A zero watermark generation method based on double-loss orthogonal decoupling and multi-scale fusion and application

Abstract

The application relates to the technical field of digital image copyright protection and information security, and discloses a zero watermark generation method based on double-loss orthogonal decoupling and multi-scale fusion and application. The method comprises the following steps: constructing a zero watermark generation model; training a multi-scale feature extraction network by using a double-channel composite data enhancement strategy to obtain a trained multi-scale feature extraction network; freezing the parameters of the trained multi-scale feature extraction network, training an orthogonal feature fusion network by using a loss function based on row-column orthogonal loss and noise consistency loss, and obtaining a trained orthogonal feature fusion network. The target host image and the target copyright image are input into the zero watermark generation model to obtain a watermark feature vector; and the watermark feature vector is subjected to binaryzation processing to output a watermark sequence. The application can balance high robustness and high distinguishability, and is suitable for high-precision and large-scale image copyright protection requirements.

Description

Technical Field

[0001] This application relates to the fields of digital image copyright protection and information security technology, and in particular to a zero-watermark generation method and its application based on dual-loss orthogonal decoupling and multi-scale fusion. Background Technology

[0002] With the rapid development of digital media industries such as short videos, digital painting, remote sensing mapping, and medical imaging, the barriers to online distribution, secondary editing, and reprinting of image materials have continued to decrease. Unauthorized use, secondary commercialization, and forgery of copyright ownership are becoming increasingly rampant, posing a severe challenge to the confirmation and traceability of digital image copyrights. Traditional embedded digital watermarking, which embeds copyright information by modifying image pixels or frequency domain coefficients, can achieve traceability, but it introduces visual distortion, significantly limiting its application in scenarios with strict constraints on image fidelity, such as medical imaging, high-precision surveying and remote sensing, and commercial photography.

[0003] Zero-watermarking technology, as a lossless copyright protection solution, extracts only the inherent stable features of the image itself and fuses them with the copyright identifier to generate a copyright certificate. It does not modify any pixels of the host image, perfectly preserving the integrity of the original image, and has become the mainstream technology for high-precision image copyright protection. Existing zero-watermarking technologies are mainly divided into two categories: one is manual feature schemes based on frequency domain transformation and invariant moments. These methods heavily rely on manual priors and lack robustness against rotation, cropping, and mixing attacks, with limited feature discrimination and a high risk of collision misjudgment in large-scale image libraries. The other is deep learning-based zero-watermarking schemes, which rely on the powerful feature representation capabilities of convolutional networks to improve robustness. However, existing methods mostly use classification loss or ordinary contrast loss, only constraining the consistency of features in images from the same source, without decoupling watermark features from different hosts and copyrights. This results in severe coupling of watermark features between different host-copyright combinations, making it difficult to meet the commercial needs of large-scale copyright libraries in terms of discrimination. Furthermore, existing deep zero-watermarking methods mostly use single-scale high-level features, losing shallow textures and local information, leading to insufficient watermark stability against local attacks such as cropping and occlusion.

[0004] In summary, existing deep zero-watermarking methods struggle to simultaneously achieve both high robustness and high discriminative power, making them unsuitable for the high-precision, large-scale image copyright protection needs. Summary of the Invention

[0005] The purpose of this invention is to address at least one deficiency in the existing technology and provide a zero-watermark generation method and application based on dual-loss orthogonal decoupling and multi-scale fusion. This invention can balance high robustness and high discriminability, while adapting to the needs of high-precision, large-scale image copyright protection.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a zero-watermark model training method based on dual-loss orthogonal decoupling and multi-scale fusion, the method comprising the following steps: A zero-watermark generation model is constructed, which includes a multi-scale feature extraction network and an orthogonal feature fusion network. The multi-scale feature extraction network includes a spatial transformation sub-network and a ConvNeXt-Small backbone sub-network. The orthogonal feature fusion network includes a multi-scale spatial pyramid attention sub-network and a residual fusion network. The multi-scale feature extraction network is trained using a dual-channel composite data augmentation strategy to obtain a trained multi-scale feature extraction network. Freeze the parameters of the trained multi-scale feature extraction network, and train the orthogonal feature fusion network using a loss function based on row and column orthogonality loss and noise consistency loss to obtain the trained orthogonal feature fusion network; The trained multi-scale feature extraction network and the trained orthogonal feature fusion network are used as the final zero-watermark generation model.

[0007] Furthermore, the spatial transformation sub-network is composed of a localization network, a mesh generator, and a bilinear sampler connected in sequence; the localization network is used to regress affine transformation parameters based on the input image; the mesh generator is used to generate sampling mesh coordinates based on the affine transformation parameters; and the bilinear sampler is used to acquire pixels from the input image based on the sampling mesh coordinates and output a geometrically corrected image.

[0008] Furthermore, the dual-channel composite data enhancement strategy includes: (1). Apply weak enhancement processing to the original images in the training dataset to obtain weakly enhanced samples; the weak enhancement processing is to randomly select one attack method from rotation, cropping, salt and pepper noise, Gaussian noise, filtering and JPEG compression to process the original image; (2). Apply strong enhancement processing to the same original image to obtain a strongly enhanced sample; the strong enhancement processing is to apply large-angle rotation, extreme proportion random cropping to the original image in sequence, and then superimpose Gaussian noise, salt and pepper noise, median filtering, Gaussian filtering and multiple attacks in JPEG compression; (3) Weakly enhanced samples and strongly enhanced samples of the same original image form positive sample pairs, while enhanced samples of different original images form negative sample pairs.

[0009] Furthermore, when training the multi-scale feature extraction network using a dual-channel composite data augmentation strategy, the SimCLR self-supervised contrastive learning framework is employed. The total training loss function is a weighted sum of the NT-Xent contrastive loss and the inverse consistency regularization loss. The NT-Xent contrastive loss is used to maximize the feature similarity of positive sample pairs and minimize the feature similarity of negative sample pairs. The inverse consistency regularization loss is used to constrain the feature consistency of the strongly augmented samples before and after applying additional affine perturbations. The total loss function is expressed as follows:

[0010] Among them, For NT-Xent contrast loss; This is the loss for inverse consistency regularization; This is the weight coefficient for the regularization term.

[0011] Furthermore, the multi-scale spatial pyramid attention subnetwork includes a channel unification module, a spatial pyramid pooling module, a squeeze-excitation channel attention module, and a hierarchical attention fusion module; The channel unification module is used to perform 1×1 convolution on the input multi-level feature maps respectively, and unify the channel dimensions of each level of feature maps to the same dimension. The spatial pyramid pooling module is used to perform average pooling on the feature maps of each level after unifying the channel, and then stitch them together with multiple grid sizes to obtain global-local descriptors for each level. The squeeze-excitation channel attention module is used to recalibrate the channel weights of the descriptors at each level; The hierarchical attention fusion module is used to perform weighted fusion of the recalibrated descriptors at each level and output multi-scale fusion features.

[0012] Furthermore, the residual fusion subnetwork includes a feature stitching unit, a linear projection branch, a multilayer perceptron branch, and a weighted synthesis unit; The feature splicing unit is used to splice the global features of the host image and the global features of the copyright image into a joint feature; The linear projection branch is used to perform a linear transformation on the joint features to obtain a first output; The multilayer perceptron branch is used to perform nonlinear mapping on the joint features to obtain a second output; The weighted synthesis unit is used to perform a weighted summation of the first output and the second output, and to perform L2 normalization on the weighted summation result to obtain the watermark feature vector located on the unit hypersphere.

[0013] Furthermore, the row-column orthogonal loss is used to constrain the watermark feature vectors generated by pairwise combinations of multiple host images and multiple copyright images in the same training batch to be pairwise orthogonal in both the row and column directions; the row direction of the watermark feature vector is the combination direction of different host images corresponding to the same copyright image, and the column direction of the watermark feature vector is the combination direction of different copyright images corresponding to the same host image; the noise consistency loss is used to constrain the cosine similarity between the features input to the orthogonal feature fusion network and the watermark feature vector outputs before and after Gaussian noise addition to be greater than or equal to a preset threshold; wherein, the mathematical expression of the row-column orthogonal loss is:

[0014] Where B represents the number of host images or copyrighted images in the training batch. The watermark feature vector generated for the i-th host image and the p-th copyright image; The mathematical expression for the noise consistency loss is:

[0015] in, The watermark vector generated for clean features. The watermark vector generated for the noisy features, The cosine similarity function; The mathematical expression of the loss function based on row-column orthogonality loss and noise consistency loss is as follows:

[0016] in, This is the weighting coefficient for noise consistency loss.

[0017] Secondly, the present invention also provides a zero-watermark generation method based on dual-loss orthogonal decoupling and multi-scale fusion, the method comprising: The target host image and the target copyright image are input into the zero watermark generation model trained by the method described in the first aspect to obtain the watermark feature vector; The watermark feature vector is binarized to output the watermark sequence.

[0018] Thirdly, the present invention also provides a copyright authentication method based on dual-loss orthogonal decoupling and multi-scale fusion, wherein the method is an application of the zero-watermark generation method based on dual-loss orthogonal decoupling and multi-scale fusion described in the second aspect, and the method includes: Using the method described in the second aspect, a first watermark feature vector corresponding to the target host image and the target copyright image is generated. The first watermark feature vector is binarized to obtain a first watermark sequence. Global features of the target copyright image are extracted to generate a scrambling key. The first watermark sequence is scrambled and encrypted using the scrambling key. The encrypted first watermark sequence is bound to the target copyright image and stored in the authentication library. The image to be verified and the image to be confirmed for copyright are input into the zero watermark generation model in the same way and binarized to obtain the second watermark sequence. A scrambling key is generated in the same way and the second watermark sequence is scrambled and encrypted. Calculate the normalized cross-correlation value between the encrypted second watermark sequence and the encrypted first watermark sequence in the authentication library. If the normalized cross-correlation value is greater than or equal to a preset threshold, it is determined that the image to be verified and the image to be confirmed have a copyright ownership relationship.

[0019] Furthermore, the scrambling key is generated as follows: extract the global feature vector of the copyrighted image, take the weighted sum of the first few dimensions of the global feature vector and round it to generate a random number seed; the scrambling encryption is a one-dimensional Fibonacci affine scrambling based on the random number seed; The one-dimensional Fibonacci affine scrambling involves iteratively performing multiple rounds of affine index rearrangement. Each round generates coefficients and offsets coprime to the length of the watermark sequence. The watermark sequence index is then rearranged after undergoing an affine transformation based on these coefficients and offsets. The affine transformation formula is as follows:

[0020] in For the original index, For the scrambled index, For the transformation parameters generated in each round, This represents the length of the watermark sequence.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention embeds a spatial transformation network at the front end of the backbone network, and combines a weak-strong dual-channel composite attack enhancement strategy with self-supervised comparative learning to improve the model’s resistance to geometric attacks such as rotation, scaling and translation from the two levels of input correction and feature learning. It can still maintain a high watermark matching degree under large-angle rotation.

[0022] (2) The present invention designs a multi-scale spatial pyramid attention module, which integrates three layers of features: shallow texture, middle structure and high semantic. It captures multi-granular local information through spatial pyramid pooling and combines channel attention and hierarchical attention with adaptive weighting to significantly improve the stability of watermarks in local destruction scenarios such as cropping and occlusion.

[0023] (3) Propose a row-column bidirectional orthogonal loss, which directly forces the watermark features of different hosts and different copyrights to decouple from each other in the vector space. Combined with a short watermark sequence, it can achieve a collision rate far lower than the existing scheme, greatly reducing the risk of mismatch under a large-scale copyright library and meeting the high uniqueness requirements of commercial rights confirmation.

[0024] (4) A dual-loss interlocking mechanism of row and column orthogonal loss and noise consistency loss is adopted. The orthogonal loss ensures high discrimination, while the noise consistency loss maintains feature robustness and training stability, thus breaking through the inherent performance bottleneck of existing schemes where "high robustness means poor discrimination, and high discrimination means poor robustness".

[0025] (5) Introduce dynamic Fibonacci affine scrambling based on copyright characteristics to achieve a Figure 1 The key is naturally bound to the copyright, requiring no additional management overhead; the scrambling transformation is a reversible bijective, which does not affect the watermark matching accuracy at all, improving security without sacrificing the performance of rights confirmation.

[0026] (6) The two-stage decoupled training architecture is adopted, and the training process is stable and easy to converge; the generated short watermark sequence has low storage overhead and fast matching calculation speed, which is suitable for the efficient retrieval and rights confirmation needs of large-scale copyright databases. Attached Figure Description

[0027] Figure 1 This is a flowchart of a zero-watermarking model training method based on dual-loss orthogonal decoupling and multi-scale fusion according to Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the spatial transformation subnetwork of Embodiment 1 of this aspect; Figure 3 This is a schematic diagram of the structure of the multi-scale spatial pyramid attention network in Embodiment 1 of the present invention; Figure 4 This is a flowchart of the Fibonacci dynamic scrambling registration and verification process based on copyright features in Embodiment 1 of the present invention; Figure 5 This is a bar chart comparing the average robust NC and collision NC under different configurations in the ablation experiment of Embodiment 2 of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] Example 1 Please see Figure 1A preferred embodiment of the present invention provides a zero-watermarking model training method based on dual-loss orthogonal decoupling and multi-scale fusion, the method comprising the following steps: S1: Construct a zero-watermark generation model, which includes a multi-scale feature extraction network and an orthogonal feature fusion network; the multi-scale feature extraction network includes a spatial transformation sub-network and a ConvNeXt-Small backbone sub-network; the orthogonal feature fusion network includes a multi-scale spatial pyramid attention sub-network and a residual fusion sub-network. S2: Train the multi-scale feature extraction network using a dual-channel composite data augmentation strategy to obtain a trained multi-scale feature extraction network; S3: Freeze the parameters of the trained multi-scale feature extraction network, and train the orthogonal feature fusion network using a loss function based on row and column orthogonal loss and noise consistency loss to obtain the trained orthogonal feature fusion network; S4: The trained multi-scale feature extraction network and the trained orthogonal feature fusion network are used as the final zero-watermark generation model.

[0030] This invention embeds a spatial transformation network at the front end of the backbone network, combining a weak-strong dual-channel composite attack enhancement strategy with self-supervised contrastive learning. This improves the model's resistance to geometric attacks such as rotation, scaling, and translation from two levels: input correction and feature learning, maintaining high watermark matching accuracy even under large-angle rotations. A multi-scale spatial pyramid attention module is designed, fusing shallow texture, mid-level structure, and high-level semantic features. It captures multi-granularity local information through spatial pyramid pooling, and combines channel attention and hierarchical attention with adaptive weighting, significantly improving watermark stability in scenarios with localized damage such as cropping and occlusion.

[0031] This paper proposes a bidirectional orthogonal loss method, which directly forces the decoupling of watermark features from different hosts and copyrights in the vector space. Combined with short watermark sequences, this achieves a collision rate far lower than existing schemes, significantly reducing the risk of mismatches in large-scale copyright databases and meeting the high uniqueness requirements for commercial rights confirmation. A dual-loss interlocking mechanism of bidirectional orthogonal loss and noise consistency loss is employed. The orthogonal loss ensures high discriminative power while the noise consistency loss maintains feature robustness and training stability, overcoming the inherent performance bottleneck of existing schemes where "high robustness equates to poor discriminative power, and vice versa." Dynamic Fibonacci affine scrambling based on copyright features is introduced to achieve a... Figure 1 The key is naturally bound to the copyright, requiring no additional management overhead; the scrambling transformation is a reversible bijective, which does not affect the watermark matching accuracy at all, improving security without sacrificing rights confirmation performance. A two-stage decoupled training architecture is adopted, ensuring a stable and easily convergent training process; the generated short watermark sequences have low storage overhead and fast matching calculation speed, adapting to the efficient retrieval and rights confirmation needs of large-scale copyright databases.

[0032] In an optional embodiment, in step S1, a zero-watermark generation model is constructed, the zero-watermark generation model including a multi-scale feature extraction network and an orthogonal feature fusion network; the multi-scale feature extraction network includes a spatial transformation sub-network and a ConvNeXt-Small backbone sub-network, and the orthogonal feature fusion network includes a multi-scale spatial pyramid attention sub-network and a residual fusion sub-network; Further, please refer to Figure 2 The spatial transformation sub-network consists of a localization network, a mesh generator, and a bilinear sampler connected sequentially. The localization network employs a lightweight convolutional structure to regress 6D affine transformation parameters from the input image. The mesh generator generates sampling mesh coordinates based on the 6D affine transformation parameters. The bilinear sampler collects pixels from the input image based on the sampling mesh coordinates and outputs a geometrically corrected image. In this embodiment, the spatial transformation network is initialized with an identity transformation to ensure stable output during the initial training phase. The mathematical expression for the affine transformation is:

[0033] In the formula Input image coordinates, These are the corrected coordinates.

[0034] Further, please refer to Figure 3 The multi-scale spatial pyramid attention subnetwork includes a channel unification module, a spatial pyramid pooling module, a squeeze-excitation channel attention module, and a hierarchical attention fusion module; The channel unification module is used to perform 1×1 convolution on the input multi-level feature maps respectively, and unify the channel dimensions of each level of feature maps to the same dimension. The spatial pyramid pooling module is used to perform average pooling on the feature maps of each level after unifying the channel, and then stitch them together with multiple grid sizes to obtain global-local descriptors for each level. The squeeze-excitation channel attention module is used to recalibrate the channel weights of the descriptors at each level; The hierarchical attention fusion module is used to perform weighted fusion of the recalibrated descriptors at each level and output multi-scale fusion features.

[0035] For example, the multi-scale spatial pyramid attention sub-network receives three feature maps (stage2, stage3, and stage4) output by ConvNeXt, with sizes of 28×28, 14×14, and 7×7, respectively. The processing flow is as follows: Channel unification: The three feature layers are each processed by a 1×1 convolution, resulting in a unified number of channels of 768. Spatial pyramid pooling: Average pooling of 1×1, 2×2, and 4×4 grids is performed on each feature layer, and the layers are concatenated to obtain the global-local feature descriptor for that level. SE channel attention: Each descriptor is processed by global average pooling and two fully connected layers to generate channel weights, and the descriptors are recalibrated by channel. Hierarchical attention fusion: The fusion weights of the three feature layers are learned through a lightweight hierarchical attention network, and the weighted sum is used to obtain a 768-dimensional multi-scale fusion feature vector. The host image and the copyright image share the weights of this module, and each outputs its own fusion feature.

[0036] Furthermore, the residual fusion subnetwork includes a feature stitching unit, a linear projection branch, a multilayer perceptron branch, and a weighted synthesis unit; The feature splicing unit is used to splice the global features of the host image and the global features of the copyright image into a joint feature; The linear projection branch is used to perform a linear transformation on the joint features to obtain a first output; The multilayer perceptron branch is used to perform nonlinear mapping on the joint features to obtain a second output; The weighted synthesis unit is used to perform a weighted summation of the first output and the second output, and to perform L2 normalization on the weighted summation result to obtain the watermark feature vector located on the unit hypersphere.

[0037] For example, the host fusion feature and the copyright fusion feature are concatenated into a 1536-dimensional joint feature, which is then input into the fusion network. The network includes a linear projection branch and an MLP residual branch, and the output is calculated as follows:

[0038] The MLP contains a 768-dimensional hidden layer, combined with LayerNorm and Dropout. The output vector is L2 normalized to obtain a 128-dimensional unit watermark vector.

[0039] This embodiment embeds a spatial transformation network at the front end of the ConvNeXt backbone and combines weak-strong dual-channel composite enhancement with inverse consistency regularization training. This enables the multi-scale feature extraction network to obtain active correction capabilities and distortion robustness against geometric deformations such as rotation, translation, and scaling, thereby ensuring the feature stability of subsequent watermark generation under extreme geometric attacks.

[0040] In an optional embodiment, in step S2, the multi-scale feature extraction network is trained using a dual-channel composite data augmentation strategy to obtain a trained multi-scale feature extraction network. The dual-channel composite data enhancement strategy includes: (1). Apply weak enhancement processing to the original images in the training dataset to obtain weakly enhanced samples; the weak enhancement processing is to randomly select one attack method from rotation, cropping, salt and pepper noise, Gaussian noise, filtering and JPEG compression to process the original image; (2). Apply strong enhancement processing to the same original image to obtain a strongly enhanced sample; the strong enhancement processing is to apply large-angle rotation, extreme proportion random cropping to the original image in sequence, and then superimpose Gaussian noise, salt and pepper noise, median filtering, Gaussian filtering and multiple attacks in JPEG compression; (3) Weakly enhanced samples and strongly enhanced samples of the same original image form positive sample pairs, while enhanced samples of different original images form negative sample pairs.

[0041] For example, weak enhancement processing: randomly select one of the following to be applied to the original image: continuous rotation from ±10° to ±30°, random cropping at a scale of 0.75 to 0.9, salt and pepper noise with a probability of 0.01, Gaussian noise with a variance of 0.01, 3×3 Gaussian filtering, and JPEG compression with a quality factor of 80. Strong enhancement processing: First, apply a large-angle rotation of ±120°~±180° with a 50% probability, then perform extreme random cropping with a ratio of 0.15~1.0, and then successively add Gaussian noise with a variance of 0.1 with a 70% probability, salt and pepper noise with a probability of 0.1 with a 70% probability, 11×11 median filtering with a 60% probability, 11×11 Gaussian filtering with a 60% probability, and JPEG compression with a quality factor of 20 with an 80% probability.

[0042] Furthermore, when training the multi-scale feature extraction network using a dual-channel composite data augmentation strategy, the SimCLR self-supervised contrastive learning framework is employed. The total training loss function is a weighted sum of the NT-Xent contrastive loss and the inverse consistency regularization loss. The NT-Xent contrastive loss is used to maximize the feature similarity of positive sample pairs and minimize the feature similarity of negative sample pairs. The inverse consistency regularization loss is used to constrain the feature consistency of the strongly augmented samples before and after applying additional affine perturbations. The total loss function is expressed as follows:

[0043] Among them, For NT-Xent contrast loss; This is the loss for inverse consistency regularization; This is the weight coefficient for the regularization term.

[0044] During training, the AdamW optimizer was used, with the backbone network learning rate set to 1-4 and the projection head learning rate set to 1-3. Cosine annealing was used for learning rate scheduling, and the training lasted for 100 epochs. After training, the feature extractor weights were retained for subsequent application stages.

[0045] This embodiment trains a multi-scale feature extraction network using a weak-strong dual-channel composite data augmentation strategy. This allows the network to be fully exposed to various single and composite attack scenarios during the self-supervised comparative learning process, thereby learning robust feature representations that are highly resistant to various signal interferences such as rotation, cropping, noise, and compression. This provides a stable multi-level feature map foundation for subsequent watermark generation.

[0046] In an optional embodiment, in step S3, the parameters of the trained multi-scale feature extraction network are frozen, and the orthogonal feature fusion network is trained using a loss function based on row and column orthogonal loss and noise consistency loss to obtain the trained orthogonal feature fusion network. The row-column orthogonality loss is used to constrain the watermark feature vectors generated by pairwise combinations of multiple host images and multiple copyright images in the same training batch to be pairwise orthogonal in both the row and column directions; the row direction of the watermark feature vector is the combination direction of different host images corresponding to the same copyright image, and the column direction of the watermark feature vector is the combination direction of different copyright images corresponding to the same host image; the noise consistency loss is used to constrain the cosine similarity between the features input to the orthogonal feature fusion network and the output watermark feature vectors before and after Gaussian noise addition to be greater than or equal to a preset threshold; wherein, the mathematical expression of the row-column orthogonality loss is:

[0047] Where B represents the number of host images or copyrighted images in the training batch. The watermark feature vector generated for the i-th host image and the p-th copyright image; The mathematical expression for the noise consistency loss is:

[0048] in, The watermark vector generated for clean features. The watermark vector generated for the noisy features, The cosine similarity function; The mathematical expression of the loss function based on row-column orthogonality loss and noise consistency loss is as follows:

[0049] in, This is the weighting coefficient for noise consistency loss.

[0050] In practice, the AdamW optimizer was used with a learning rate of 1-3, and cosine annealing was employed for training for 150 epochs, resulting in a well-trained orthogonal feature fusion network.

[0051] This embodiment uses row-column orthogonal loss to force watermark features of different host-copyright combinations to decouple in the feature space, significantly reducing the probability of watermark collisions. At the same time, it combines noise consistency loss to enhance the anti-interference ability of the fusion network against feature perturbations, maintaining the robust stability of watermark features while ensuring high discriminability.

[0052] Finally, the trained multi-scale feature extraction network and the trained orthogonal feature fusion network are used as the final zero-watermark generation model.

[0053] Furthermore, embodiments of the present invention also provide a zero-watermark generation method based on dual-loss orthogonal decoupling and multi-scale fusion, the method comprising: The target host image and the target copyright image are input into the zero watermark generation model trained by the method to obtain the watermark feature vector; The watermark feature vector is binarized to output the watermark sequence.

[0054] Example 2 This embodiment is an application of the method proposed in Embodiment 1, specifically a copyright authentication method based on dual-loss orthogonal decoupling and multi-scale fusion. The method includes: Using the method described in Example 1, a first watermark feature vector corresponding to the target host image and the target copyright image is generated. The first watermark feature vector is binarized to obtain a first watermark sequence. Global features of the target copyright image are extracted to generate a scrambling key. The first watermark sequence is scrambled and encrypted using the scrambling key. The encrypted first watermark sequence is bound to the target copyright image and stored in the authentication database. The image to be verified and the image to be confirmed for copyright are input into the zero watermark generation model in the same way and binarized to obtain the second watermark sequence. A scrambling key is generated in the same way and the second watermark sequence is scrambled and encrypted. Calculate the normalized cross-correlation value between the encrypted second watermark sequence and the encrypted first watermark sequence in the authentication library. If the normalized cross-correlation value is greater than or equal to a preset threshold, it is determined that the image to be verified and the image to be confirmed have a copyright ownership relationship.

[0055] Furthermore, the scrambling key is generated as follows: extract the global feature vector of the copyrighted image, take the weighted sum of the first few dimensions of the global feature vector and round it to generate a random number seed; the scrambling encryption is a one-dimensional Fibonacci affine scrambling based on the random number seed; The one-dimensional Fibonacci affine scrambling involves iteratively performing multiple rounds of affine index rearrangement. Each round generates coefficients and offsets coprime to the length of the watermark sequence. The watermark sequence index is then rearranged after undergoing an affine transformation based on these coefficients and offsets. The affine transformation formula is as follows:

[0056] in For the original index, For the scrambled index, For the transformation parameters generated in each round, This represents the length of the watermark sequence.

[0057] For example, please refer to Figure 4 In practical applications, it includes two phases: the registration phase and the verification phase. I. Registration Phase: The steps are as follows: (1). Enter the host image to be registered and the corresponding copyright logo image; (2). Input the two images into the feature extractor and the SPP-PAN module respectively to obtain their respective multi-scale fusion features; (3). After concatenating the two fusion features, input them into the fusion network to generate a 128-dimensional floating-point watermark vector; (4). Perform sign function binarization on each element of the vector, mapping positive numbers to +1 and negative numbers to -1, to obtain a binary watermark sequence; (5). Extract the global features of the copyright image, take the first 8 components, sum them by weight according to the sequence number, multiply by 106, round down, and then take the modulus of 231-1 to generate a random seed for scrambling; drive 5 rounds of one-dimensional Fibonacci affine scrambling to encrypt the binary watermark sequence, and randomly generate coefficients a and offset b that are coprime to the watermark length of 128 in each round, perform index affine transformation and rearrange the sequence; (6) Bind the encrypted watermark sequence with the corresponding copyright information and store it in a third-party copyright authentication database.

[0058] II. Verification Phase, the steps are as follows: (1) Input the disputed image to be verified and the copyrighted image for which rights are claimed; (2). Generate the watermark sequence to be verified by following the same feature extraction, fusion, binarization and scrambling process as the registration phase; (3) Retrieve the corresponding registered watermark sequence from the copyright database and calculate the normalized cross-correlation value between the two:

[0059] Where L is the length of the watermark sequence, which is 128. For the k-th position of the registration sequence, The k-th element of the sequence to be verified; (4). Judgment rule: If NC≥0.7, the copyright is judged to match; otherwise, it is judged to not match. The threshold of 0.7 is the optimal compromise value determined by combining the heterogeneous watermark collision distribution and the robust distribution of homogeneous attacks, taking into account both extremely low false positive rate and low false negative rate.

[0060] Example 3 Please see Figure 5 This embodiment verifies the performance of the method through experiments. The Mini-ImageNet dataset is used as the host image source, and 38,400 images are randomly selected as the training set, 9,600 images as the validation set, and 12,000 images as the test set. 200 different school emblem images are used as copyright images.

[0061] For a fixed copyright image, watermarks are generated across all test hosts. The absolute values ​​of the collision NC (noise, collision, and index) are calculated pairwise and averaged to obtain the fixed copyright collision NC. For a fixed host image, watermarks are generated across all copyrights, and the average collision NC is calculated. The fixed copyright collision NC is 0.141, and the fixed host collision NC is 0.172.

[0062] Typical attack tests were selected to test watermark matching NC. Some results are as follows: 180° rotation: NC=0.956, variance 0.1 Gaussian noise: NC=0.855, 25% random cropping: NC=0.758.

[0063] Ablation experiments validated the contributions of each module: removing STN reduced the average NC of rotational attacks by approximately 0.02; removing SPP-PAN reduced the NC of clipping and occlusion attacks; removing composite enhancement generally reduced the robustness of various attacks; removing row-column orthogonal loss boosted the collision NC to over 0.97, completely losing discriminative power but improving robust NC to 0.992, validating the trade-off between robustness and discriminative power and the core role of orthogonal loss; removing noise consistency loss caused training collapse, but the collision NC remained above 0.8, validating the necessity of the dual-loss interlocking mechanism.

[0064] By applying a tiny perturbation to a single pixel of the copyrighted image to generate a new scrambling key, and then encrypting the same watermark separately, the bit error rate is calculated. The average BER is approximately 0.505, which is close to the theoretical value of a random sequence. This proves that the scrambling mechanism has extremely strong key sensitivity and can effectively resist forgery attacks.

[0065] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A zero-watermarking model training method based on dual-loss orthogonal decoupling and multi-scale fusion, characterized in that, The method includes the following steps: A zero-watermark generation model is constructed, which includes a multi-scale feature extraction network and an orthogonal feature fusion network. The multi-scale feature extraction network includes a spatial transformation sub-network and a ConvNeXt-Small backbone sub-network. The orthogonal feature fusion network includes a multi-scale spatial pyramid attention sub-network and a residual fusion network. The multi-scale feature extraction network is trained using a dual-channel composite data augmentation strategy to obtain a trained multi-scale feature extraction network. Freeze the parameters of the trained multi-scale feature extraction network, and train the orthogonal feature fusion network using a loss function based on row and column orthogonality loss and noise consistency loss to obtain the trained orthogonal feature fusion network; The trained multi-scale feature extraction network and the trained orthogonal feature fusion network are used as the final zero-watermark generation model.

2. The method according to claim 1, characterized in that, The spatial transformation subnetwork consists of a localization network, a mesh generator, and a bilinear sampler connected in sequence; the localization network is used to regress the affine transformation parameters based on the input image. The mesh generator is used to generate sampling mesh coordinates based on the affine transformation parameters; the bilinear sampler is used to collect pixels from the input image based on the sampling mesh coordinates and output a geometrically corrected image.

3. The method according to claim 1, characterized in that, The dual-channel composite data enhancement strategy includes: (1). Apply weak enhancement processing to the original images in the training dataset to obtain weakly enhanced samples; the weak enhancement processing is to randomly select one attack method from rotation, cropping, salt and pepper noise, Gaussian noise, filtering and JPEG compression to process the original image; (2). Apply strong enhancement processing to the same original image to obtain a strong enhancement sample; the strong enhancement processing is to apply large-angle rotation, extreme proportion random cropping to the original image in sequence, and then superimpose Gaussian noise, salt and pepper noise, median filtering, Gaussian filtering and multiple attacks in JPEG compression; (3) Weakly enhanced samples and strongly enhanced samples of the same original image form positive sample pairs, while enhanced samples of different original images form negative sample pairs.

4. The method according to claim 3, characterized in that, When training the multi-scale feature extraction network using a dual-channel composite data augmentation strategy, the SimCLR self-supervised contrastive learning framework is employed. The total loss function during training is a weighted sum of the NT-Xent contrastive loss and the inverse consistency regularization loss. The NT-Xent contrastive loss is used to maximize the feature similarity of positive sample pairs and minimize the feature similarity of negative sample pairs. The inverse consistency regularization loss is used to constrain the feature consistency of the strongly augmented samples before and after applying additional affine perturbations. The total loss function is expressed as follows: Among them, For NT-Xent contrast loss; This is the loss for inverse consistency regularization; This is the weight coefficient for the regularization term.

5. The method according to claim 1, characterized in that, The multi-scale spatial pyramid attention subnetwork includes a channel unification module, a spatial pyramid pooling module, a squeeze-excitation channel attention module, and a hierarchical attention fusion module. The channel unification module is used to perform 1×1 convolution on the input multi-level feature maps respectively, and unify the channel dimensions of each level of feature maps to the same dimension. The spatial pyramid pooling module is used to perform average pooling on the feature maps of each level after unifying the channel, and then stitch them together with multiple grid sizes to obtain global-local descriptors for each level. The squeeze-excitation channel attention module is used to recalibrate the channel weights of the descriptors at each level; The hierarchical attention fusion module is used to perform weighted fusion of the recalibrated descriptors at each level and output multi-scale fusion features.

6. The method according to claim 5, characterized in that, The residual fusion subnetwork includes a feature stitching unit, a linear projection branch, a multilayer perceptron branch, and a weighted synthesis unit. The feature splicing unit is used to splice the global features of the host image and the global features of the copyright image into a joint feature; The linear projection branch is used to perform a linear transformation on the joint features to obtain a first output; The multilayer perceptron branch is used to perform nonlinear mapping on the joint features to obtain a second output; The weighted synthesis unit is used to perform a weighted summation of the first output and the second output, and to perform L2 normalization on the weighted summation result to obtain the watermark feature vector located on the unit hypersphere.

7. The method according to claim 6, characterized in that, The row-column orthogonality loss is used to constrain the watermark feature vectors generated by pairwise combinations of multiple host images and multiple copyright images in the same training batch to be pairwise orthogonal in both the row and column directions; the row direction of the watermark feature vector is the combination direction of different host images corresponding to the same copyright image, and the column direction of the watermark feature vector is the combination direction of different copyright images corresponding to the same host image; the noise consistency loss is used to constrain the cosine similarity between the features input to the orthogonal feature fusion network and the output watermark feature vectors before and after Gaussian noise addition to be greater than or equal to a preset threshold; wherein, the mathematical expression of the row-column orthogonality loss is: Where B represents the number of host images or copyrighted images in the training batch. The watermark feature vector generated for the i-th host image and the p-th copyright image; The mathematical expression for the noise consistency loss is: in, The watermark vector generated for clean features. The watermark vector generated for the noisy features, The cosine similarity function; The mathematical expression of the loss function based on row-column orthogonality loss and noise consistency loss is as follows: in, This is the weighting coefficient for noise consistency loss.

8. A zero-watermark generation method based on dual-loss orthogonal decoupling and multi-scale fusion, characterized in that, The method includes: Input the target host image and the target copyright image into the zero watermark generation model trained by the method described in any one of claims 1-7 to obtain the watermark feature vector; The watermark feature vector is binarized to output the watermark sequence.

9. A copyright authentication method based on dual-loss orthogonal decoupling and multi-scale fusion, characterized in that, The method is an application of the zero-watermark generation method based on dual-loss orthogonal decoupling and multi-scale fusion as described in claim 8, characterized in that the method includes: Using the method described in claim 8, a first watermark feature vector corresponding to the target host image and the target copyright image is generated. The first watermark feature vector is binarized to obtain a first watermark sequence. Global features of the target copyright image are extracted to generate a scrambling key. The first watermark sequence is scrambled and encrypted using the scrambling key. The encrypted first watermark sequence is bound to the target copyright image and stored in the authentication library. The image to be verified and the image to be confirmed for copyright are input into the zero watermark generation model in the same way and binarized to obtain the second watermark sequence. A scrambling key is generated in the same way and the second watermark sequence is scrambled and encrypted. Calculate the normalized cross-correlation value between the encrypted second watermark sequence and the encrypted first watermark sequence in the authentication library. If the normalized cross-correlation value is greater than or equal to a preset threshold, it is determined that the image to be verified and the image to be confirmed have a copyright ownership relationship.

10. The method according to claim 9, characterized in that, The scrambling key is generated as follows: extract the global feature vector of the copyrighted image, take the weighted sum of the first few dimensions of the global feature vector and round it to generate a random number seed; the scrambling encryption is a one-dimensional Fibonacci affine scrambling based on the random number seed; The one-dimensional Fibonacci affine scrambling involves iteratively performing multiple rounds of affine index rearrangement. Each round generates coefficients and offsets coprime to the length of the watermark sequence. The watermark sequence index is then rearranged after undergoing an affine transformation based on these coefficients and offsets. The affine transformation formula is as follows: in For the original index, For the scrambled index, For the transformation parameters generated in each round, This represents the length of the watermark sequence.

Images