An invisible watermark removal method based on parallel multi-resolution structure

Abstract

The present invention discloses an invisible watermark removal algorithm based on a parallel multi-resolution structure, in the field of image processing technology, characterized in that the method comprises the following steps: S1: constructing an invisible watermark removal network; the invisible watermark removal network comprises six stages and two modules, the two modules being a mask perception module and an attention guidance module; S2: constructing a loss function; S3: inputting a watermarked image and a mask into the network to achieve watermark removal. The present invention relates to the field of image processing technology, and in particular, to an invisible watermark removal algorithm based on a parallel multi-resolution structure. The present invention can efficiently eliminate embedded invisible watermarks without compromising image quality. The original practical value of the original image is retained, and the technology is not easily detected, providing a safe and efficient watermark removal solution.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to an invisible watermark removal algorithm based on a parallel multi-resolution structure. Background Art

[0002] In today's world where digital media is widely disseminated and shared, invisible watermarking technology has become an important tool for protecting intellectual property and ensuring data integrity. Invisible watermarking technology allows information to be discreetly embedded into digital media files for the purpose of protecting copyright, data tracking, and integrity verification. Attacks and defenses against invisible watermarks have become an important topic in the field of digital media security. As invisible watermarking technology continues to improve, invisible watermark removal technology is also evolving, bringing greater challenges to the protection of intellectual property rights in digital media.

[0003] After detecting the watermark information, the traditional invisible watermark removal algorithm can remove the watermark information from the digital image media to a certain extent. However, due to the negligible difference between the watermarked image and the non-watermarked image, after using the watermark removal network, in order to ensure the accurate removal of the watermark information, the image after the attack will be irreversibly damaged, which may destroy the integrity and readability of the watermarked image, thus affecting its practical application value. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an invisible watermark removal algorithm based on a parallel multi-resolution structure, which can efficiently remove the embedded invisible watermark without damaging the image quality. The original practical value of the original image is retained, and the technology is not easily detected, providing a safe and efficient watermark removal solution.

[0005] The present invention adopts the following technical solutions to achieve the invention objectives:

[0006] An invisible watermark removal algorithm based on a parallel multi-resolution structure is characterized by comprising the following steps:

[0007] S1: Construct invisible watermark removal network;

[0008] The invisible watermark removal network consists of six stages and two modules, the two modules are the mask perception module and the attention guidance module;

[0009] Each stage consists of four branches corresponding to four different resolutions;

[0010] The first five stages execute a mask-aware module at the end of each stage;

[0011] Before the last stage, an attention-guided fusion module is used;

[0012] The main task of the mask perception module is to accurately locate and focus on the specific area of ​​the image containing the watermark during the watermark removal process, so as to improve the efficiency and accuracy of the entire removal process. The module achieves this goal by creating a mask that can accurately identify the distribution of the watermark in the image and its characteristic performance at different resolution levels. Using this mask, the network can intelligently adjust its resource allocation and processing focus to concentrate on in-depth processing of the watermark area while avoiding unnecessary processing of the parts of the image that are not embedded with the watermark, effectively improving processing speed and resource utilization efficiency;

[0013] The core function of the attention guidance module is to enhance the efficiency and accuracy of the network in detecting and removing invisible watermarks. This module dynamically calculates the attention weights between feature maps at different resolutions to achieve intelligent guidance and selective focusing of the network focus, allowing the network to adaptively adjust its focus to focus on key areas in the image, especially those containing watermarks. This adaptive adjustment mechanism not only improves the efficiency of identifying and removing watermarks, but also optimizes the resource allocation of the network by reducing the processing of non-critical information. In addition, the attention guidance fusion module significantly improves the network's generalization ability for images of different types and qualities by emphasizing features related to watermarks while ignoring irrelevant interference information, thereby enhancing the robustness of the watermark removal algorithm.

[0014] S2: construct loss function;

[0015] S3: Input the watermarked image and mask into the network to achieve watermark removal.

[0016] As a further limitation of the present technical solution, the core concept of the mask perception module is to generate a mask that can identify the position and influence range of the watermark in the image and effectively reflect the characteristics of the watermark at different resolutions.

[0017] As a further limitation of the present technical solution, the invisible watermark removal network uses the mask to adjust its processing strategy, allocating more attention to the area where the watermark exists, while reducing the processing of the non-watermark area, thereby improving the efficiency of watermark removal.

[0018] As a further limitation of the present technical solution, the attention guidance module allows the network to dynamically and selectively focus on information of different resolutions by calculating attention weights between feature maps of different levels.

[0019] As a further limitation of the technical solution, the specific steps of the mask perception module are:

[0020] S111: First, pixels with higher priority are restored, and the mask values ​​are updated as follows:

[0021] (1);

[0022] in: It's pixels The current mask value of

[0023] is the priority to be defined, Only calculate on the border of the masked area;

[0024] is a hyperparameter, its value is 0.5;

[0025] is the priority of the boundary pixels The maximum value of

[0026] For the convolution window Center Pixel , its priority Defined as:

[0027] (2);

[0028] Among them, the resolution level Respectively represent 256 2 ,28 2 ,64 2 , 32 2 Feature map size;

[0029] S112: The defined priorities are the common priorities of different layers with different resolutions and specific resolution priorities ;

[0030] Calculate the number of pixels in the current convolution window that are not covered by the mask, surrounding pixels The more unmasked pixels there are, the The more it can be restored, because there is more contextual information to reduce uncertainty, and the closer it can be to the original image;

[0031] It is a specific resolution priority, which also means that the low-resolution layer and the high-resolution layer are different. As a high-resolution layer, the bottom branch Treated as a low-resolution layer.

[0032] As a further limitation of the technical solution, it is noted that the specific steps of the boot module are:

[0033] S121: First, high resolution The feature map is downsampled to the lowest resolution size;

[0034] S122: The downsampled feature map Connect them and calculate the attention score map based on this ;

[0035] S123: Applying attention score maps to feature maps of different resolutions , get the global structure information:

[0036] (3);

[0037] in: ;

[0038] yes No. item;

[0039] It is patches for compositing resolution levels No. Patches , since the higher resolution feature map is larger than 32 2 , The size of 8 2 ,4 2 , 2 2 , , capturing local patch information at each resolution respectively;

[0040] The size of 8 2 ,4 2 , 2 2 , capturing local patch information at each resolution respectively;

[0041] S124: Finally, structural information With feature map Fusion to get each resolution level The output feature map of :

[0042] (4).

[0043] As a further limitation of the technical solution, the specific steps of S2 are:

[0044] S21: Using reconstruction loss To calculate the final output With the original input image Distance between:

[0045] (5);

[0046] in: and It refers to the L1 norm, which represents the sum of the absolute values ​​of the pixel differences between two images;

[0047] S22: In the invisible watermark removal network, lower resolution images are also generated , , as the output of each stage; for the output of each stage And the corresponding original image Using the original reconstruction loss :

[0048] (6);

[0049] S23: In order to make the output image closer to the original image, a generative adversarial network is used to generate the output Using the Discriminator , discriminator The adversarial loss is defined as:

[0050] (7);

[0051] in: It is the discriminator, a network used to distinguish generated images from real images in the generative adversarial network;

[0052] Represents all original input images Computing the Discriminator The mathematical expectation of the log-probability of the output;

[0053] Indicates that all output results Computing the Discriminator The mathematical expectation of the log probability of identifying it as a fake image;

[0054] The adversarial nature of the multi-resolution network is defined as:

[0055] (8);

[0056] S24: The total loss function is:

[0057] (9);

[0058] in: and is the parameter;

[0059] When optimizing a generator, for ; When optimizing the discriminator, yes .

[0060] Compared with the prior art, the advantages and positive effects of the present invention are: the present invention integrates information of different scales using a multi-resolution network, so that it can handle global and local watermark embedding problems. The network consists of multiple branches of different resolutions, each branch focusing on a specific watermark removal task. The low-resolution branch is intended to process the global structure of the watermark, while the high-resolution branch targets the fine-grained texture details in the watermark content. These branches work in parallel and are integrated through a multi-resolution fusion mechanism, allowing watermarks to be removed while preserving image quality and content integrity. The main function of the mask-aware fusion module of the present invention is to selectively focus on the watermark-related areas in the image when processing invisible watermarks to improve the performance and accuracy of watermark removal, and the network achieves a good balance between preserving image content and removing watermarks. This also helps to enhance the robustness of the watermark removal algorithm. The use of an attention-guided fusion module helps the network to more effectively detect and remove invisible watermarks, thereby improving the performance and accuracy of watermark removal. The attention-guided fusion module allows the network to dynamically and selectively focus on information of different resolutions. Finally, the removal of invisible watermark information is completed. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the invisible watermark removal network structure of the present invention. DETAILED DESCRIPTION

[0062] A specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation.

[0063] The present invention comprises the following steps:

[0064] S1: Construct invisible watermark removal network;

[0065] The invisible watermark removal network consists of six stages and two modules, the two modules are the mask perception module and the attention guidance module;

[0066] In the invisible watermark removal network based on parallel multi-resolution structure, the mask-aware fusion module plays a vital role. Its main function is to selectively focus on the watermark-related areas in the image when processing invisible watermarks to improve the performance and accuracy of watermark removal.

[0067] The attention guided fusion module is a key component of the multi-resolution invisible watermark removal network. Its main role is to help the network detect and remove invisible watermarks more effectively, thereby improving the performance and accuracy of watermark removal.

[0068] Each stage consists of four branches corresponding to four different resolutions;

[0069] The first five stages execute a mask-aware module at the end of each stage;

[0070] Before the last stage, an attention-guided fusion module is used;

[0071] The core concept of the mask-aware module is to generate a mask that can identify the location and impact range of the watermark in the image, effectively reflecting the characteristics of the watermark at different resolutions.

[0072] The invisible watermark removal network uses the mask to adjust its processing strategy, allocating more attention to the areas where watermarks exist and reducing the processing of non-watermark areas, thereby improving the efficiency of watermark removal.

[0073] Through the mask-aware fusion module, the network achieves a good balance between preserving image content and removing watermarks. This also helps to enhance the robustness of the watermark removal algorithm, because the network can dynamically adjust to process images of different types and qualities without affecting the effectiveness of watermark removal. The introduction of this module is of great significance to improving the capabilities of invisible watermark removal technology.

[0074] The specific steps of the mask perception module are:

[0075] S111: First, pixels with higher priority are restored, and the mask values ​​are updated as follows:

[0076] (1);

[0077] in: It's pixels The current mask value of

[0078] is the priority to be defined, Only calculate on the border of the masked area;

[0079] is a hyperparameter, its value is 0.5;

[0080] is the priority of the boundary pixels The maximum value of

[0081] For the convolution window Center Pixel , its priority Defined as:

[0082] (2);

[0083] Among them, the resolution level Respectively represent 256 2 ,28 2 ,64 2 , 32 2 Feature map size;

[0084] S112: The defined priorities are the common priorities of different layers with different resolutions and specific resolution priorities ;

[0085] Calculate the number of pixels in the current convolution window that are not covered by the mask, surrounding pixels The more unmasked pixels there are, the The more it can be restored, because there is more contextual information to reduce uncertainty, and the closer it can be to the original image;

[0086] It is a specific resolution priority, which also means that the low-resolution layer and the high-resolution layer are different. As a high-resolution layer, the bottom branch Treat as a low-resolution layer. Low Resolution The network can focus on structural information because the receptive field of low-resolution feature maps is large, thus helping to collect global structural information. , we want these branches to focus on texture information. Here we use the context residual information to represent the texture information in the feature map.

[0087] The attention guidance module allows the network to dynamically and selectively focus on information at different resolutions by computing attention weights between feature maps at different levels.

[0088] This means that the network can adaptively adjust its focus based on the image content and the location of the watermark, thus improving the efficiency of identifying and removing watermarks. In addition, this module helps improve the generalization ability of the network by reducing unnecessary information interference and emphasizing features related to watermarks. This helps the network better adapt to images of different types and qualities, thereby improving the robustness of watermark removal.

[0089] Note that the specific steps for booting the module are:

[0090] In order to further fuse the high-level features (low resolution) and low-level features (high resolution) of the multi-resolution network, a self-attention-based attention guidance module is proposed to fuse the attention map calculated at each resolution into a shared attention map applicable to all resolutions. Through multi-resolution attention fusion, the low-resolution branch can better learn the global relationship from different scales, and the high-resolution branch can more comprehensively understand the overall structure and local texture of the image from the low resolution, so as to complete the entire image better and faster.

[0091] S121: First, high resolution The feature map is downsampled to the lowest resolution Size ;

[0092] S122: The downsampled feature map Connect them and calculate the attention score map based on this ;

[0093] S123: Applying attention score maps to feature maps of different resolutions , get the global structure information:

[0094] (3);

[0095] in: ;

[0096] yes No. item;

[0097] It is patches for compositing resolution levels No. Patches , since the higher resolution feature map is larger than 32 2 , The size of 8 2 ,4 2 , 2 2 , , capturing local patch information at each resolution respectively;

[0098] The size of 8 2 ,4 2 , 2 2 , capturing local patch information at each resolution respectively;

[0099] S124: Finally, structural information With feature map Fusion to get each resolution level The output feature map of :

[0100] (4).

[0101] In summary, in the invisible watermark removal network based on parallel multi-resolution, the role of the attention guided fusion module is to optimize the feature selection process, improve the accuracy and performance of watermark detection and removal, and improve the generalization ability of the network in different invisible watermark application scenarios. The introduction of this module is crucial to improving the effectiveness of the watermark removal method.

[0102] S2: Construct loss function.

[0103] The specific steps of S2 are:

[0104] S21: Using reconstruction loss To calculate the final output With the original input image Distance between:

[0105] (5);

[0106] in: and It refers to the L1 paradigm, which represents the sum of the absolute values ​​of the pixel differences between two images and is a way to evaluate image similarity;

[0107] S22: In the invisible watermark removal network, lower resolution images are also generated , , as the output of each stage; for the output of each stage And the corresponding original image Using the original reconstruction loss :

[0108] (6);

[0109] S23: In order to make the output image closer to the original image, a generative adversarial network is used to generate the output Using the Discriminator , discriminator The adversarial loss is defined as:

[0110] (7);

[0111] in: It is the discriminator, a network used to distinguish generated images from real images in the generative adversarial network;

[0112] Represents all original input images Computing the Discriminator The mathematical expectation of the log-probability of the output;

[0113] Indicates that all output results Computing the Discriminator The mathematical expectation of the log probability of identifying it as a fake image;

[0114] The adversarial nature of the multi-resolution network is defined as:

[0115] (8);

[0116] S24: The total loss function is:

[0117] (9);

[0118] in: and For the parameter, set the parameter to 10;

[0119] When optimizing a generator, for ; When optimizing the discriminator, yes .

[0120] S3: Input the watermarked image and mask into the network to achieve watermark removal.

[0121] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. An invisible watermark removal method based on a parallel multi-resolution structure, characterized in that: The following steps are involved: S1: Construct invisible watermark removal network; The invisible watermark removal network consists of six stages and two modules, the two modules are the mask perception module and the attention guidance module; Each stage consists of four branches corresponding to four different resolutions; The first five stages execute a mask-aware module at the end of each stage; Before the last stage, an attention-guided fusion module is used; S2: construct loss function; S3: Input the watermarked image and mask into the network to remove the watermark; The specific steps of the mask perception module are: S111: First, pixels with higher priority are restored, and the mask values ​​are updated as follows: （1）； in: It's pixels The current mask value of is the priority to be defined, Only calculate on the border of the masked area; is a hyperparameter, its value is 0.5; is the priority of the boundary pixels The maximum value of For the convolution window Center Pixel , its priority Defined as: （2）； Among them, the resolution level Respectively represent 256 2 ,28 2 ,64 2 , 32 2 Feature map size; S112: The defined priorities are the common priorities of different layers with different resolutions and specific resolution priorities ; Calculate the number of pixels in the current convolution window that are not covered by the mask, surrounding pixels The more unmasked pixels there are, the The more it can be restored, because there is more contextual information to reduce uncertainty, and the closer it can be to the original image; It is a specific resolution priority, which also means that the low-resolution layer and the high-resolution layer are different. As a high-resolution layer, the bottom branch Treated as a low-resolution layer; Note that the specific steps for booting the module are: S121: First, high resolution The feature map is downsampled to the lowest resolution size; S122: The downsampled feature map Connect them and calculate the attention score map based on this ; S123: Applying attention score maps to feature maps of different resolutions , get the global structure information: （3）； in: ; yes No. item; It is patches for compositing resolution levels No. Patches , since the higher resolution feature map is larger than 32 2 , The size of 8 2 ,4 2 , 2 2 , , capturing local patch information at each resolution respectively; The size of 8 2 ,4 2 , 2 2 , capturing local patch information at each resolution respectively; S124: Finally, structural information With feature map Fusion to get each resolution level The output feature map of : （4）。 2. The invisible watermark removal method based on parallel multi-resolution structure according to claim 1 is characterized by: The core concept of the mask-aware module is to generate a mask that identifies the location and impact range of the watermark in the image, effectively reflecting the characteristics of the watermark at different resolutions.

3. The invisible watermark removal method based on parallel multi-resolution structure according to claim 2 is characterized by: The invisible watermark removal network uses the mask to adjust its processing strategy, allocating more attention to the areas where watermarks exist and reducing the processing of non-watermark areas, thereby improving the efficiency of watermark removal.

4. The invisible watermark removal method based on parallel multi-resolution structure according to claim 1 is characterized by: The attention guidance module allows the network to dynamically and selectively focus on information at different resolutions by computing attention weights between feature maps at different levels.

5. The invisible watermark removal method based on parallel multi-resolution structure according to claim 1 is characterized in that: The specific steps of S2 are: S21: Using reconstruction loss To calculate the final output With the original input image Distance between: （5）； in: and It refers to the L1 norm, which represents the sum of the absolute values ​​of the pixel differences between two images; S22: In the invisible watermark removal network, lower resolution images are also generated , , as the output of each stage; for the output of each stage And the corresponding original image Using the original reconstruction loss : （6）； S23: In order to make the output image closer to the original image, a generative adversarial network is used to generate the output Using the Discriminator , discriminator The adversarial loss is defined as: （7）； in: It is the discriminator, a network used to distinguish generated images from real images in the generative adversarial network; Represents all original input images Computing the Discriminator The mathematical expectation of the log-probability of the output; Indicates that all output results Computing the Discriminator The mathematical expectation of the log probability of identifying it as a fake image; The adversarial nature of the multi-resolution network is defined as: （8）； S24: The total loss function is: （9）； in: and is the parameter; When optimizing a generator, for ; When optimizing the discriminator, yes .

Images