model to recover the embedded watermark information from the watermarked image. Perform end-to-end training with the embedder to minimize the watermark reconstruction error and improve the stealth and robustness of the overall system.

[0014] Preferably, the decoding structure of the watermark extractor includes multiple upsampling convolutional layers, each layer is followed by a batch normalization layer and a GELU activation function, for gradually recovering the watermark encoding features.

[0015] Preferably, the method of the present invention further comprises constructing a watermark information dataset and introducing a variety of image perturbation operations (such as cropping, scaling, compression, etc.) during the training phase to enhance the robustness of the network against attacks.

[0016] Preferably, a composite loss function is introduced during the training process, including absolute difference loss, mean square error loss, gradient loss and connectivity loss, to comprehensively evaluate the watermark recovery effect and optimize the model performance.

[0017] In addition, the present invention also provides: a computer program product, comprising a computer-readable storage medium and instructions, wherein the instructions implement the above-mentioned image watermark information hiding method when executed by a processor.

[0018] An electronic device includes a processor, a memory and an input / output interface, wherein the memory stores computer program instructions for executing the above method, and the interface is used to receive an image to be processed and output an embedded or extracted watermark result.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] Strong concealment: Using Swin Transformer to model the deep semantic features of images, watermark embedding is "imperceptible" and has minimal impact on image quality, making it suitable for e-commerce product image display.

[0021] High robustness: Through joint training and composite loss function optimization, the system can effectively resist common image processing operations such as image compression, cropping, and filtering, ensuring that watermark information is stable and extractable.

[0022] Systematic design: From watermark encoding, image generation, neural network embedding and extraction, loss function design to dataset construction, we provide a complete end-to-end solution that is suitable for practical e-commerce scenarios.

[0023] Strong scalability: The method is compatible with other generative models and various neural network structures, and supports iterative optimization for future scenarios such as image content synthesis and copyright management. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the present invention;

[0025] Figure 2 Schematic diagram of the encoder-decoder portion of the algorithm model of the present invention;

[0026] Figure 3 It is a schematic diagram of the watermark removal process in the algorithm model of the present invention; DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0028] See also Figure 1-Figure 3 As shown, the present invention provides the following technical solutions:

[0029] A generative image watermark information hiding method based on multi-channel neural network. Figure 1 As shown, including:

[0030] S1. Prepare the generated image;

[0031] S2. Watermark information preprocessing, multi-channel neural network watermark embedding;

[0032] S3. Watermark extraction and joint training;

[0033] Implementation Method 1

[0034] This embodiment provides an image watermark information embedding system based on a multi-channel neural network. Figure 2 The system includes:

[0035] Input image module: used to receive the original image to be embedded with watermark;

[0036] Watermark information input module: used to receive watermark data, which can be information in any format such as text, image, QR code, etc.

[0037] Encoder (multi-channel neural network embedder): This module uses the improved ConvNeXt network as the backbone structure, accesses the multi-channel input mechanism, fuses the image and watermark information, and outputs the watermarked image;

[0038] The above encoder structure supports multi-scale and multi-channel parallel information processing, which improves the concealment and robustness of the watermark after embedding.

[0039] Output module: outputs image files embedded with watermark information.

[0040] Implementation method 2:

[0041] like Figure 3 As shown, the present invention further provides a watermark information extraction and verification process to ensure that the embedded information can be completely extracted and verified for consistency, which mainly includes:

[0042] Watermarked image input module: receives watermarked images from the outside;

[0043] Watermark extraction neural network: A decoder model based on the Transformer structure can extract embedded watermark information from images;

[0044] Watermark removal module: Based on the U-Net network structure design, it is used to remove watermark information from the image to restore the original image;

[0045] Image restoration module: outputs the image after stripping the watermark;

[0046] Verification module: compares the extracted watermark information with the original watermark and verifies their consistency through hash similarity or structural similarity measurement.

Claims

1. A generative image watermark information hiding method based on a multi-channel neural network, characterized by: The steps include: S1: Generate image preparation, using the self-developed generative image model to generate the target image as the carrier for watermark embedding; S2: Multi-channel neural network watermark embedding, building a multi-channel neural network watermark embedder to embed the pre-encoded watermark information into the generated target image to generate a watermarked image containing a steganographic watermark; S3: Design a corresponding watermark extraction neural network model to recover the embedded watermark information from the watermarked image, and optimize the embedder and decoder through end-to-end joint training to ensure the concealment and robustness of the watermark.

2. The method for hiding image watermark information based on a multi-channel neural network according to claim 1, characterized in that: The step S2 also includes the following steps: Watermark information preprocessing, encoding conversion of the original watermark information to generate watermark code.

3. The method for hiding image watermark information based on a multi-channel neural network according to claim 1, characterized in that: The watermark encoder uses Swin Transformer as the backbone network and performs encoding after removing its classifier module.

4. A generative image watermark information hiding method based on a multi-channel neural network, characterized in that: The SwinTransformer module includes: The feature projection module is used to convert the multi-scale features output by the encoder into a unified dimensional space; Multi-head self-attention module to capture long-range dependencies within the image; Feedforward neural network module for deep processing of the results of multi-head self-attention; Residual connection and layer normalization modules are used to enhance training stability and improve convergence speed.

5. The method for hiding image watermark information based on a multi-channel neural network according to claim 1, characterized in that: The decoder includes multiple upsampling convolution layers, each of which is followed by a batch normalization layer and a GELU activation function, for gradually generating watermark information encoding results.

6. The method for hiding image watermark information based on a multi-channel neural network according to any one of claims 1 to 5, characterized in that: Including constructing a watermark information dataset.

7. The method for hiding image watermark information based on a multi-channel neural network according to claim 6, characterized in that: The following steps are involved: For images with watermarks, neural network methods are used for training and removal.

8. The method for hiding image watermark information based on a multi-channel neural network according to claim 7, characterized in that: It also includes the introduction of loss functions in the training process, which include the sum of absolute differences, mean square error, gradient loss and connectivity loss, which are used to evaluate the watermark removal effect and optimize the model parameters.

9. A computer program product, characterized in that: The method comprises a computer-readable storage medium and computer program instructions, wherein the computer program instructions implement the method according to any one of claims 1 to 8 when executed by a processor.

10. An electronic device, characterized in that: comprising a processor, a memory and an input / output interface, wherein the memory stores computer program instructions; When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 8 is implemented; The input and output interface is used to receive the AI ​​generated image to be processed and output the image with watermark information added, and to receive the image with watermark information and output the AI ​​generated image with the watermark removed.

Images