Picture tracing method based on blind watermark
By embedding encrypted watermark information into the frequency coefficients of the Y channel of an image, the problem of traditional tracing methods being susceptible to damage due to slight image changes and explicit watermarks is solved, achieving stable and secure tracing in complex scenarios.
Patent Information
- Application Number
- CN202510999180.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-21
- Publication Date
- 2025-11-07
AI Technical Summary
Existing technologies struggle to effectively trace the source of images during their dissemination, especially when image content is slightly altered, visible watermarks are easily damaged, or metadata is easily lost. Traditional tracing methods fail in these situations, failing to guarantee the stability and security of the tracing process.
A blind watermarking-based method is adopted, which embeds the watermark information into the intermediate frequency coefficients of the DCT in the Y channel of the image, uses AES-128 encryption and PBKDF2 key derivation function to generate unique identification information, and combines 64×64 block division and DCT transformation to achieve the concealment and stability of the watermark, ensuring accurate extraction even when the image is slightly modified.
It achieves resistance to minor changes, is difficult to remove or destroy, and can be traced without the original image, ensuring the security and accuracy of traceability information and adapting to complex dissemination scenarios.
Smart Images
Figure CN120912411A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of digital image processing, and particularly relates to a picture tracing method based on blind watermarking. BACKGROUND
[0002] In the current Internet environment, large-scale dissemination of picture content is increasingly frequent. With the development of social media, content platforms and instant messaging tools, picture information is extremely easy to be downloaded, forwarded and even tampered with, leading to difficulties in copyright protection and difficulties in monitoring illegal dissemination. Traditional picture tracing methods include file fingerprints, metadata, and explicit watermarking.
[0003] Traditional tracing methods such as file fingerprint-based methods cannot cope with slight changes in image content, such as re-encoding, resizing or light compression, which will change the file fingerprint and make it impossible to locate the original source of publication. The metadata method relies on additional information, but when transmitted on social media platforms, the information is usually automatically stripped, resulting in information loss. Explicit watermarking is visually noticeable and easy to remove through cropping, blocking, AI repair and other means, making it difficult to ensure its stability and effectiveness in complex dissemination scenarios. Therefore, these methods all have the common defect of being easily damaged and difficult to track the image dissemination path for a long time. SUMMARY
[0004] In view of the above-mentioned shortcomings of the prior art, the present application provides a picture tracing method based on blind watermarking.
[0005] To achieve the above-mentioned purposes, the present application is realized by the following technical solutions: The present application provides a picture tracing method based on blind watermarking, comprising a watermark embedding step and a watermark extraction step. The watermark embedding step comprises: Image preprocessing: OpenCV is used to read the image, and the image is converted to a 512x512 size image; Watermark information generation: the unique identification information composed of platform ID, user ID and upload timestamp is formed by string concatenation, and is processed by uniform encoding format, and then encrypted using AES-128 encryption method combined with fixed initialization vector and key derivation function; The encrypted watermark information is converted to a binary bit stream by a string to binary stream function, and then filled into a 64x64 0,1 two-dimensional array, with the filling logic being filled row by row, with 0 added if the bits are insufficient, and truncated if the bits are excessive; The pixel values of the Y channel of the preprocessed image are converted to a MAT object using OpenCV; The Y channel is divided into 64x64 blocks by a MAT object, two medium frequency positions in the DCT result of each 8x8 block are selected, the coefficient values are compared, and whether to exchange or fine tune the relationship of the two is determined according to the corresponding watermark bit; IDCT transformation is performed on the blocks to restore the image; The image embedded with the watermark information is saved; The watermark extraction step comprises: The Y channel pixel values of the image containing the watermark are loaded as a MAT object by using OpenCV; The Y channel pixel values are divided into 64x64 blocks by a MAT object; DCT transformation is performed on the blocks; The embedded position of each block is located, the corresponding DCT coefficient is read, and the watermark bit is recovered according to the size relationship. If A>B, the bit is recorded as 1, otherwise as 0. Finally, the entire watermark bit stream is reconstructed and filled into a 64x64 two-dimensional array; The watermark binary information of the two-dimensional array is converted into a binary stream, and then into encrypted text information. The original watermark information is decrypted by AES.
[0006] Preferably, in the watermark embedding step 5), the medium frequency positions are (2,3) and (3,2).
[0007] Preferably, in the watermark embedding step 5), the relationship of the two is determined according to the corresponding watermark bit, specifically: If the watermark bit to be embedded is 1, A>B and the difference is greater than or equal to delta are required: If A is already greater than B and A-B is greater than or equal to delta, no modification is made; If A is greater than B but the difference is less than delta, fine tuning is performed as A=B+delta; If A≤B, A_new=B+delta / 2 and B_new=B-delta / 2 are set; If the watermark bit to be embedded is 0, A<B and the difference is greater than or equal to delta are required: If A is already less than B and B-A is greater than or equal to delta, no modification is made; If A is less than B but the difference is less than delta, fine tuning is performed as A=B-delta; If A≥B, exchange and fine tuning are performed as A_new=B-delta / 2 and B_new=B+delta / 2; Wherein, A represents the coefficient value of position (2,3) in the DCT block, B represents the coefficient value of position (3,2) in the DCT block, and delta is the minimum embedding strength.
[0008] Preferably, the value of the delta is between 3 and 10.
[0009] Preferably, in the watermark information generation step, the key derivation function is PBKDF2.
[0010] Preferably, in the watermark extraction step 4), since the watermark information is uniformly distributed by blocking, even if the picture is cut and partially blocked, the approximate information of the watermark can be recovered; at the same time, by extracting the watermark information from the medium frequency position of the DCT result, even if the picture is compressed, the watermark information can be extracted.
[0011] Preferably, the watermark information is embedded by AES encryption, so even if the watermark information is illegally extracted, it cannot be decrypted.
[0012] Preferably, in the watermark embedding process, by controlling the embedding strength factor, the balance between robustness and visual effect is optimized, which can ensure the stability of the watermark while avoiding image edge jumping.
[0013] Compared with the known prior art, the technical scheme provided by the present application has the following beneficial effects: I. Stronger resistance to slight content changes: This method embeds the watermark in the medium frequency coefficients of the DCT of the image Y channel, and controls the embedding strength delta, so that the watermark is deeply integrated with the image content. Even if the image is slightly processed such as re-encoding, resizing, light compression, etc., the relative relationship of the medium frequency coefficients (A>B or A<B) can still be accurately identified, effectively avoiding the problems of invalid file fingerprint and easy loss of metadata in traditional methods, and the watermark can still be stably extracted after the image is slightly changed, ensuring the continuity of the traceability.
[0014] II. Watermark is difficult to remove or destroy: This method realizes the advantage that the watermark is difficult to remove or destroy by virtue of the concealment and distribution design of the watermark. The watermark is embedded in the frequency domain of the image and is invisible, avoiding the problem that the explicit watermark is easy to be perceived and removed; at the same time, the watermark information is uniformly distributed in the image blocks in the form of 64x64 binary array, and each block corresponds to an independent watermark bit. Even if part of the image is cut or blocked, the undamaged blocks can still provide effective information, overcoming the defects of traditional methods that explicit watermark is easy to be destroyed and metadata can be actively deleted.
[0015] III. Blind extraction can be realized without original drawing: The method belongs to blind watermarking technology. In the embedding stage, through standardized procedures (such as uniforming the image to 512*512 size, using 64*64 block, fixed DCT medium frequency embedding position, etc.), the watermark embedding position and rule have uniqueness; in the extraction, only through the same block, DCT transformation and medium frequency coefficient comparison, the watermark bit stream can be directly recovered, without relying on the original image as a reference, solving the problem that the traditional non-blind watermarking method cannot trace the source without the original image, and adapting to complex transmission scenarios.
[0016] IV. The traceability information is more secure: The watermark information of the method is composed of platform ID, user ID and upload timestamp, which can uniquely correspond to the source, subject and time of the image, ensuring the accuracy of traceability; at the same time, AES-128 encryption is combined with PBKDF2 key derivation and fixed initialization vector, and the anti-brute force cracking ability is enhanced through multiple iterations, even if the watermark is illegally extracted, it is also difficult to decrypt its content, avoiding the problem that the metadata or explicit watermark information in the traditional method is easy to be tampered with, forged or leaked, and ensuring the security of the traceability information. BRIEF DESCRIPTION OF DRAWINGS
[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, below will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.
[0018] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, below will combine the drawings in the embodiments of the present application to clearly and completely describe the technical scheme in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0020] The present application will be further described below in combination with embodiments.
[0021] Embodiment: Reference Figure 1 A picture traceability method based on blind watermarking, comprising a watermark embedding step and a watermark extraction step; The watermark embedding step comprises: Image preprocessing: OpenCV is used to read the image, and the image is converted to a 512x512 size image. The size selection can ensure the uniformity of the subsequent block, and lay the foundation for the standardized process of watermark embedding; Watermark information generation: The unique identification information composed of platform ID, user ID and upload timestamp is formed by string splicing. This identification information can uniquely correspond to the upload subject, source platform and time dimension of the picture. After processing by a unified encoding format (such as UTF-8), it is encrypted using the AES-128 encryption method, combined with a fixed initialization vector (IV) and a key derivation function PBKDF2. PBKDF2 enhances the key resistance to brute force cracking through multiple iterations, and the fixed IV ensures the consistency of the encryption process. The encrypted watermark information is converted to a binary bit stream through a string-to-binary stream function, and then filled into a 64x64 0,1 two-dimensional array. The filling logic is to fill in each row, and if the bits are insufficient, 0 is added, and if the bits are too many, they are truncated. This operation makes the watermark information structure accurately correspond to the image block structure, ensuring that each block embeds independent watermark bits. The preprocessed image is converted from the RGB color space to the YCrCb color space using OpenCV, and the Y channel pixel values are extracted and converted to a MAT object. The Y channel is selected because it carries image brightness information, has less impact on visual perception, is suitable for embedding watermarks, and is not easily detected. The Y channel is divided into 64x64 blocks through the MAT object, and each block is 8x8 pixels. DCT (Discrete Cosine Transform) is performed on each 8x8 block to convert the pixel values from the spatial domain to the frequency domain. In the DCT result, select the (2,3) and (3,2) medium frequency positions. This position avoids both the low-frequency coefficients that affect the overall outline of the image and the high-frequency coefficients that are easily damaged by compression, which can improve the watermark's resistance to interference. Compare the coefficient values A at the (2,3) position and B at the (3,2) position, and decide whether to exchange or fine-tune their relationship according to the corresponding watermark bit: If the watermark bit to be embedded is 1, then A > B and the difference is greater than or equal to delta: If A is already greater than B and A-B is greater than or equal to delta, no modification is made. If A > B but the difference is less than delta, fine-tune A = B + delta. If A ≤ B, set A_new = B + delta / 2 and B_new = B-delta / 2. If the watermark bit to be embedded is 0, then A < B and the difference is greater than or equal to delta: If A is already less than B and B-A is greater than or equal to delta, no modification is made. If A < B but the difference is not enough delta, fine-tune to A = B - delta; If A ≥ B, swap and fine-tune to A_new = B - delta / 2, B_new = B + delta / 2; Wherein, delta is the minimum embedding strength, the value is between 3-10, which can ensure the stability of the watermark information while ensuring the visual quality of the image; IDCT (Inverse Discrete Cosine Transform) is performed on the block to convert the frequency domain information back to the spatial domain, restore the Y channel pixel value of the image, and then combine Cr and Cb channels to reorganize the complete image; Save the image embedded with the watermark, which is invisible in the image and does not affect the normal viewing of the image; The watermark extraction step includes: The image containing the watermark is loaded by OpenCV, which is converted from RGB color space to YCrCb color space, and the pixel value of the Y channel is extracted as a MAT object, which is consistent with the channel selection in the embedding step, ensuring the accuracy of watermark information extraction; The Y channel is divided into 64x64 blocks through the MAT object, and the block division method is completely consistent with the embedding step, ensuring the accuracy of the watermark extraction position; DCT transform is performed on the block to convert the pixel value to the frequency domain, preparing the frequency domain coefficients for extracting the watermark bits; Locate the embedded position of each block, i.e. (2,3) and (3,2), read the corresponding DCT coefficients A and B, and recover the watermark bit according to the size relationship: if A > B, record the bit as 1; otherwise, record it as 0. Since the watermark information is evenly distributed in the image through 64x64 blocks, even if the image is cut and partially blocked to some extent, the undamaged block can still provide effective bit information, and the approximate information of the watermark can be recovered. At the same time, since the information is extracted from the medium frequency position, the coefficients in this area are complete during image compression, so even if the image is compressed, the watermark information can still be extracted, and finally the entire watermark bit stream is reconstructed and filled into a 64x64 two-dimensional array; The binary information of the two-dimensional array is converted into a binary stream in row order, then converted into encrypted text information through encoding, and finally decrypted using the same key and AES-128 decryption algorithm as the embedding step to obtain the original watermark information, realizing image tracing.
[0022] In the watermark embedding process, by controlling the embedding strength factor delta (between 3-10), the balance between robustness and visual effect can be optimized: if the delta value is too small, the watermark is weak against interference and easy to lose in image processing; if the delta value is too large, it will cause local pixel mutation of the image, resulting in visible distortion. This value range can ensure the stability of the watermark while avoiding image edge jump, ensuring that the quality of the embedded watermark image is not significantly affected.
[0023] The working principle of the present application is as follows: The present application is based on a blind watermark-based picture tracing method, the core of which is to embed a watermark containing unique identification information into a picture in an invisible manner, and to accurately extract the watermark information without the original picture, thereby realizing the tracing of the picture. The operation principle revolves around the two key processes of watermark embedding and extraction, as follows: I. Watermark embedding principle The adaptability principle of image preprocessing: OpenCV is used to read the image and convert it to a uniform size of 512*512, which is for the subsequent standardized processing of image blocking and watermark information embedding. The uniform image size can ensure the uniformity and consistency of the blocking, so that the embedding position and method of the watermark information have a fixed reference, avoiding the chaos of embedding rules caused by the difference in image size, and laying a foundation for the accurate extraction of the watermark.
[0024] Security and adaptability generation principle of watermark information: The watermark information is composed of platform ID, user ID, upload timestamp and other unique identification information, which is formed by string splicing and uniform encoding processing to form a unique original watermark content. This original content is directly related to the identity information of the picture, and is the core basis for tracing.
[0025] The original watermark information is encrypted by using the AES-128 encryption method combined with a fixed initialization vector (IV) and a key derivation function (such as PBKDF2). The high-strength encryption characteristics of the AES encryption algorithm make it difficult to crack the content of the watermark information even if the watermark information is extracted illegally, ensuring the security of the watermark information. At the same time, the encrypted information is more suitable for embedding in the image in binary form.
[0026] Binary conversion and padding principle of watermark information: the encrypted watermark information is converted into a binary bit stream and filled into a 6464 0,1 two-dimensional array, the filling logic is to fill in each row, and the insufficient part is filled with 0 and the excessive part is truncated. This process is to match the structure of the watermark information with the structure of the subsequent image blocking, and the 6464 two-dimensional array corresponds to the 64*64 blocks into which the image will be divided, ensuring that each block can correspond to a watermark bit, and realizing the uniform distribution of the watermark information in the image.
[0027] The principle of embedding selection based on Y channel: the Y channel pixel value of the preprocessed image is converted into a MAT object by OpenCV, and the Y channel is selected for watermark embedding. In the YCrCb color space of the image, the Y channel represents the brightness information, and its influence on human visual is relatively stable. Embedding watermark in this channel is not easy to be detected, and the visual quality of the image can be better guaranteed.
[0028] The principle of embedding based on block and DCT transformation: The Y channel is divided into 64*64 blocks of 8*8, and DCT (Discrete Cosine Transform) is performed on each block. DCT can convert the image from spatial domain to frequency domain. The low-frequency coefficients of the image concentrate most of the energy, representing the overall outline of the image, and the high-frequency coefficients correspond to the details of the image, which are easily affected by compression and other operations. Selecting medium-frequency coefficients such as positions (2, 3) and (3, 2) for watermark embedding can avoid the defects of excessive influence of low-frequency coefficients on the overall quality of the image and the vulnerability of high-frequency coefficients to compression damage, and enhance the anti-interference ability of the watermark.
[0029] According to the watermark bit (0 or 1), the size relationship of the two medium-frequency coefficients (A and B) is compared, and the corresponding exchange or fine tuning is performed to make the coefficient relationship correspond to the watermark bit. For example, when embedding bit 1, ensure that A > B and the difference is not less than the minimum embedding strength delta; when embedding bit 0, ensure that A < B and the difference is not less than delta. The value of delta is between 3-10. By controlling this embedding strength factor, the stability of the watermark is guaranteed while avoiding obvious impact on the visual effect of the image, achieving a balance between robustness and visual effect.
[0030] The principle of image recovery by IDCT transformation: IDCT (Inverse Discrete Cosine Transform) is performed on the block after embedding the watermark, and the frequency domain information is converted back to the spatial domain, thereby recovering the image after embedding the watermark. At this time, the watermark has been integrated into the image in an invisible way, and the image is saved to complete the watermark embedding process.
[0031] II. Watermark extraction principle The principle of targeted extraction of Y channel: the image containing the watermark is loaded by OpenCV, and the pixel value of the Y channel is extracted as a MAT object. Since the watermark information is embedded in the Y channel, processing this channel can accurately lock the area where the watermark information is located.
[0032] The principle of information extraction preparation based on block and DCT transformation: the Y channel pixel value is divided into 64*64 blocks, and DCT is performed on each block to convert it to the frequency domain, preparing for the extraction of watermark information, because the watermark information is stored in the medium-frequency coefficients in the frequency domain.
[0033] The principle of recovering the watermark bit is as follows: locating the middle frequency position of embedding the watermark in each block such as (2, 3) and (3, 2), reading the corresponding DCT coefficients A and B, and recovering the watermark bit according to the size relationship of the two - if A > B, the bit is 1; otherwise, it is 0. Since the watermark information is uniformly distributed in the blocks, even if the image is cut or partially blocked, most of the watermark information can be recovered from the complete block, so that the approximate watermark content can be obtained. At the same time, the middle frequency coefficient is relatively stable in the image compression process, so even if the picture is compressed, the watermark bit can be successfully extracted, and finally the 64*64 two-dimensional array form of the watermark bit stream can be reconstructed.
[0034] The principle of decrypting and restoring the watermark information is as follows: converting the two-dimensional array of the watermark binary information into a binary stream, then into encrypted text information, and finally obtaining the original watermark information through AES decryption. Since the watermark information is encrypted by AES before embedding, only the authorized party with the decryption key can obtain the original watermark information, even if the watermark is illegally extracted, the content cannot be cracked, further ensuring the security of the watermark information.
[0035] Through the above principles, the present application realizes the stable, safe and invisible embedding of the watermark into the image under the premise of ensuring the image quality, and can still accurately extract the watermark information after the image is cut, compressed and the like, thereby effectively realizing the identity identification and traceability of the picture, and improving the copyright protection ability and content supervision efficiency of the digital picture.
[0036] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the protection scope of the technical solutions of the embodiments of the present application.
Claims
1. A blind watermark-based picture provenance method, characterized in that, The watermark embedding step and the watermark extracting step are included. The watermark embedding step includes: Image preprocessing: reading an image by using OpenCV and converting the image into an image with a size of 512*512; Watermark information generation: forming unique identification information composed of a platform ID, a user ID and an upload timestamp by string splicing, processing the unique identification information through a unified encoding format, and encrypting the unique identification information by using an AES-128 encryption mode in combination with a fixed initialization vector and a key derivation function; Converting the encrypted watermark information into a binary bit stream by using a string-to-binary stream function, and filling the binary bit stream into a 64*64 two-dimensional array of 0 and 1, the filling logic being filling in each row, and if the bits are insufficient, 0 is supplemented, and if the bits are excessive, the bits are truncated; Converting pixel values of a Y channel of the preprocessed image into a MAT object by using OpenCV; Dividing the Y channel into 64*64 blocks through the MAT object, selecting two medium-frequency positions in a DCT result of each 8*8 block, comparing the coefficient values of the two medium-frequency positions, and deciding whether to exchange or fine-tune the relationship between the two medium-frequency positions according to corresponding watermark bits; IDCT transforming the blocks to recover the image; Saving the image in which the watermark information is embedded; The watermark extracting step includes: Loading Y channel pixel values of the image containing the watermark as a MAT object by using OpenCV; Dividing the Y channel pixel values into 64*64 blocks through the MAT object; DCT transforming the blocks; Positioning to the embedding position of each block, reading corresponding DCT coefficients, recovering watermark bits according to the size relationship, if A>B, recording the bit as 1, otherwise recording the bit as 0, finally reconstructing the entire watermark bit stream, and filling the watermark bit stream into a 64*64 two-dimensional array; Converting the watermark binary information of the two-dimensional array into a binary stream, then into encrypted text information, and then decrypting the text information by using AES to obtain original watermark information.
2. The blind watermark-based picture provenance method of claim 1, wherein, In the watermark embedding step 5), the medium-frequency positions are (2, 3) and (3, 2).
3. The blind watermark-based picture provenance method of claim 1, wherein, In the watermark embedding step 5), the deciding whether to exchange or fine-tune the relationship between the two medium-frequency positions according to corresponding watermark bits is specifically: If the watermark bit to be embedded is 1, A>B and the difference is greater than or equal to delta are required: If A is already greater than B and A-B is greater than or equal to delta, no modification is made; If A is greater than B but the difference is less than delta, A is fine-tuned to B+delta; If A is less than B, A_new is set to B+delta / 2 and B_new is set to B-delta / 2; If the watermark bit to be embedded is 0, A<B and the difference is greater than or equal to delta are required: If A is already less than B and B-A is greater than or equal to delta, no modification is made; If A is less than B but the difference is less than delta, A is fine-tuned to B-delta; If A is greater than or equal to B, A_new and B_new are exchanged and fine-tuned to B-delta / 2 and B+delta / 2, respectively; Wherein, A represents the coefficient value of position (2, 3) in the DCT block, B represents the coefficient value of position (3, 2) in the DCT block, and delta is the minimum embedding strength.
4. The blind watermark-based picture provenance method of claim 3, wherein, The value of the delta is between 3 and 10.
5. The blind watermark based picture provenance method of claim 1, wherein, In the watermark information generation step, the key derivation function is PBKDF2.
6. The blind watermark-based picture provenance method of claim 1, wherein, In the watermark extraction step 4), since the watermark information is evenly distributed by block, even if the picture is cut and partially blocked, the approximate information of the watermark can be recovered; at the same time, by extracting the watermark information from the medium frequency position of the DCT result, even if the picture is compressed, the watermark information can be extracted.
7. The blind watermark based picture provenance method of claim 1, wherein, The watermark information is embedded by AES encryption, so even if the watermark information is illegally extracted, it cannot be decrypted.
8. The blind watermark based picture provenance method of claim 1, wherein, In the watermark embedding process, by controlling the embedding strength factor, the balance between robustness and visual effect is optimized, which can ensure the stability of the watermark while avoiding image edge jump.