Composite watermark low-coupling embedding and efficient identification method based on wavelet transform

Abstract

The invention discloses a composite watermark low-coupling embedding and efficient identification method based on wavelet transform. The method comprises the following steps: carrying out wavelet transform on a carrier image and dividing sub-band priorities based on visual sensitivity; carrying out differential preprocessing on the copyright watermark, the authentication watermark and the content description watermark; embedding the watermarks into the corresponding sub-bands according to the priorities and the security requirements, and performing noise suppression; generating a watermark-containing image through wavelet inverse transformation; and extracting a verification watermark by adopting a hierarchical cascade identification architecture, and dynamically adjusting a matching threshold value according to an attack type. According to the method, the robustness of the watermark when facing noise, compression and geometric attacks is improved, the watermark embedding capacity is increased, and the recognition accuracy and efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of digital watermarking technology, and particularly relates to a low-coupling embedding and efficient recognition method for composite watermarking based on wavelet transform. Background Technology

[0002] With the rapid development of digital media technology, the dissemination and sharing of multimedia data such as images and videos has become increasingly convenient, but it also faces problems such as copyright infringement, content tampering, and malicious dissemination. Digital watermarking technology, as an effective means of multimedia protection, embeds specific information (watermark) into the carrier data to achieve functions such as confirming copyright ownership and verifying content integrity.

[0003] Existing composite watermarking technologies have the following shortcomings in the frequency domain embedding and recognition process: Insufficient robustness: Most methods use fixed subbands to embed watermarks, without considering the differences in visual sensitivity of subbands in different frequency domains. This makes the watermarks easy to be destroyed when faced with attacks such as noise and compression, resulting in low recognition accuracy. Watermark capacity is limited: Traditional multi-watermark embedding methods do not match watermark type with subband priority, and high-priority subband resources are not used reasonably, resulting in limited capacity when multiple watermarks are embedded at the same time. Low recognition efficiency: The recognition stage often uses a single feature matching algorithm, which requires traversing and detecting all sub-bands, which is time-consuming and difficult to meet real-time requirements. Poor attack adaptability: It lacks a dynamic adjustment mechanism for different attack types, and its robustness in identifying attacks decreases significantly in complex attack scenarios.

[0004] Therefore, designing a composite watermark frequency domain embedding and recognition method with high robustness, large capacity, high efficiency and strong attack adaptability has become a key requirement in the current field of digital watermarking technology. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform, comprising: The carrier image is decomposed into multiple frequency domain sub-bands, including low-frequency and high-frequency sub-bands, by wavelet transform, and the priority of the multiple frequency domain sub-bands is divided based on visual sensitivity. Adaptive preprocessing is performed on the composite watermark, which includes copyright watermark, authentication watermark, and content description watermark; Based on the priority division results and the security requirements of different types of watermarks, different types of watermarks to be embedded are embedded into the corresponding frequency domain sub-bands, and noise suppression is performed on the embedding region during the embedding process. Perform inverse wavelet transform on the frequency domain sub-band after embedding the watermark to generate a carrier image with composite watermark; After performing wavelet transform on the carrier image containing composite watermarks, a hierarchical cascaded recognition architecture is used to extract and verify the composite watermarks sequentially, and the matching threshold is dynamically adjusted according to the attack type during the recognition process.

[0007] Optionally, the priority of multiple frequency domain sub-bands can be divided based on visual sensitivity, including: The visual sensitivity of each frequency domain sub-band is calculated based on an improved visual perception weighting algorithm. The visual perception weight is calculated by weighting the contrast features, edge intensity features, and texture complexity features of the sub-band. According to the visual sensitivity from high to low, the frequency domain sub-bands are divided into high-priority sub-bands, medium-priority sub-bands, and low-priority sub-bands.

[0008] Optionally, adaptive preprocessing may be performed on the composite watermark, including: Error correction encoding is applied to the copyright watermark to add redundancy; hashing is performed on the authentication watermark to generate a fixed-length authentication sequence; and binary encoding is performed on the content description watermark to generate a binary stream.

[0009] Optionally, different types of watermarks to be embedded can be embedded into their corresponding frequency domain subbands, including: The copyright watermark is embedded in a high-priority sub-band, a quantization index modulation algorithm is used, and the quantization step size is adaptively adjusted according to the visual sensitivity of the sub-band. The authentication watermark is embedded in the priority sub-band, the difference expansion algorithm is used, and pixel value overflow is avoided by pixel pair grouping and difference adjustment strategy. The content description watermark is embedded in a low-priority sub-band, and a coefficient replacement algorithm based on discrete cosine transform is used.

[0010] Optionally, noise suppression is performed on the embedding region during the embedding process, including: using Wiener filtering to filter the high-frequency subband used for embedding the watermark.

[0011] Optionally, a hierarchical cascaded recognition architecture is used to sequentially extract and verify composite watermarks, including: The first step is to calculate the percentage difference between the energy of the watermarked subband and the energy of the original subband, and then filter candidate subbands based on the percentage difference in subband energy. The second step involves calculating the normalized correlation coefficient between the extracted copyright watermark and the original copyright watermark, and then verifying the copyright watermark based on the normalized correlation coefficient. The third step involves calculating the Hamming distance between the extracted authentication watermark hash value and the original hash value, and verifying the authentication watermark based on the Hamming distance comparison. The fourth level involves decoding the extracted content description watermark binary stream and verifying the content description watermark based on the text matching rate after decoding.

[0012] Optionally, the matching threshold can be dynamically adjusted based on the attack type during the identification process, including: The noise variance, compression ratio, and geometric deformation parameters of the carrier image containing the composite watermark are extracted and input into a pre-trained support vector machine classification model to identify the attack type; based on the identified attack type, the matching threshold of the corresponding layer in the hierarchical cascaded identification architecture is dynamically adjusted.

[0013] Optionally, when a hybrid attack is identified, a multi-feature fusion matching mechanism is activated, including: Extract edge and texture features from the carrier image containing the composite watermark; construct a multi-feature vector by combining energy features, content features, hash value features, decoding features, edge features, and texture features; calculate the total matching score using a weighted voting mechanism, and determine that the watermark recognition is successful if the total score is greater than or equal to a preset threshold.

[0014] Optionally, before performing wavelet transform on the carrier image, a denoising process is also included. The denoising process uses an adaptive median filtering algorithm, which identifies and replaces noisy pixels by judging the relationship between the extreme values ​​and the median values ​​of pixels within the filtering window. If the relationship cannot be determined, the filtering window is adaptively expanded.

[0015] Optionally, after adaptive preprocessing of the composite watermark, the method further includes embedding synchronization information, including: concatenating the synchronization code to the head of the watermark sequence and embedding it together during embedding; and determining the starting position of the watermark data by matching the synchronization code during extraction.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention addresses the problems of insufficient robustness, limited watermark capacity, and low recognition efficiency in frequency domain embedding of existing composite watermarking technologies. By constructing a frequency domain low-coupling embedding model based on wavelet transform and a hierarchical cascaded recognition architecture, it achieves collaborative embedding and efficient extraction of multiple types of watermarks (copyright watermarks, authentication watermarks, and content description watermarks).

[0017] This invention performs wavelet transform on the carrier image and divides the frequency domain sub-bands into priorities using an improved visual perception weighting algorithm. Then, based on the security and robustness requirements of different watermarks, it allocates the optimal embedding sub-bands and embedding strengths with low coupling. Finally, in the recognition stage, it adopts a hierarchical cascaded matching algorithm, combined with sub-band energy features and a watermark feature library, to achieve rapid positioning and accurate recognition of composite watermarks.

[0018] This invention achieves high accuracy in recognizing composite watermarks when facing noise attacks, compression attacks, and geometric attacks, and its recognition speed is significantly improved compared to traditional methods. It can be widely used in image copyright protection, content authentication, and multimedia source tracing scenarios. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the low-coupling embedding and efficient recognition method of various watermarks in the wavelet transform frequency domain according to embodiments of the present invention. Figure 2 This is a flowchart of the carrier image preprocessing process according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform, including: The carrier image is decomposed into multiple frequency domain sub-bands, including low-frequency and high-frequency sub-bands, by wavelet transform, and the priority of the multiple frequency domain sub-bands is divided based on visual sensitivity. Adaptive preprocessing is performed on the composite watermark, which includes copyright watermark, authentication watermark, and content description watermark; Based on the priority division results and the security requirements of different types of watermarks, different types of watermarks to be embedded are embedded into the corresponding frequency domain sub-bands, and noise suppression is performed on the embedding region during the embedding process. The noise suppression of the embedding region during the embedding process includes: using Wiener filtering to filter the high-frequency sub-band used for embedding watermarks.

[0023] Perform inverse wavelet transform on the frequency domain sub-band after embedding the watermark to generate a carrier image with composite watermark; After performing wavelet transform on the carrier image containing composite watermarks, a hierarchical cascaded recognition architecture is used to extract and verify the composite watermarks sequentially, and the matching threshold is dynamically adjusted according to the attack type during the recognition process.

[0024] The input carrier image undergoes grayscale normalization and denoising. A two-dimensional discrete wavelet transform (2D-DWT) is used to decompose the carrier image into low-frequency sub-bands (LL) and high-frequency sub-bands (LH, HL, HH), constructing a frequency domain sub-band set. The denoising process employs an adaptive median filtering algorithm, which identifies and replaces noisy pixels by judging the relationship between the extreme values ​​and median values ​​of pixels within the filtering window. If the relationship cannot be determined, the filtering window is adaptively expanded. Figure 2 As shown.

[0025] The low-coupling median filtering algorithm is adopted, and the specific process is as follows: Initialize the filter window size (initial window is 3×3, maximum window is 7×7); calculate the median, minimum, and maximum values ​​of the pixels within the window; If Min < Med < Max, then determine whether the center pixel of the window is noise (if the center pixel value is < Min or > Max, it is determined to be noise and replaced with Med; otherwise, the original pixel is retained); if Med = Min or Med = Max, then expand the window to the next level and repeat the above steps until the window reaches the maximum size or the noise reduction is completed. The noise suppression ratio (NSR) of the denoised image is ≥20dB, ensuring the accuracy of subband feature extraction in subsequent wavelet transform.

[0026] Specifically, the priority of multiple frequency domain sub-bands is divided based on visual sensitivity, including: The visual sensitivity of each frequency domain sub-band is calculated based on an improved visual perception weighting algorithm. The visual perception weight is calculated by weighting the contrast features, edge intensity features, and texture complexity features of the sub-band. According to the visual sensitivity from high to low, the frequency domain sub-bands are divided into high-priority sub-bands, medium-priority sub-bands, and low-priority sub-bands.

[0027] Furthermore, based on the improved Visual Perception Weight (VPW) algorithm, the visual sensitivity of each frequency domain sub-band is calculated, and the sub-bands are divided into high-priority sub-bands (LL, LH1), medium-priority sub-bands (HL1, HH1), and low-priority sub-bands (LH2, HL2, HH2) according to the sensitivity from high to low. The improved Visual Perception Weights (VPW) algorithm is specifically calculated using the following formula: ; Wherein, Contrast(i,j) is the contrast feature of sub-band (i,j), calculated by the standard deviation of pixel gray values ​​within the sub-band; Edge(i,j) is the edge intensity feature, obtained by edge detection of each sub-band using the Sobel operator; Texture(i,j) is the texture complexity feature, obtained by calculating the entropy value of the sub-band using the gray-level co-occurrence matrix (GLCM); α, β, and γ are weight coefficients, and α+β+γ=1, which are optimized iteratively using a genetic algorithm (with 100 iterations and a population size of 50) to obtain the optimal values ​​(α=0.45, β=0.35, γ=0.2).

[0028] Specifically, adaptive preprocessing is performed on composite watermarks, including: Error correction encoding is applied to the copyright watermark to add redundancy; hashing is performed on the authentication watermark to generate a fixed-length authentication sequence; and binary encoding is performed on the content description watermark to generate a binary stream.

[0029] Furthermore, the copyright watermark (binary sequence), authentication watermark (hash value sequence), and content description watermark (text encoded sequence) are preprocessed respectively. The copyright watermark is added with error correction redundancy through RS encoding, the authentication watermark is generated with a fixed-length hash value through the SHA-256 algorithm, and the content description watermark is converted into a binary stream through Base64 encoding.

[0030] After adaptive preprocessing of the composite watermark, the process also includes embedding synchronization information, including: concatenating the synchronization code to the head of the watermark sequence and embedding it during the embedding process; and determining the starting position of the watermark data by matching the synchronization code during extraction.

[0031] Furthermore, the specific method is as follows: Generate a 16-bit binary synchronization code (generated using a pseudo-random sequence with a sequence period of 1024), and concatenate the synchronization code with the headers of the copyright watermark, authentication watermark, and content description watermark to form a watermark sequence with synchronization information. During the embedding phase, the synchronization code is embedded together with the corresponding watermark into the specified sub-band (the copyright watermark synchronization code is embedded into the LL sub-band, the authentication watermark synchronization code is embedded into the HL1 sub-band, and the content description watermark synchronization code is embedded into the LH2 sub-band). During the extraction stage, the starting position of the watermark is first located by matching the synchronization code (the correlation between the extracted synchronization code and the original synchronization code is calculated, and the synchronization is considered successful when the correlation is ≥0.9). Then, the subsequent watermark data is extracted to avoid watermark extraction failure caused by factor band offset. The synchronization success rate is no less than 99%.

[0032] Specifically, different types of watermarks to be embedded are embedded into their corresponding frequency domain sub-bands, including: The copyright watermark is embedded in a high-priority sub-band, a quantization index modulation algorithm is used, and the quantization step size is adaptively adjusted according to the visual sensitivity of the sub-band. The authentication watermark is embedded in the priority sub-band, the difference expansion algorithm is used, and pixel value overflow is avoided by pixel pair grouping and difference adjustment strategy. The content description watermark is embedded in a low-priority sub-band, and a coefficient replacement algorithm based on discrete cosine transform is used.

[0033] Furthermore, based on the security requirements of composite watermarking, the copyright watermark is embedded in the high-priority sub-band (using the improved Quantization Index Modulation (QIM) algorithm to adjust the quantization step size with low coupling), the authentication watermark is embedded in the medium-priority sub-band (using the Differential Expansion (DE) algorithm to avoid pixel overflow), and the content description watermark is embedded in the low-priority sub-band (based on the coefficient replacement algorithm of Discrete Cosine Transform (DCT) to control embedding distortion).

[0034] The improved Quantization Indexed Modulation (QIM) algorithm adjusts the quantization step size with low coupling in the following way: Wherein, Δ0 is the initial quantization step size (range 5-10, dynamically set according to the gray level of the carrier image; for a 256-level gray image, Δ0=8 by default); k is the adjustment coefficient (range 1-3, determined through cross-validation; 1.5 is used when the carrier image texture complexity is high, and 2.5 is used when the texture is simple); VPW(i,j) is the visual perception weight of subband (i,j). The higher the visual sensitivity of the subband, the smaller the quantization step size, and the embedding distortion is controlled within the range of peak signal-to-noise ratio (PSNR) ≥ 35dB.

[0035] When using the Differential Expansion (DE) algorithm for authentication watermark embedding, a pixel pair grouping and difference adjustment strategy is employed to avoid pixel overflow. The specific steps are as follows: The pixels of the medium priority sub-bands (HL1, HH1) are divided into several pixel pairs (x,y) in 2×2 blocks; Calculate the mean of pixel pairs and difference ; Based on the 1-bit binary information 'b' of the authentication watermark, adjust the difference to... ; Calculate the adjusted pixel value ; If x′ or y′ exceeds the pixel grayscale range (0-255), the pixel pair is marked as "cannot be embedded", the position is skipped and an adjacent pixel pair is selected for recalculation, ensuring that there is no overflow of subband pixels after embedding and that the effective utilization rate of pixel pairs is ≥95%.

[0036] Specifically, a hierarchical cascaded recognition architecture is used to sequentially extract and verify composite watermarks, including: The first step is to calculate the percentage difference between the energy of the watermarked subband and the energy of the original subband, and then filter candidate subbands based on the percentage difference in subband energy. The second step involves calculating the normalized correlation coefficient between the extracted copyright watermark and the original copyright watermark, and then verifying the copyright watermark based on the normalized correlation coefficient. The third step involves calculating the Hamming distance between the extracted authentication watermark hash value and the original hash value, and verifying the authentication watermark based on the Hamming distance comparison. The fourth level involves decoding the extracted content description watermark binary stream and verifying the content description watermark based on the text matching rate after decoding.

[0037] Furthermore, the watermarked image is first subjected to inverse wavelet transform, and then the watermark features are extracted layer by layer according to subband priority; a hierarchical cascaded recognition architecture is constructed. The first level filters candidate watermark subbands through subband energy matching, the second level verifies the copyright watermark through correlation matching (NC), the third level confirms the authentication watermark through hash value comparison, and the fourth level extracts the content description watermark through encoding and decoding matching. The judgment conditions and execution logic for each level of the hierarchical cascaded recognition architecture are as follows: Level 1 (Subband Energy Matching): Calculate the percentage difference between the watermarked subband energy Ew and the original subband energy Eo. If the value is less than the threshold θ1 (θ1=5%), it is determined to be a candidate subband, and subbands with excessive energy difference are excluded. Level 2 (Relevance Matching): Calculate the correlation coefficient between the extracted copyright watermark and the original copyright watermark. (W) o For the original watermark, W e To extract the watermark, W o W e (These are the corresponding mean values). When NC ≥ θ2 (θ2 = 0.85), the copyright watermark is considered to be successfully matched. Level 3 (Hash Value Comparison): Calculate the Hamming distance between the extracted authentication hash value and the original authentication hash value. (Where ho is the original hash value, he is the extracted hash value, and n is the hash value length), when HD≤3, the authentication watermark verification is successful, and it is determined that the carrier content has not been tampered with; Level 4 (Encoding / Decoding Matching): Base64 decoding is performed on the extracted content description watermark binary stream. The character matching rate between the decoded text and the original text is compared. When the matching rate is greater than 98%, the content description watermark extraction is deemed valid.

[0038] Specifically, during the identification process, the matching threshold is dynamically adjusted based on the attack type, including: The noise variance, compression ratio, and geometric deformation parameters of the carrier image containing the composite watermark are extracted and input into a pre-trained support vector machine classification model to identify the attack type; based on the identified attack type, the matching threshold of the corresponding layer in the hierarchical cascaded identification architecture is dynamically adjusted.

[0039] Specifically, when a hybrid attack is identified, a multi-feature fusion matching mechanism is activated, including: Extract edge and texture features from the carrier image containing the composite watermark; construct a multi-feature vector by combining energy features, content features, hash value features, decoding features, edge features, and texture features; calculate the total matching score using a weighted voting mechanism, and determine that the watermark recognition is successful if the total score is greater than or equal to a preset threshold.

[0040] Furthermore, the execution process for attack type determination includes: Feature extraction: Extract the noise variance of the watermarked image (obtained by calculating the mean square error between the image grayscale value and the smoothed image grayscale value), compression ratio (obtained by analyzing the ratio of the quantization table of the JPEG image to the original image data size), and geometric deformation parameters (calculated by feature point matching algorithm and detected by Hough transform). Type classification: Input the extracted features into a pre-trained support vector machine (SVM) classification model (kernel function uses radial basis function (RBF), penalty coefficient C=10, gamma parameter=0.1), and output the attack type (noise attack, compression attack, geometric attack or hybrid attack). Threshold adjustment: The matching thresholds at each level are dynamically adjusted according to the attack type. For example, the correlation matching threshold θ2 is reduced to 0.8 for noise attacks, θ2 is kept at 0.85 for compression attacks, and the subband energy matching threshold θ1 is relaxed to 8% for geometric attacks. The attack type identification accuracy of this module is no less than 92%.

[0041] Furthermore, the watermark extraction and recognition also includes an abnormal attack handling mechanism. When the attack type judgment module identifies a mixed attack (two or more attack types exist simultaneously), the following operations are performed: Initiate multi-feature fusion matching: In addition to subband energy, correlation, hash value, and encoding / decoding features, additional edge features (detected by the Canny operator) and texture features (calculated by Local Binary Pattern (LBP)) of the watermarked image are extracted to construct a multi-feature vector; A weighted voting mechanism is adopted: weights are assigned to the matching results of each feature (energy matching weight 0.2, correlation matching weight 0.3, hash value comparison weight 0.25, encoding and decoding matching weight 0.15, and edge and texture feature matching weights 0.05 each), and the total matching score is calculated. Set a matching threshold for hybrid attacks (total score ≥ 0.8). When the total score reaches the threshold, the watermark recognition is considered successful. This mechanism can improve the accuracy of composite watermark recognition under hybrid attacks by 5%-8%.

[0042] The overall process of the method of the present invention includes six core steps: carrier image preprocessing, composite watermark preprocessing, frequency domain sub-band priority division, low-coupling watermark embedding, watermark extraction and recognition, and robustness optimization. Each step works together to achieve efficient embedding and reliable recognition of composite watermarks.

[0043] Key technical details: (1) Frequency domain sub-band priority division; Based on an improved Visual Perception Weighting (VPW) algorithm, this method comprehensively considers subband contrast, edge strength, and texture complexity to achieve accurate subband priority division. Contrast features are calculated using the standard deviation of subband pixel grayscale values, edge strength features are detected using the Sobel operator, and texture complexity features are measured using the entropy value of the Gray-Level Co-occurrence Matrix (GLCM). A genetic algorithm is used to optimize the weight coefficients, ensuring a high degree of matching between the priority division and the characteristics of the Human Visual System (HVS), providing a basis for subsequent loosely coupled embedding.

[0044] (2) Low-coupling watermark embedding model; To address the varying needs of different watermark types, a differentiated embedding strategy was designed: Copyright watermark: As core protected information, it needs to be highly robust, therefore high-priority sub-bands (LL, LH1) are embedded. An improved QIM algorithm is adopted to dynamically adjust the quantization step size according to the visual sensitivity of the sub-band, so as to minimize embedding distortion while ensuring robustness. Authentication watermark: Used for content integrity verification, it needs to have fast detection capabilities and is embedded in a high-priority sub-band (HL1, HH1). It employs the difference expansion (DE) algorithm, which expands the watermark by the difference between adjacent pixels to avoid pixel overflow while ensuring embedding capacity. Content description watermark: Used to record metadata such as image capture time and location, embedded in low-priority sub-bands (LH2, HL2, HH2). Based on the DCT transform coefficient replacement algorithm, high-frequency coefficients with minimal visual impact are selected for replacement, controlling embedding distortion within an acceptable range (peak signal-to-noise ratio PSNR ≥ 35dB).

[0045] (3) Hierarchical cascaded identification architecture; To improve recognition efficiency, a four-level hierarchical cascaded recognition architecture is designed: Level 1 (Subband Energy Matching): Quickly screens candidate subbands, excludes subbands whose energy is severely distorted due to attacks, and reduces the scope of subsequent detection; Level 2 (Relevance Matching): For copyright watermarks, calculate the correlation coefficient (NC) between the extracted watermark and the original watermark to verify copyright ownership; Level 3 (Hash value comparison): For authentication watermarks, the hash value is extracted by comparing the Hamming distance with the original hash value to determine whether the content has been tampered with; Level 4 (Encoding / Decoding Matching): For content description watermarks, valid metadata is extracted through Base64 decoding and text integrity verification.

[0046] This architecture significantly reduces identification time and improves real-time performance through a hierarchical strategy of "coarse screening - fine matching".

[0047] (4) Robust optimization mechanism; Embedding stage: A low-coupling filter module is introduced, and Wiener filtering is used to suppress high-frequency subband noise, thereby improving the stability of watermark embedding; Identification phase: An attack type judgment module is added to identify attack types based on the SVM classification model and dynamically adjust the feature matching threshold (e.g., reduce the NC threshold to 0.8 for noise attacks and keep the threshold at 0.85 for compression attacks) to ensure robust identification under complex attacks.

[0048] Experimental verification: To verify the effectiveness of the method of the present invention, the following experiments were conducted: Experimental environment: PYCHARM2022, carrier images used Lena, Peppers, Baboon (512×512, 256 grayscale levels), watermarks include 8×8 binary copyright watermark, 256-bit authentication hash value, and 64-byte content description watermark; Attack types: Gaussian noise (variance 0.01), JPEG compression (quality factor 50), rotation attack (10°), scaling attack (0.8x); Evaluation metrics: watermark recognition accuracy (Acc), peak signal-to-noise ratio (PSNR), and recognition time (T).

[0049] The experimental results are shown in Table 1 below: Table 1 Example 1: The carrier image is Lena (512×512); Carrier image preprocessing: The Lena image is grayscale normalized (pixel values ​​are mapped to [0,1]), and 3-level 2D-DWT decomposition is used to obtain LL3 (low frequency), LH1-LH3, HL1-HL3, and HH1-HH3 (high frequency) subbands; Composite watermark preprocessing: The copyright watermark is an 8×8 binary sequence (10101100...), which is extended to 128 bits by RS encoding (bitrate 1 / 2); the authentication watermark generates a 256-bit hash value for the Lena image using SHA-256; the content description watermark is "Shooting time: 2024-05-20, Location: Laboratory", which is converted to a 128-bit binary stream by Base64 encoding; Frequency domain sub-band priority division: The weight of each sub-band is calculated by the improved VPW algorithm. LL3 (VPW=0.92) and LH1 (VPW=0.88) are high priority, HL1 (VPW=0.75) and HH1 (VPW=0.70) are medium priority, and the rest are low priority. Low-coupling watermark embedding: Copyright watermark is embedded in LL3 subband with quantization step size Δ=2; Authentication watermark is embedded in HL1 subband with DE algorithm to expand the difference to 2; Content description watermark is embedded in LH2 subband, replacing the high-frequency coefficients of DCT. Watermark Extraction and Recognition: A 3-level IDWT is performed on the watermarked image. The first level filters subbands with an energy difference of <5% (LL3, LH1, HL1, LH2). The second level calculates the copyright watermark NC=0.92≥0.85, indicating a successful match. The third level authenticates the watermark with a Hamming distance HD=1≤3, indicating successful verification. The fourth level decodes the content description watermark, ensuring the text is complete and the extraction is effective. Robustness test: Gaussian noise (variance 0.01) was added to the watermarked image. The attack type judgment module identified it as a noise attack. The NC threshold was adjusted to 0.8. Finally, the copyright watermark Acc=96.3%, the authentication watermark Acc=95.8%, and the content description watermark Acc=95.2%.

[0050] Example 2: The carrier image is a Baboon (512×512, complex texture image); (1) Carrier image preprocessing; Baboon images have complex textures (hair areas account for over 60%), so noise reduction needs to be enhanced during the preprocessing stage. Gray-level normalization: Maps pixel values ​​from 0,255 to 0,1 to avoid excessive fluctuations in gray-level values ​​in texture areas that could affect subband decomposition; Low-coupling median filtering: The initial window is 3×3. When a dense hair area is detected (texture complexity > 0.8, judged by GLCM entropy value), the window is automatically expanded to 5×5. The noise suppression ratio (NSR) after denoising reaches 22dB. 2D-DWT decomposition: A 3-level decomposition is adopted to obtain the following sub-band sets: LL3 (low frequency, preserving texture contours), LH1-LH3 (horizontal details, including hair edges), HL1-HL3 (vertical details, including hair texture), and HH1-HH3 (diagonal details, including texture noise).

[0051] (2) Composite watermark preprocessing; To address the vulnerability of complex texture images to attacks, the watermark redundancy design has been optimized: Copyright watermark: A 16×16 binary sequence (including the copyright owner's identifier), extended to 512 bits by RS encoding (code rate 2 / 3) to improve error correction capability; Authentication watermark: A 256-bit hash value is generated from the grayscale matrix of the Baboon image using SHA-256, and an additional 8-bit CRC check code is added to enhance the accuracy of tamper detection; The content description watermark reads: "Shooting device: Canon EOS R6, resolution: 512×512, shooting scene: animal specimen", which is then converted to a 256-bit binary stream using Base64 encoding to increase the amount of metadata information.

[0052] (3) Frequency domain sub-band priority division; The sub-band weights were calculated using the improved VPW algorithm, and the results are shown in Table 2. Table 2 (4) Low-coupling watermark embedding; Adjust the embedding parameters according to the sub-band characteristics of the texture region: Copyright watermark: Embedded LL3 subband, improved QIM algorithm, initial quantization step size Δ0=6, adjustment coefficient k=2.2 (because LL3 subband VPW=0.95, calculated as follows). ), control embedding distortion; Authentication watermark: Embedded in HL1 subband, the DE algorithm uses 3×3 pixel block grouping (normally 2×2) to reduce the probability of pixel pair overflow in texture areas, with an effective embedding rate of 96%; Content description watermark: Embedded LH2 subband, after DCT transformation, select the 5th-8th order high frequency coefficients to replace (usually the 3rd-6th order) to avoid affecting texture details, PSNR=34.8dB after embedding.

[0053] (5) Watermark extraction and recognition (JPEG compression attack, quality factor 40); Subband energy matching: Screen subbands with an energy difference of <5% (LL3, LH1, HL1, LH2), and exclude HH2 (energy loss of 12%). Relevance matching: Copyright watermark NC=0.88≥0.85 (threshold not adjusted due to compression attack), match successful; Hash value comparison: The Hamming distance between the extracted authentication watermark hash value and the original value is HD=2≤3, the CRC check passes, and the content is determined to be tampered with; Encoding / decoding matching: The text matching rate after decoding the content description watermark is 99.2%, and the extraction is effective; Recognition time: 15.3ms, a 43% improvement over the traditional method (26.8ms).

[0054] It should be noted that in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0055] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform, characterized in that, Includes the following steps: The carrier image is decomposed into multiple frequency domain sub-bands, including low-frequency and high-frequency sub-bands, by wavelet transform, and the priority of the multiple frequency domain sub-bands is divided based on visual sensitivity. Adaptive preprocessing is performed on the composite watermark, which includes copyright watermark, authentication watermark, and content description watermark; Based on the priority division results and the security requirements of different types of watermarks, different types of watermarks to be embedded are embedded into the corresponding frequency domain sub-bands, and noise suppression is performed on the embedding region during the embedding process. Perform inverse wavelet transform on the frequency domain sub-band after embedding the watermark to generate a carrier image with composite watermark; After performing wavelet transform on the carrier image containing the composite watermark, a hierarchical cascaded recognition architecture is used to extract and verify the composite watermark sequentially, and the matching threshold is dynamically adjusted according to the attack type during the recognition process.

2. The wavelet transform-based composite watermarking low-coupling embedding and efficient recognition method according to claim 1, characterized in that, Prioritizing multiple frequency domain sub-bands based on visual sensitivity includes: The visual sensitivity of each frequency domain sub-band is calculated based on an improved visual perception weighting algorithm. The visual perception weight is calculated by weighting the contrast features, edge intensity features, and texture complexity features of the sub-band. According to the visual sensitivity from high to low, the frequency domain sub-bands are divided into high-priority sub-bands, medium-priority sub-bands, and low-priority sub-bands.

3. The low-coupling embedding and efficient recognition method for composite watermarking based on wavelet transform according to claim 1, characterized in that, Adaptive preprocessing of composite watermarks includes: Error correction encoding is applied to the copyright watermark to add redundancy; hashing is performed on the authentication watermark to generate a fixed-length authentication sequence; and binary encoding is performed on the content description watermark to generate a binary stream.

4. The wavelet transform-based composite watermarking low-coupling embedding and efficient recognition method according to claim 2, characterized in that, Different types of watermarks to be embedded are embedded into their corresponding frequency domain sub-bands, including: The copyright watermark is embedded in a high-priority sub-band, a quantization index modulation algorithm is used, and the quantization step size is adaptively adjusted according to the visual sensitivity of the sub-band. The authentication watermark is embedded in the priority sub-band, the difference expansion algorithm is used, and pixel value overflow is avoided by pixel pair grouping and difference adjustment strategy. The content description watermark is embedded in a low-priority sub-band, and a coefficient replacement algorithm based on discrete cosine transform is used.

5. The low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform according to claim 1, characterized in that, Noise suppression is performed on the embedding region during the embedding process, including: using Wiener filtering to filter the high-frequency subband used for embedding the watermark.

6. The low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform according to claim 1, characterized in that, A hierarchical cascaded recognition architecture is used to sequentially extract and verify composite watermarks, including: The first step is to calculate the percentage difference between the energy of the watermarked subband and the energy of the original subband, and then filter candidate subbands based on the percentage difference in subband energy. The second step involves calculating the normalized correlation coefficient between the extracted copyright watermark and the original copyright watermark, and then verifying the copyright watermark based on the normalized correlation coefficient. The third step involves calculating the Hamming distance between the extracted authentication watermark hash value and the original hash value, and verifying the authentication watermark based on the Hamming distance comparison. The fourth level involves decoding the extracted content description watermark binary stream and verifying the content description watermark based on the text matching rate after decoding.

7. The low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform according to claim 6, characterized in that, The matching threshold is dynamically adjusted based on the attack type during the identification process, including: The noise variance, compression ratio, and geometric deformation parameters of the carrier image containing the composite watermark are extracted and input into a pre-trained support vector machine classification model to identify the attack type; based on the identified attack type, the matching threshold of the corresponding layer in the hierarchical cascaded identification architecture is dynamically adjusted.

8. The low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform according to claim 7, characterized in that, When a hybrid attack is identified, a multi-feature fusion matching mechanism is activated, including: Extract edge and texture features from the carrier image containing the composite watermark; construct a multi-feature vector by combining energy features, content features, hash value features, decoding features, edge features, and texture features; calculate the total matching score using a weighted voting mechanism, and determine that the watermark recognition is successful if the total score is greater than or equal to a preset threshold.

9. The low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform according to claim 1, characterized in that, Before performing wavelet transform on the carrier image, a denoising process is also included. The denoising process uses an adaptive median filtering algorithm, which identifies and replaces noisy pixels by judging the relationship between the extreme values ​​and the median values ​​of pixels within the filtering window. If it cannot be determined, the filtering window is adaptively expanded.

10. The low-coupling embedding and efficient identification method for composite watermarking based on wavelet transform according to claim 3, characterized in that, After adaptive preprocessing of the composite watermark, the process also includes embedding synchronization information, including: concatenating the synchronization code to the head of the watermark sequence and embedding it during the embedding process; and determining the starting position of the watermark data by matching the synchronization code during extraction.

Images