Watermark identification detection method and device for video image, equipment and storage medium

Abstract

The invention discloses a video image watermark identification detection method and device, equipment and a storage medium, which can improve the reliability of a detection result by combining a plurality of watermark detection technologies so as to better adapt to various watermark styles and embedding modes and improve the adaptability and flexibility of detection. The method comprises the following steps: acquiring a video image to be detected; extracting a target watermark feature of the to-be-detected video image, and performing template watermark recognition on the target watermark feature by adopting a pre-constructed watermark identification template library to obtain a template recognition result; extracting target metadata and target text data of the to-be-detected video image, and respectively identifying the target metadata and the target text information to obtain a metadata identification result and a text identification result; according to the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result, the template recognition result, the metadata recognition result and the text recognition result are integrated, and a target watermark identification detection result is obtained.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for detecting watermarks in video images. Background Art

[0002] With the rapid development of artificial intelligence technology, generative artificial intelligence models are able to generate realistic and high-quality images and video content. However, technological advances have also brought security risks such as copyright infringement and the spread of false information. In order to meet these challenges and ensure the compliance and traceability of image content, relevant technicians usually use watermark detection technology to detect watermarks on video images. By detecting the watermark in the image, the copyright of the image can be determined, unauthorized use and dissemination can be prevented, and the legitimate rights and interests of the copyright owner can be protected. At the same time, the watermark can also be used as a basis for verifying the authenticity of the image. By detecting the presence of the watermark, it can be determined whether it has been tampered with or forged, which helps prevent the spread of false information and reduce security risks.

[0003] In the related technology, traditional detection methods usually assume that the watermark and the content are independent of each other, and use traditional Fourier transform or wavelet transform technology to extract effective features and then determine whether there is a watermark.

[0004] In the process of implementing this application, the applicant found that the related technology has at least the following problems:

[0005] In complex watermark scenarios, watermarks have anti-tampering capabilities. Using traditional detection methods for analysis is not only difficult to extract effective features, but may also damage the features of the watermark, resulting in the traditional detection method being unable to accurately identify whether there is a watermark in the video image to be detected. Summary of the invention

[0006] In view of this, the present application provides a method, device, equipment and storage medium for detecting watermarks in video images. The main purpose is to solve the problem that in the current complex watermark scenarios, the watermark has an anti-tampering function, and when using traditional detection methods for analysis, it is not only difficult to extract effective features, but also may cause damage to the features of the watermark, resulting in the traditional detection method being unable to accurately identify whether there is a watermark in the video image to be detected.

[0007] According to the first aspect of the present application, a method for detecting a watermark in a video image is provided, the method comprising:

[0008] Acquire a video image to be detected, where the video image to be detected is a generative artificial intelligence video image;

[0009] Extracting the target watermark feature of the video image to be detected, and using a pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result;

[0010] Extracting target metadata and target text data of the video image to be detected, and respectively identifying the target metadata and the target text information to obtain a metadata recognition result and a text recognition result;

[0011] According to the weights respectively corresponding to the template recognition result, the metadata recognition result and the text recognition result, the template recognition result, the metadata recognition result and the text recognition result are integrated to obtain a target watermark identification detection result.

[0012] Optionally, the extracting the target watermark feature of the video image to be detected, using a pre-built watermark identification template library to perform template watermark recognition on the target watermark feature, and obtaining a template recognition result includes:

[0013] Performing edge detection and texture analysis on the video image to be detected to determine edge information and texture features corresponding to the video image to be detected;

[0014] Determine a color histogram corresponding to the video image to be detected, extract deep features of the video image to be detected based on a pre-trained deep neural network, and determine the target watermark feature according to the edge information, the texture feature, the color histogram and the deep feature;

[0015] A pre-constructed watermark identification template library is used to perform template watermark recognition on the target watermark feature to obtain a template recognition result, wherein the watermark identification template library stores multiple specified watermark features, and the types of the specified watermark features include but are not limited to text watermark types and image watermark types.

[0016] Optionally, the using a pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result includes:

[0017] Determine a first feature vector corresponding to the target watermark feature, determine a second feature vector of each designated watermark feature, calculate a similarity between the first feature vector and the second feature vector based on a template matching algorithm, and associate the similarity with the designated watermark feature;

[0018] Comparing the similarity associated with each specified watermark feature with a preset similarity threshold;

[0019] If there is a designated watermark feature with a similarity greater than the similarity threshold, then determine the watermark identification template corresponding to the designated watermark feature, and determine a measurable function value indicating the existence of the watermark identification, and generate the template recognition result according to the measurable function value and the watermark information of the watermark identification template, wherein the watermark information is used to indicate the artificial intelligence large model for generating the video image to be detected;

[0020] If there is no designated watermark feature with a similarity greater than the similarity threshold, a measurable function value indicating that there is no watermark identifier is determined, and the template recognition result is generated according to the measurable function value.

[0021] Optionally, extracting target metadata of the to-be-detected video image, identifying the target metadata, and obtaining a metadata identification result includes:

[0022] Extracting target metadata of the video image to be detected, parsing the target metadata to obtain metadata information, wherein the metadata information includes but is not limited to EXIF ​​data, creation date, and device information;

[0023] The metadata information is identified using the predefined metadata format standard, wherein the metadata format standard records standard fields corresponding to a plurality of artificial intelligence models;

[0024] If a standard field exists in the metadata information, determining an artificial intelligence model corresponding to the standard field and a measurable function value indicating the presence of a watermark identifier, and generating the metadata recognition result according to the measurable function value and the model identifier of the artificial intelligence model;

[0025] If the standard field does not exist in the metadata information, a measurable function value indicating that there is no watermark identifier is determined, and the metadata identification result is generated according to the measurable function value.

[0026] Optionally, extracting target text information of the video image to be detected, recognizing the target text information, and obtaining a text recognition result includes:

[0027] Using optical character recognition technology to extract the target text information from the video image to be detected;

[0028] Using natural language processing technology to identify keywords in the target text information;

[0029] If a keyword related to the watermark is identified, an artificial intelligence model matching the keyword and a measurable function value indicating the presence of the watermark are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model;

[0030] If no keyword related to the watermark is identified, a measurable function value indicating that the watermark does not exist is determined, and the text recognition result is generated according to the measurable function value.

[0031] Optionally, extracting target text information of the video image to be detected, recognizing the target text information, and obtaining a text recognition result includes:

[0032] Based on a keyword matching algorithm, the target text information is identified using predefined watermark keywords;

[0033] If the recognition result indicates that a watermark keyword exists in the target text information, an artificial intelligence model matching the watermark keyword and a measurable function value indicating the presence of a watermark identifier are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model;

[0034] If the recognition result indicates that the watermark keyword does not exist in the target text information, a measurable function value indicating that the watermark identifier does not exist is determined, and the text recognition result is generated according to the measurable function value.

[0035] Optionally, the step of integrating the template recognition result, the metadata recognition result and the text recognition result according to the weights respectively corresponding to the template recognition result, the metadata recognition result and the text recognition result to obtain the target watermark identification detection result includes:

[0036] Acquire training data, the training data including training samples, detection results of the training samples and actual labels of the training samples, the actual labels are used to indicate whether a watermark exists;

[0037] The template recognition result, metadata recognition result and text recognition result in the detection result are combined into a feature vector, and a machine recognition model is used to learn the relationship between the feature vector and the actual label to obtain an initial weight corresponding to each recognition result, and based on a cross-validation method, training data is continuously obtained to adjust the initial weight of each recognition result to obtain weights corresponding to the template recognition result, the metadata recognition result and the text recognition result respectively;

[0038] The predictable functions in the template recognition result, the metadata recognition result and the text recognition result are determined respectively, and each predictable function is integrated according to the weight corresponding to each recognition result to obtain the target watermark identification detection result.

[0039] According to a second aspect of the present application, a video image watermark detection device is provided, the device comprising:

[0040] An acquisition module, used to acquire a video image to be detected, wherein the video image to be detected is a generative artificial intelligence video image;

[0041] The first recognition module is used to extract the target watermark feature of the video image to be detected, and use the pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result;

[0042] The second recognition module is used to extract the target metadata and target text data of the video image to be detected, and recognize the target metadata and the target text information respectively to obtain a metadata recognition result and a text recognition result;

[0043] The integration module is used to integrate the template recognition result, the metadata recognition result and the text recognition result according to the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result respectively, so as to obtain the target watermark identification detection result.

[0044] Optionally, the first recognition module is used to perform edge detection and texture analysis on the video image to be detected, determine the edge information and texture features corresponding to the video image to be detected; determine the color histogram corresponding to the video image to be detected, and extract the deep features of the video image to be detected based on a pre-trained deep neural network, and determine the target watermark features according to the edge information, the texture features, the color histogram and the deep features; use a pre-built watermark identification template library to perform template watermark recognition on the target watermark features to obtain a template recognition result, wherein the watermark identification template library stores multiple specified watermark features, and the types of the specified watermark features include but are not limited to text watermark types and image watermark types.

[0045] Optionally, the first recognition module is used to determine a first feature vector corresponding to the target watermark feature, determine a second feature vector of the specified watermark feature for each specified watermark feature, calculate the similarity between the first feature vector and the second feature vector based on a template matching algorithm, and associate the similarity with the specified watermark feature; compare the similarity associated with each specified watermark feature with a preset similarity threshold; if there is a specified watermark feature with a similarity greater than the similarity threshold, determine a watermark identification template corresponding to the specified watermark feature, and determine a measurable function value indicating the existence of a watermark identification, and generate the template recognition result according to the measurable function value and the watermark information of the watermark identification template, the watermark information is used to indicate the generation of an artificial intelligence large model for the video image to be detected; if there is no specified watermark feature with a similarity greater than the similarity threshold, determine a measurable function value indicating the absence of a watermark identification, and generate the template recognition result according to the measurable function value.

[0046] Optionally, a second identification module is used to extract target metadata of the video image to be detected, parse the target metadata to obtain metadata information, the metadata information including but not limited to EXIF ​​data, creation date, and device information; use the predefined metadata format standard to identify the metadata information, and the metadata format standard records standard fields corresponding to multiple artificial intelligence models; if there is a standard field in the metadata information, determine the artificial intelligence model corresponding to the standard field and a measurable function value indicating the presence of a watermark identifier, and generate the metadata identification result according to the measurable function value and the model identifier of the artificial intelligence model; if there is no standard field in the metadata information, determine a measurable function value indicating the absence of a watermark identifier, and generate the metadata identification result according to the measurable function value.

[0047] Optionally, a second recognition module is used to extract the target text information from the video image to be detected using optical character recognition technology; and to recognize keywords in the target text information using natural language processing technology; if keywords related to the watermark are recognized, an artificial intelligence model matching the keywords and a measurable function value indicating the presence of a watermark are determined, and the text recognition result is generated based on the measurable function value and the model identifier of the artificial intelligence model; if keywords related to the watermark are not recognized, a measurable function value indicating the absence of a watermark is determined, and the text recognition result is generated based on the measurable function value.

[0048] Optionally, the second recognition module is used to identify the target text information based on a keyword matching algorithm and using predefined watermark keywords; if the recognition result indicates that the target text information contains a watermark keyword, then an artificial intelligence model that matches the watermark keyword and a measurable function value indicating the existence of a watermark identifier are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model; if the recognition result indicates that the target text information does not contain a watermark keyword, then a measurable function value indicating the absence of a watermark identifier is determined, and the text recognition result is generated according to the measurable function value.

[0049] Optionally, an integration module is used to obtain training data, the training data including training samples, detection results of the training samples and actual labels of the training samples, the actual labels are used to indicate whether a watermark exists; the template recognition results, metadata recognition results and text recognition results in the detection results are combined into a feature vector, and a machine recognition model is used to learn the relationship between the feature vector and the actual label to obtain an initial weight corresponding to each recognition result; and based on a cross-validation method, training data is continuously obtained to adjust the initial weight of each recognition result to obtain weights corresponding to the template recognition result, the metadata recognition result and the text recognition result respectively; predictable functions in the template recognition result, the metadata recognition result and the text recognition result are respectively determined, and each predictable function is integrated according to the weight corresponding to each recognition result to obtain the target watermark identification detection result.

[0050] According to a third aspect of the present application, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods described in the first aspect when executing the computer program.

[0051] According to a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0052] By means of the above technical scheme, the present application provides a method, device, equipment and storage medium for detecting watermarks of video images. The present application first obtains a video image to be detected, and the video image to be detected is a generative artificial intelligence video image. Further, the target watermark feature of the video image to be detected is extracted, and the target watermark feature is identified by a pre-built watermark template library to obtain a template recognition result. Next, the target metadata and target text data of the video image to be detected are extracted, and the target metadata and target text information are identified respectively to obtain metadata recognition results and text recognition results. Finally, according to the weights corresponding to the template recognition results, metadata recognition results and text recognition results, the template recognition results, metadata recognition results and text recognition results are integrated to obtain the target watermark detection result. Different watermark styles and embedding methods may require different recognition methods. The embodiments of the present application provide a variety of technical means for watermark detection, and there is complementarity between different technical means. The template-based recognition results, metadata analysis results and text information extraction results are integrated to make the final watermark detection judgment. The combination of multiple technical means can improve the reliability of detection results, thereby better adapting to various watermark styles and embedding methods, and improving the adaptability and flexibility of detection.

[0053] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0055] Figure 1 A schematic diagram of a video image watermark detection method provided in an embodiment of the present application is shown;

[0056] Figure 2 A schematic diagram of a video image watermark detection method provided in an embodiment of the present application is shown;

[0057] Figure 3 A schematic diagram of a video image watermark detection method provided in an embodiment of the present application is shown;

[0058] Figure 4 A schematic diagram of the structure of a video image watermark detection device provided in an embodiment of the present application is shown;

[0059] Figure 5 A schematic diagram of the device structure of a computer device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0060] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0061] With the rapid development of artificial intelligence technology, generative artificial intelligence models are able to generate realistic and high-quality images and video content. However, this also brings potential security risks, such as copyright infringement and the spread of false information. In order to solve these problems and ensure the compliance and traceability of image content, effective image security assessment methods are required, and watermark detection is an important technology.

[0062] The "Regulations on the Management of Deep Synthesis of Internet Information Services" issued by the Cyberspace Administration of China in 2022 requires that if deep synthesis service providers provide the following deep synthesis services that may cause confusion or misunderstanding among the public, they shall prominently mark the generated or edited information content in reasonable locations and areas. The "Interim Measures for the Management of Generative Artificial Intelligence Services" issued by the Cyberspace Administration of China in 2023 requires that generative artificial intelligence service providers shall mark the generated content such as pictures and videos. The present method provides a generative artificial intelligence video image watermark detection method. Watermark detection is a key technology in the image security assessment system, and its function is to identify whether there is a watermark in the image. Existing watermark detection methods are mainly based on image processing technologies, such as Fourier transform, wavelet transform, etc. However, these methods are generally difficult to adapt to complex and changeable watermark styles and embedding methods, and are easily interfered by factors such as image compression and noise. Therefore, an embodiment of the present application provides a watermark detection method for video images, such as Figure 1 As shown, the method includes:

[0063] 101. Obtain a video image to be detected, where the video image to be detected is a generative artificial intelligence video image.

[0064] Generative AI has a wide range of applications in the fields of video and images. Among them, generative adversarial networks (GANs) are the most commonly used generative models, variational autoencoders (VAEs) generate new images by learning the latent space of data, and are usually used for image compression and generation, and MoCoGAN focuses on video generation, which can create high-quality dynamic images. However, with the rapid development of generative artificial intelligence technology, the importance of watermark detection has become increasingly prominent. Watermarks are not only used to identify copyright and origin, but also to prevent the abuse of content to a certain extent. Combined with watermark detection technology, it is helpful to effectively identify and track these generated video images, thereby safeguarding the rights of the original creator. Therefore, in an embodiment of the present application, the video image to be detected for which a watermark needs to be detected is specifically a video image generated by generative artificial intelligence.

[0065] 102. Extract target watermark features of the video image to be detected, use a pre-built watermark identification template library to perform template watermark recognition on the target watermark features, and obtain a template recognition result.

[0066] Among them, the watermark template library stores a variety of currently known generative large model watermarks. By calculating the similarity between the target watermark features and the watermark features of each watermark template, it is determined whether there is a watermark template matching the target watermark features in the watermark template library, and then the template recognition result is obtained. The template recognition result is used to indicate whether a watermark is detected in the template recognition dimension.

[0067] 103. Extract target metadata and target text data of the video image to be detected, recognize the target metadata and target text information respectively, and obtain metadata recognition results and text recognition results.

[0068] Among them, metadata is used to describe, explain, locate or manage the properties and characteristics of other data. In the field of video and image processing, metadata usually contains various information about the video or image itself, which helps to verify the copyright ownership and content integrity of the video or image. For example, some generative artificial intelligence software will automatically embed specific watermark information in the generated content. If the metadata shows that such a tool is used, the corresponding watermark may exist. In other words, some tools specifically used to generate or identify watermarks may add specific logos or tags when generating metadata, such as keywords such as "watermarked" or "contains watermarks" to indicate that the file contains a watermark.

[0069] Furthermore, considering that watermark text usually appears in fixed positions of images, such as corners, edges or central areas. By identifying the target text data in the image, the existing watermark content can be detected. Combined with specific text features, position, format and context information, it can be effectively determined whether there is a watermark in the video image to be detected. This method, combined with visual feature analysis and metadata verification, can provide more comprehensive watermark detection results.

[0070] 104. According to the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result, the template recognition result, the metadata recognition result and the text recognition result are integrated to obtain the target watermark identification detection result.

[0071] In the embodiments of the present application, different weights are assigned to each detection method according to its reliability and accuracy, and then multiple detection results are integrated by weighted summation, which can reduce the errors and blind spots that may exist in a single detection method, thereby improving the accuracy of the overall detection results.

[0072] The method provided by the embodiment of the present application can first obtain a video image to be detected, and the video image to be detected is a generative artificial intelligence video image. Further, the target watermark feature of the video image to be detected is extracted, and the target watermark feature is identified by a pre-built watermark identification template library to obtain a template identification result. Next, the target metadata and target text data of the video image to be detected are extracted, and the target metadata and target text information are identified respectively to obtain metadata identification results and text identification results. Finally, according to the weights corresponding to the template identification results, metadata identification results and text identification results, the template identification results, metadata identification results and text identification results are integrated to obtain the target watermark identification detection result. Different watermark styles and embedding methods may require different identification methods. The embodiment of the present application provides a variety of technical means for watermark identification detection. There is complementarity between different technical means, and the template-based identification results, metadata analysis results and text information extraction results are integrated to make the final watermark detection judgment. The combination of multiple technical means can improve the reliability of the detection results, and thus can better adapt to various watermark styles and embedding methods, and improve the adaptability and flexibility of detection.

[0073] The present application embodiment provides a method for detecting a watermark in a video image. Figure 2 As shown, the method includes:

[0074] 201. Obtain a video image to be detected, where the video image to be detected is a generative artificial intelligence video image.

[0075] With the rapid development of artificial intelligence technology, generative artificial intelligence models can generate realistic and high-quality images and video content. However, technological advances have also brought security risks such as copyright infringement and the spread of false information. In order to meet these challenges and ensure the compliance and traceability of image content, relevant technicians usually use watermark detection technology to detect watermarks on video images. By detecting the watermark in the image, the copyright of the image can be determined, unauthorized use and dissemination can be prevented, and the legitimate rights and interests of the copyright owner can be protected. At the same time, the watermark can also be used as a basis for verifying the authenticity of the image. By detecting whether the watermark exists, it can be determined whether it has been tampered with or forged, which helps to prevent the spread of false information and reduce security risks. At present, traditional detection methods usually assume that the watermark is independent of the content. The traditional Fourier transform or wavelet transform technology can be used to extract effective features and then determine whether there is a watermark. However, the applicant recognizes that in complex watermark scenarios, the watermark has an anti-tampering function. When using traditional detection methods for analysis, it is not only difficult to extract effective features, but also may damage the features of the watermark, resulting in the inability of traditional detection methods to accurately identify whether there is a watermark in the video image to be detected. Therefore, the present application provides a method for detecting watermarks of video images. The present application first obtains a video image to be detected, and the video image to be detected is a generative artificial intelligence video image. Further, the target watermark feature of the video image to be detected is extracted, and the target watermark feature is identified by a pre-built watermark template library to obtain a template recognition result. Next, the target metadata and target text data of the video image to be detected are extracted, and the target metadata and target text information are identified respectively to obtain metadata recognition results and text recognition results. Finally, according to the weights corresponding to the template recognition results, metadata recognition results and text recognition results, the template recognition results, metadata recognition results and text recognition results are integrated to obtain the target watermark detection result. Different watermark styles and embedding methods may require different recognition methods. The embodiments of the present application provide a variety of technical means for watermark detection, and there is complementarity between different technical means. The template-based recognition results, metadata analysis results and text information extraction results are integrated to make the final watermark detection judgment. The combination of multiple technical means can improve the reliability of the detection results, and thus can better adapt to various watermark styles and embedding methods, and improve the adaptability and flexibility of detection.

[0076] In this step, the relevant technicians can upload the video image to be detected based on the visual interface of the terminal device. It should be noted that the terminal device can be a smart device such as a mobile phone, computer or tablet computer. This application does not specifically limit the model of the terminal device. After the watermark detection system detects that the relevant technicians have uploaded the original video image, in order to improve the accuracy and efficiency of the subsequent processing steps, such as Figure 3 As shown, it is necessary to obtain the original video image for image preprocessing, and then obtain the video image to be detected. The steps of image preprocessing include but are not limited to image denoising, image enhancement and size adjustment. The specific process of image preprocessing is as follows:

[0077] First, the original video image is denoised to remove the noise in the original video image and obtain the video image to be enhanced. In video image processing, noise refers to unexpected signals or interference that are unrelated to the content of the original image. In the actual operation process, the classic filtering algorithms such as Gaussian filtering and median filtering can be selected according to the type and distribution characteristics of the noise to remove the noise in the original video image. This application does not specifically limit the filtering algorithm. In this step, by removing the noise in the original video image, the interference of noise on OCR (Optical Character Recognition) recognition and watermark feature extraction can be reduced, thereby improving the operating efficiency of subsequent processing steps.

[0078] Furthermore, in order to enhance the visual effect of the image and make the watermark information clearer and more recognizable, histogram equalization, contrast adjustment and other methods can be used to process the enhanced video image to obtain the image to be adjusted. It should be noted that histogram equalization is to adjust the intensity distribution of image pixels so that they are evenly distributed throughout the grayscale range, thereby improving the contrast and detail visibility of the image. Contrast adjustment is to adjust the intensity range of image pixels through linear or nonlinear transformation, thereby enhancing the contrast of the image. In the actual operation process, relevant technical personnel can select a suitable image enhancement method according to actual needs, and this application does not specifically limit the image enhancement method used.

[0079] Finally, according to the requirements of OCR technology and the needs of subsequent processing, the image size is adjusted to a suitable resolution to improve processing speed and recognition accuracy. For example, the resolution of the image to be adjusted is adjusted to 300-400 DPI, and the size is adjusted to 1000x1500 pixels, and then the video image to be detected is obtained. By adjusting the size and resolution reasonably, the performance of the OCR system can be significantly improved.

[0080] In the embodiment of the present application, through the above-mentioned image preprocessing steps, the watermark detection system can effectively remove noise in the original video image, enhance the visual effect of the image, and adjust the size and resolution of the image, so as to obtain high-quality video images to be detected, thereby improving the accuracy and efficiency of subsequent processing.

[0081] 202. Extract target watermark features of the video image to be detected, use a pre-built watermark identification template library to perform template watermark recognition on the target watermark features, and obtain a template recognition result.

[0082] In the embodiment of the present application, watermark feature extraction is performed on the video image to be detected, and feature information related to the watermark is extracted from the possible candidate areas to obtain the target watermark feature. Further, the target watermark feature is subjected to template watermark recognition through the watermark identification template library to determine whether there is a specified watermark feature similar to the target watermark feature in the watermark identification library, and then a template recognition result is generated according to the watermark information corresponding to the specified watermark feature. The specific process of determining the template recognition result is as follows:

[0083] First, use the preset algorithms (such as edge detection, shape detection, etc.) to perform edge detection and texture analysis on the video image to be detected, and determine the edge information and texture features corresponding to the video image to be detected. Then determine the color histogram corresponding to the video image to be detected, and extract the deep features of the video image to be detected based on the pre-trained deep neural network, and determine the target watermark features based on the edge information, texture features, color histogram and deep features.

[0084] Furthermore, if Figure 3As shown, a pre-constructed watermark identification library is determined, and the watermark identification library is used to perform template watermark recognition on the target watermark feature to determine the template recognition result. It should be noted that the watermark identification template library is pre-constructed by relevant technical personnel based on a variety of currently known generative large model watermark identifications. Specifically, relevant technical clerks collect watermark identification samples of different generative models, including but not limited to text watermarks, graphic watermarks, etc., and then standardize the samples, including resizing, color conversion, etc., to ensure the consistency of watermarks in the template library, and finally construct a watermark identification template library. The watermark identification template library stores multiple watermark identification templates, and the watermark features corresponding to the watermark identification templates are designated watermark features. The types of designated watermark features include but are not limited to text watermark features and image watermark features. Specifically, the first feature vector corresponding to the target watermark feature is determined, and for each designated watermark feature, the second feature vector of the designated watermark feature is determined. Based on the template matching algorithm, such as the normalized cross-correlation algorithm, the square difference matching algorithm, etc., the similarity between the first feature vector and the second feature vector is calculated, and the similarity is associated with the designated watermark feature. The similarity associated with each specified watermark feature is compared with a preset similarity threshold. It should be noted that the similarity threshold can adopt the default value of the watermark identification detection system, or it can adopt the value set by relevant technical personnel according to actual needs. The present application does not specifically limit the value of the similarity threshold. If there is a specified watermark feature with a similarity greater than the similarity threshold, the watermark identification template corresponding to the specified watermark feature is determined, and the measurable function value indicating the existence of the watermark identification is determined, and the template recognition result is generated according to the measurable function value and the watermark information of the watermark identification template, wherein the measurable function value can represent the probability or confidence of the existence of the watermark in the image or video, and the watermark information is used to indicate the generation of the artificial intelligence large model of the video image to be detected. If there is no specified watermark feature with a similarity greater than the similarity threshold, the measurable function value indicating the absence of the watermark identification is determined, and the template recognition result is generated according to the measurable function value. In addition, with the addition of new watermark samples, the watermark identification template library can be continuously updated, thereby improving the adaptability of the watermark identification detection system.

[0085] 203. Extract target metadata of the video image to be detected, identify the target metadata, and obtain a metadata identification result.

[0086] In the embodiment of the present application, the target metadata of the video image to be detected can be parsed using an existing standard metadata parsing tool or library (such as ExifTool). In order to facilitate subsequent operations, the format of the parsed target metadata needs to be standardized, and then Figure 3 As shown, by analyzing the standardized metadata information to identify whether there are possible watermark clues, the metadata recognition result is obtained.

[0087] Specifically, the target metadata of the video image to be detected is extracted, and the target metadata is parsed to obtain metadata information. In the actual operation process, the metadata information includes but is not limited to EXIF ​​data, creation date, and device information. Then, the metadata information is identified using a predefined metadata format standard. It should be noted that the metadata format standard records standard fields corresponding to multiple artificial intelligence models. In the metadata information of the video image generated by the artificial intelligence model, these standardized metadata fields can provide key parameters and information about the generated content. For example, the device information in the metadata information has a field AI_Generator_Model_XYZ, and the field indicates the name of the artificial intelligence model that generates the video image. Further, if there is a standard field in the metadata information, the artificial intelligence model corresponding to the standard field and the measurable function value indicating the presence of the watermark mark are determined. In the process of watermark detection and identification, "measurable function value" generally refers to a value used to quantify or represent a specific attribute (such as the existence, non-existence, or confidence of related features of a watermark). This value can be calculated based on the extracted features and similarity. The metadata identification result is generated based on the measurable function value and the model identifier of the artificial intelligence model. If there is no standard field in the metadata information, a measurable function value indicating that there is no watermark is determined, and a metadata recognition result is generated based on the measurable function value. For example, the extracted metadata is parsed to obtain specific metadata information EXIF ​​data: {"Camera model": "AI_Model_XYZ", "Exposure time": "1 / 1000sec"}, creation date: "2024-10-20T14:30:00", device information: "AI_Model_XYZ_Device_01". The predefined metadata format standard is AI_Model_XYZ: ["Camera model", "Device information"]. The "Camera model" and "Device information" in the metadata information are matched with the standard fields of AI_Model_XYZ. If there is a standard field (for example, "Camera model" and "Device information" match the standard field), the corresponding artificial intelligence model (AI_Model_XYZ) is determined, and a positive measurable function value (for example, 0.85) is generated to indicate the probability of the presence of a watermark. If the standard field does not exist in the metadata information, a zero value or a negative value (eg, 0) is generated, indicating the probability that there is no watermark identification.

[0088] 204. Extract target text information from the video image to be detected, recognize the target text information, and obtain a text recognition result.

[0089] In the embodiments of the present application, Figure 3As shown in the figure, the watermark detection system uses OCR technology to extract text information from the video image to be detected. Then, by performing natural language processing on the target text information, or matching the target text information with a set of keywords related to the watermark, such as copyright information, image source, creator name, etc., it searches for text features that match the watermark in the target text information and generates text recognition results.

[0090] Specifically, the target text information is extracted from the video image to be detected by using optical character recognition technology. The keywords in the target text information are identified by using natural language processing technology. If the keywords related to the watermark are identified, the artificial intelligence model matching the keywords and the measurable function value indicating the presence of the watermark are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model. If the keywords related to the watermark are not identified, the measurable function value indicating the absence of the watermark is determined, and the text recognition result is generated according to the measurable function value. Or based on the keyword matching algorithm, the target text information is identified by using predefined watermark keywords, such as copyright information, image source, creator name, etc. If the recognition result indicates that the watermark keyword exists in the target text information, the artificial intelligence model matching the watermark keyword and the measurable function value indicating the presence of the watermark are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model. If the recognition result indicates that the watermark keyword does not exist in the target text information, the measurable function value indicating the absence of the watermark is determined, and the text recognition result is generated according to the measurable function value.

[0091] 205. According to the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result, the template recognition result, the metadata recognition result and the text recognition result are integrated to obtain the target watermark identification detection result.

[0092] In the embodiments of the present application, Figure 3As shown, according to the importance of different recognition results, the weights corresponding to the template recognition result, metadata recognition result and text recognition result are set. According to the weights corresponding to the template recognition result, metadata recognition result and text recognition result, the template recognition result, metadata recognition result and text recognition result are integrated to obtain the target watermark identification detection result. It should be noted that the setting of the weight can be optimized based on experience or through training data. Specifically, the ensemble learning technology is used to obtain the weights for determining different results in the actual operation process. Training data can be obtained, and the training data includes training samples, detection results of training samples and actual labels of training samples, wherein the actual label is used to indicate whether there is a watermark. The template recognition result, metadata recognition result and text recognition result in the detection result are combined into a feature vector, and the relationship between the feature vector and the actual label is learned by a machine recognition model to obtain the initial weight corresponding to each recognition result. Then, based on the cross-validation method, the training data is continuously obtained to adjust the initial weight of each recognition result to obtain the weights corresponding to the template recognition result, metadata recognition result and text recognition result. For example, the weight of the template recognition result is 0.5, the weight of the metadata recognition result is 0.3, and the weight of the text recognition result is 0.2. Determine the predictable functions in the template recognition result, metadata recognition result, and text recognition result respectively, integrate each predictable function according to the weight corresponding to each recognition result, and obtain the target watermark detection result. The embodiment of the present application realizes comprehensive judgment and recognition of watermarks in artificial intelligence generated images by integrating template recognition results, metadata recognition results, and text recognition results. This method not only takes into account the visual content of the image, but also the metadata and embedded text information of the image, thereby improving the accuracy and reliability of recognition.

[0093] The method provided by the embodiment of the present application can first obtain a video image to be detected, and the video image to be detected is a generative artificial intelligence video image. Further, the target watermark feature of the video image to be detected is extracted, and the target watermark feature is identified by a pre-built watermark identification template library to obtain a template identification result. Next, the target metadata and target text data of the video image to be detected are extracted, and the target metadata and target text information are identified respectively to obtain metadata identification results and text identification results. Finally, according to the weights corresponding to the template identification results, metadata identification results and text identification results, the template identification results, metadata identification results and text identification results are integrated to obtain the target watermark identification detection result. Different watermark styles and embedding methods may require different identification methods. The embodiment of the present application provides a variety of technical means for watermark identification detection. There is complementarity between different technical means, and the template-based identification results, metadata analysis results and text information extraction results are integrated to make the final watermark detection judgment. The combination of multiple technical means can improve the reliability of the detection results, and thus can better adapt to various watermark styles and embedding methods, and improve the adaptability and flexibility of detection.

[0094] Further, as Figure 1 In a specific implementation of the method, the present application provides a video image watermark detection device, such as Figure 4 As shown, the device includes: an acquisition module 401 , a first identification module 402 , a second identification module 403 , and an integration module 404 .

[0095] The acquisition module 401 is used to acquire a video image to be detected, wherein the video image to be detected is a generative artificial intelligence video image;

[0096] The first recognition module 402 is used to extract the target watermark feature of the video image to be detected, and use the pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result;

[0097] The second recognition module 403 is used to extract the target metadata and target text data of the video image to be detected, and recognize the target metadata and the target text information respectively to obtain a metadata recognition result and a text recognition result;

[0098] The integration module 404 is used to integrate the template recognition result, the metadata recognition result and the text recognition result according to the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result, so as to obtain the target watermark identification detection result.

[0099] In a specific application scenario, the first recognition module 402 is used to perform edge detection and texture analysis on the video image to be detected, determine the edge information and texture features corresponding to the video image to be detected; determine the color histogram corresponding to the video image to be detected, and extract the deep features of the video image to be detected based on a pre-trained deep neural network, and determine the target watermark features according to the edge information, the texture features, the color histogram and the deep features; use a pre-constructed watermark identification template library to perform template watermark recognition on the target watermark features to obtain a template recognition result, wherein the watermark identification template library stores multiple specified watermark features, and the types of the specified watermark features include but are not limited to text watermark types and image watermark types.

[0100] In a specific application scenario, the first recognition module 402 is used to determine the first feature vector corresponding to the target watermark feature, determine the second feature vector of the specified watermark feature for each specified watermark feature, calculate the similarity between the first feature vector and the second feature vector based on a template matching algorithm, and associate the similarity with the specified watermark feature; compare the similarity associated with each specified watermark feature with a preset similarity threshold; if there is a specified watermark feature with a similarity greater than the similarity threshold, determine the watermark identification template corresponding to the specified watermark feature, and determine a measurable function value indicating the existence of a watermark identification, and generate the template recognition result according to the measurable function value and the watermark information of the watermark identification template, wherein the watermark information is used to indicate the generation of an artificial intelligence large model for the video image to be detected; if there is no specified watermark feature with a similarity greater than the similarity threshold, determine a measurable function value indicating the absence of a watermark identification, and generate the template recognition result according to the measurable function value.

[0101] In a specific application scenario, the second identification module 403 is used to extract the target metadata of the video image to be detected, parse the target metadata to obtain metadata information, and the metadata information includes but is not limited to EXIF ​​data, creation date, and device information; the metadata information is identified using the predefined metadata format standard, and the metadata format standard records standard fields corresponding to multiple artificial intelligence models; if there is a standard field in the metadata information, the artificial intelligence model corresponding to the standard field and a measurable function value indicating the presence of a watermark identifier are determined, and the metadata identification result is generated according to the measurable function value and the model identifier of the artificial intelligence model; if there is no standard field in the metadata information, the measurable function value indicating the absence of a watermark identifier is determined, and the metadata identification result is generated according to the measurable function value.

[0102] In a specific application scenario, the second recognition module 403 is used to extract the target text information from the video image to be detected using optical character recognition technology; to recognize keywords in the target text information using natural language processing technology; if keywords related to the watermark are recognized, an artificial intelligence model matching the keywords and a measurable function value indicating the presence of a watermark are determined, and the text recognition result is generated based on the measurable function value and the model identifier of the artificial intelligence model; if keywords related to the watermark are not recognized, a measurable function value indicating the absence of a watermark is determined, and the text recognition result is generated based on the measurable function value.

[0103] In a specific application scenario, the second recognition module 403 is used to identify the target text information based on a keyword matching algorithm and using predefined watermark keywords; if the recognition result indicates that the target text information contains a watermark keyword, then an artificial intelligence model that matches the watermark keyword and a measurable function value indicating the existence of a watermark identifier are determined, and the text recognition result is generated based on the measurable function value and the model identifier of the artificial intelligence model; if the recognition result indicates that the target text information does not contain a watermark keyword, then a measurable function value indicating the absence of a watermark identifier is determined, and the text recognition result is generated based on the measurable function value.

[0104] In a specific application scenario, the integration module 404 is used to obtain training data, wherein the training data includes training samples, detection results of the training samples and actual labels of the training samples, wherein the actual labels are used to indicate whether a watermark exists; the template recognition results, metadata recognition results and text recognition results in the detection results are combined into a feature vector, and a machine recognition model is used to learn the relationship between the feature vector and the actual label to obtain an initial weight corresponding to each recognition result, and based on a cross-validation method, the training data is continuously obtained to adjust the initial weight of each recognition result to obtain the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result respectively; the predictable functions in the template recognition result, the metadata recognition result and the text recognition result are respectively determined, and each predictable function is integrated according to the weight corresponding to each recognition result to obtain the target watermark identification detection result.

[0105] The device provided by the embodiment of the present application can first obtain a video image to be detected, and the video image to be detected is a generative artificial intelligence video image. Further, the target watermark feature of the video image to be detected is extracted, and the target watermark feature is identified by a pre-built watermark identification template library to obtain a template identification result. Next, the target metadata and target text data of the video image to be detected are extracted, and the target metadata and target text information are identified respectively to obtain metadata identification results and text identification results. Finally, according to the weights corresponding to the template identification results, metadata identification results and text identification results, the template identification results, metadata identification results and text identification results are integrated to obtain the target watermark identification detection result. Different watermark styles and embedding methods may require different identification methods. The embodiment of the present application provides a variety of technical means for watermark identification detection. There is complementarity between different technical means. Template-based identification results, metadata analysis results and text information extraction results are integrated to make a final watermark detection judgment. The combination of multiple technical means can improve the reliability of the detection results, and thus can better adapt to various watermark styles and embedding methods, and improve the adaptability and flexibility of detection.

[0106] It should be noted that for other corresponding descriptions of the functional units involved in the video image watermark detection device provided in the embodiment of the present application, reference can be made to Figures 1 to 3 The corresponding description in will not be repeated here.

[0107] In an exemplary embodiment, see Figure 5 , and also provides a device, which includes a communication bus, a processor, a memory and a communication interface, and may also include an input and output interface and a display device, wherein each functional unit can communicate with each other through the bus. The memory stores a computer program, and the processor is used to execute the program stored in the memory and execute the watermark detection method of the video image in the above embodiment.

[0108] A computer-readable storage medium stores a computer program, which implements the steps of the video image watermark detection method when executed by a processor.

[0109] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by hardware, or by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0110] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily required for implementing the present application.

[0111] Those skilled in the art will appreciate that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the implementation scenario description, or can be changed accordingly and located in one or more devices different from the implementation scenario. The modules in the above implementation scenario can be combined into one module, or can be further split into multiple sub-modules.

[0112] The above application serial numbers are for description only and do not represent the advantages or disadvantages of the implementation scenarios.

[0113] The above disclosure only discloses several specific implementation scenarios of the present application. However, the present application is not limited thereto, and any changes that can be conceived by technicians in this field should fall within the scope of protection of the present application.

Claims

1. A method for detecting watermarks in video images, characterized in that: include: Acquire a video image to be detected, where the video image to be detected is a generative artificial intelligence video image; Extracting the target watermark feature of the video image to be detected, and using a pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result; Extracting target metadata and target text data of the video image to be detected, and respectively identifying the target metadata and the target text information to obtain a metadata recognition result and a text recognition result; According to the weights respectively corresponding to the template recognition result, the metadata recognition result and the text recognition result, the template recognition result, the metadata recognition result and the text recognition result are integrated to obtain a target watermark identification detection result.

2. The method according to claim 1, characterized in that: The step of extracting the target watermark feature of the video image to be detected, and using a pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result includes: Performing edge detection and texture analysis on the video image to be detected to determine edge information and texture features corresponding to the video image to be detected; Determine a color histogram corresponding to the video image to be detected, extract deep features of the video image to be detected based on a pre-trained deep neural network, and determine the target watermark feature according to the edge information, the texture feature, the color histogram and the deep feature; A pre-constructed watermark identification template library is used to perform template watermark recognition on the target watermark feature to obtain a template recognition result, wherein the watermark identification template library stores multiple specified watermark features, and the types of the specified watermark features include but are not limited to text watermark types and image watermark types.

3. The method according to claim 2, characterized in that The step of using the pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result includes: Determine a first feature vector corresponding to the target watermark feature, determine a second feature vector of each designated watermark feature, calculate a similarity between the first feature vector and the second feature vector based on a template matching algorithm, and associate the similarity with the designated watermark feature; Comparing the similarity associated with each specified watermark feature with a preset similarity threshold; If there is a designated watermark feature with a similarity greater than the similarity threshold, then determine the watermark identification template corresponding to the designated watermark feature, and determine a measurable function value indicating the existence of the watermark identification, and generate the template recognition result according to the measurable function value and the watermark information of the watermark identification template, wherein the watermark information is used to indicate the artificial intelligence large model for generating the video image to be detected; If there is no designated watermark feature with a similarity greater than the similarity threshold, a measurable function value indicating that there is no watermark identifier is determined, and the template recognition result is generated according to the measurable function value.

4. The method according to claim 1, characterized in that The step of extracting target metadata of the video image to be detected, identifying the target metadata, and obtaining a metadata identification result includes: Extracting target metadata of the video image to be detected, parsing the target metadata to obtain metadata information, wherein the metadata information includes but is not limited to EXIF ​​data, creation date, and device information; The metadata information is identified using the predefined metadata format standard, wherein the metadata format standard records standard fields corresponding to a plurality of artificial intelligence models; If a standard field exists in the metadata information, determining an artificial intelligence model corresponding to the standard field and a measurable function value indicating the presence of a watermark identifier, and generating the metadata recognition result according to the measurable function value and the model identifier of the artificial intelligence model; If the standard field does not exist in the metadata information, a measurable function value indicating that there is no watermark identifier is determined, and the metadata identification result is generated according to the measurable function value.

5. The method according to claim 1, characterized in that The step of extracting target text information of the video image to be detected, identifying the target text information, and obtaining a text recognition result includes: Using optical character recognition technology to extract the target text information from the video image to be detected; Using natural language processing technology to identify keywords in the target text information; If a keyword related to the watermark is identified, an artificial intelligence model matching the keyword and a measurable function value indicating the presence of the watermark are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model; If no keyword related to the watermark is identified, a measurable function value indicating that the watermark does not exist is determined, and the text recognition result is generated according to the measurable function value.

6. The method according to claim 1, characterized in that The step of extracting target text information of the video image to be detected, identifying the target text information, and obtaining a text recognition result includes: Based on a keyword matching algorithm, the target text information is identified using predefined watermark keywords; If the recognition result indicates that a watermark keyword exists in the target text information, an artificial intelligence model matching the watermark keyword and a measurable function value indicating the presence of a watermark identifier are determined, and the text recognition result is generated according to the measurable function value and the model identifier of the artificial intelligence model; If the recognition result indicates that the watermark keyword does not exist in the target text information, a measurable function value indicating that the watermark identifier does not exist is determined, and the text recognition result is generated according to the measurable function value.

7. The method according to claim 1, characterized in that The step of integrating the template recognition result, the metadata recognition result and the text recognition result according to the weights respectively corresponding to the template recognition result, the metadata recognition result and the text recognition result to obtain the target watermark identification detection result includes: Acquire training data, the training data including training samples, detection results of the training samples and actual labels of the training samples, the actual labels are used to indicate whether a watermark exists; The template recognition result, metadata recognition result and text recognition result in the detection result are combined into a feature vector, and a machine recognition model is used to learn the relationship between the feature vector and the actual label to obtain an initial weight corresponding to each recognition result, and based on a cross-validation method, training data is continuously obtained to adjust the initial weight of each recognition result to obtain weights corresponding to the template recognition result, the metadata recognition result and the text recognition result respectively; The predictable functions in the template recognition result, the metadata recognition result and the text recognition result are determined respectively, and each predictable function is integrated according to the weight corresponding to each recognition result to obtain the target watermark identification detection result.

8. A video image watermark detection device, characterized in that: include: An acquisition module, used to acquire a video image to be detected, wherein the video image to be detected is a generative artificial intelligence video image; The first recognition module is used to extract the target watermark feature of the video image to be detected, and use the pre-built watermark identification template library to perform template watermark recognition on the target watermark feature to obtain a template recognition result; The second recognition module is used to extract the target metadata and target text data of the video image to be detected, and recognize the target metadata and the target text information respectively to obtain a metadata recognition result and a text recognition result; The integration module is used to integrate the template recognition result, the metadata recognition result and the text recognition result according to the weights corresponding to the template recognition result, the metadata recognition result and the text recognition result respectively, so as to obtain the target watermark identification detection result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images