Image matching technology-based film and television infringement evaluation method and system

Abstract

The invention relates to the technical field of film and television infringement evaluation, in particular to a film and television infringement evaluation method and system based on an image matching technology. The method comprises the following steps: collecting to-be-evaluated film and television image data; through constructing an adaptive multi-scale convolution denoising network and an image enhancement algorithm, film and television image data are subjected to denoising, enhancement and cutting processing so that image quality is improved and sizes are unified. Extracting image color, texture and shape features, constructing a multi-modal convolution feature extraction network, and improving the matching accuracy by using an adaptive weighted image matching algorithm; after image matching, similarity is calculated, and infringement probability evaluation is carried out in combination with legal standards; an evaluation result is displayed through a report, and copyright party processing suggestions are provided; in addition, a template image matching method is introduced, and first-order and second-order matching codes are used for carrying out rapid preliminary matching on the images, so that the high efficiency and accuracy of evaluation are ensured. The method comprehensively considers image quality, feature matching and legal standards, and has high practicability and flexibility.

Description

Technical Field

[0001] The present invention relates to the technical field of film and television infringement evaluation, and specifically relates to a film and television infringement evaluation method and system based on image matching technology. Background Art

[0002] With the rapid development of digital content, the copyright protection problem of film and television works has become increasingly prominent. Especially on Internet platforms, piracy and infringement behaviors emerge in an endless stream. The infringement of film and television works not only seriously affects the economic interests of creators and copyright holders, but also causes a vicious cycle in content creation. Traditional copyright protection and infringement detection methods often rely on manual review, which is inefficient, costly and prone to overlooking some subtle infringement behaviors.

[0003] A Chinese invention patent with the publication number CN112149744B discloses a method for quickly identifying infringing pictures, including: preprocessing the pictures to be tested in a picture library; B): calculating the histogram similarity between the preprocessed pictures to be tested and the copyright pictures. If the histogram similarity is greater than the first threshold, it is determined that the preprocessed pictures to be tested are key pictures; C): obtaining the feature grids of the key pictures and the copyright pictures, comparing the image similarity within the feature grids, obtaining similar feature grids, and taking the key pictures with the number of similar feature grids greater than the grid number threshold as suspected infringing pictures; D): calculating the similarity between the grayscale pictures of the suspected infringing pictures and the grayscale pictures of the copyright pictures, and determining the suspected infringing pictures with the similarity greater than the threshold as infringing pictures. The substantial effect of the present invention is: through the preprocessing, overall detection, local detection and precise detection of pictures, picture infringement determination is carried out, the efficiency of identifying infringing pictures is improved, and the objectivity of picture infringement determination is enhanced.

[0004] Due to the diversity of the forms of expression of film and television works, infringing content is often hidden in complex scenes and details, and traditional technical means are often difficult to effectively identify. This makes it difficult for copyright holders to conduct comprehensive monitoring and protection in the face of a large amount of data, further exacerbating the dilemma of copyright protection. Therefore, how to quickly and accurately identify infringement behaviors in a large amount of film and television content and improve the efficiency and accuracy of infringement detection has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of the present invention is to address the problems in the background art and propose a film and television infringement evaluation method and system based on image matching technology.

[0006] The technical solution of the present invention: A film and television infringement evaluation method based on image matching technology includes the following specific implementation steps: S1. Obtain the image data of the film and television works to be evaluated; S2. Construct an adaptive multi-scale convolutional denoising network. Through multi-scale convolution, noise adaptive weight adjustment, and denoising fusion, improve the accuracy of image feature extraction and matching. Combine enhanced contrast, ROI cropping, and size unification to optimize the efficiency of film and television infringement assessment; S3. By constructing a multi-modal convolutional neural network, extract color, texture, and shape features respectively, and dynamically calculate their weights for image analysis. After extracting features using convolutional operations, fuse different features to generate the final representation of the image, construct an adaptive weighted image matching model, select the image to be evaluated and the template image, and calculate the feature similarity; S4. Construct an infringement assessment model. Use the image matching similarity and legal standards to calculate the infringement probability. Use the Sigmoid function to transform the similarity, and judge whether there is infringement according to the set threshold. When the infringement probability exceeds the threshold, output the result of "infringement", otherwise it is "no infringement"; S5. Automatically generate an infringement assessment report, including the infringement probability, information on the infringed work, and infringement details; S6. Display the infringement assessment report through the user interface, provide information on the infringing work and handling suggestions. If the infringement probability is high, that is, exceeds 95%, it is recommended to give a warning or legal treatment, and provide specific corrective measures; S7. According to the data update of the monitoring platform or partners, regularly download the image data of newly released film and television works from the film and television database and add it to the database.

[0007] Preferably, the adaptive multi-scale convolutional denoising network includes: Input image layer: The input image size is , and feature extraction is performed through an initial convolutional layer; Multi-scale convolutional layer: Use convolutional kernels of different sizes to perform convolution on the image to extract features at different scales; at each scale, after the convolution operation, an activation function and batch normalization are performed; Adaptive weight adjustment layer: Each convolutional layer of the network adaptively adjusts the weight according to the noise distribution in the image. During the denoising process, regions with more noise, that is, blurred regions in the image, are given higher weights to enhance the denoising effect; while in the detail regions, the weights are lower to avoid losing the details of the image. The calculation formula for the adaptive weight is as follows: ; ; In the formula, represents the adaptive weight at position in the th scale; represents the saliency of the image at position , and is obtained through the image gradient is calculated from the amplitude; and respectively represent the gradients of the image in the and directions; represents the adjustment coefficient, controlling the sensitivity of the noise weight; Denoising fusion layer: Fuses the denoising information from different scale convolutional layers. By means of weighted averaging, the denoising results of different scales are combined to obtain a comprehensive denoising result. The fusion formula is: ; In the formula, represents the denoising result of the th scale; Output layer: The final output image is generated through a convolutional layer to obtain the denoised image .

[0008] Preferably, the denoising process is as follows: S31. Define the noisy image: ; In the formula, represents the noisy image; represents the original image; represents the additive noise; S32. Train the adaptive multi-scale convolutional denoising network, and optimize it using the L2 loss function: ; In the formula, represents the loss function; represents the denoised image; S33. The adaptive multi-scale convolutional denoising network performs supervised learning through a large number of noisy image and clear image pairs, and is optimized using the Adam optimizer. The adaptive multi-scale convolutional denoising network is trained by minimizing the L2 loss function.

[0009] Preferably, the fusion process of fusing different features to generate the final representation of the image is as follows: S41. Calculate the dynamic weights of color, texture, and shape features respectively: ; ; ; ; ; ; In the formula, represents the probability of color in the image; Indicates the number of color categories; Indicates the gradient value of the image at position ; Indicates the number of all pixel points in the image; Indicates the mean value of the gradient value; Indicates the second-order gradient of the image at position ; Indicates the unit step function. If indicates an edge, then = 1, otherwise 0; , and respectively represent the weights of color, texture, and shape features; S42. The color feature extraction processes the input image through a convolution operation. The formula is as follows: ; In the formula, Indicates the color feature of the image at position ; Indicates the weight of the color convolution kernel; Indicates the RGB or Lab color value of the image at position ; Indicates the radius of the convolution kernel; S43. The texture feature is extracted by calculating the magnitude of the image gradient. The formula is as follows: ; In the formula, Indicates the texture feature of the image at position ; Indicates the weight of the texture convolution kernel; Indicates the gradient value of the image at position , that is, edge or texture information; S44. The extraction of the shape feature is completed by calculating the second-order gradient of the image. The formula is as follows: ; In the formula, Indicates the shape feature of the image at position ; Indicates the weight of the shape convolution kernel; Indicates the second-order gradient value of the image at position ; S45. The extracted color, texture, and shape features are weighted and fused to synthesize the final feature representation of the image. The fusion method is: ; In the formula, Indicates the final multi-modal feature of the image; , and represent the weights of color, texture, and shape features respectively.

[0010] Preferably, the calculation process for calculating feature similarity is as follows: S51. Based on the matching method between the image to be evaluated and the template image, select two images to be evaluated and ; Among them, is the image to be evaluated; is the selected template image in the database; S52. Obtain the feature vectors and of the two images and to be matched, and calculate the similarity and of the images : ; In the formula, and represent the L2 norm of the vector.

[0011] Preferably, the implementation process of the matching method based on the image to be evaluated and the template image is as follows: S61. The user customizes and selects the template image , and automatically generates a first-order matching code for the template image selected by the user; Among them, is the generator of the cyclic group ; represents a bilinear mapping, ; is a cyclic group; S62. Automatically select the image to be evaluated, and automatically generate a second-order matching code , and its generation process is as follows: S6201. Encode the image to be evaluated to obtain binary data Image; S6202. Select a random number , and calculate the matching parameter C: ; In the formula, and represent randomly selected matching code generation factors, , ; is a cyclic group The order of; is the integer ring modulo ; S6203. Calculate the second-order matching code :

[0012] ; S63. If , it indicates that the image to be evaluated matches the template image successfully. Then, infringement assessment is performed, and the assessment result is fed back to the user.

[0013] Preferably, the calculation process of the infringement probability is as follows: ; ; In the formula, represents the similarity between the images and ; represents the probability of infringement, with the value range being [0, 1]. 0 indicates no infringement, and 1 indicates complete infringement; and represent parameters set according to the legal standards or industry rules of infringement assessment, used to adjust the impact of the matching similarity on infringement assessment; represents function.

[0014] Preferably, the ROI cropping process is as follows: Based on the automatic selection of the salient region, select the visual saliency based on the image, that is, crop by detecting the region in the image that attracts the most attention. The formula is: ; In the formula, represents the saliency of the image at the position , calculated through the magnitude of the image gradient ; and respectively represent the gradients of the image in the x and y directions.

[0015] The technical solution of the present invention: A film and television infringement assessment system based on image matching technology, which is used to execute the above-mentioned film and television infringement assessment method based on image matching technology, includes: An image acquisition module, used to acquire images in the target film and television works through a camera, screen capture software, or other devices; An image preprocessing module, used to perform preprocessing operations on the acquired images; An image matching module for extracting and matching features of images of target film and television works using an image matching algorithm and comparing the similarity between images; An infringement identification module for analyzing whether the images in a film and television work are similar or repeated to the existing content in other film and television works based on the image matching result and determining whether there is an infringement; A result evaluation module for evaluating the infringement result and generating an infringement evaluation report; A user interface module for providing a user interaction interface, displaying the evaluation result and relevant suggestions, and facilitating user understanding and processing; A database for storing template images for matching.

[0016] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: The present invention designs a film and television infringement evaluation method and system based on image matching technology. Through multi-dimensional image processing, feature fusion, dynamic evaluation, and a flexible matching mechanism, it effectively improves the accuracy and efficiency of film and television infringement evaluation, and while ensuring the accuracy of the evaluation result, provides higher system adaptability and flexibility, which makes it have significant technical advantages and promotion value in practical applications: (1) Efficient and accurate film and television infringement evaluation: Through image matching technology, combined with multi-modal feature extraction and an adaptive weighted matching mechanism, it can effectively improve the accuracy and efficiency of film and television infringement evaluation; By using a multi-scale convolutional denoising network (AMC-DNNet) and an adaptive weighted image matching method, it can better process complex image data, ensure accurate matching of images under different scales and noise conditions, and thus improve the robustness of the system; (2) Adaptive image preprocessing: Through operations such as denoising, enhancement, and cropping, the features of the image can be better highlighted in subsequent matching, improving the reliability of image matching; In particular, using an adaptive multi-scale convolutional network (AMC-DNNet) for denoising can dynamically adjust the denoising strategy according to the noise characteristics of the image, ensuring the retention of details rather than blind denoising, and significantly improving the effect of subsequent feature extraction; (3) Multi-modal feature fusion: By constructing a multi-modal convolutional neural network, it can extract image features from multiple dimensions such as color, texture, and shape, and dynamically adjust the weights according to the matching degree of the features; This feature fusion method enables image matching to consider not only single-dimensional features but also comprehensive types of features, thus more comprehensively reflecting the similarity of images and improving the accuracy of film and television infringement evaluation; (4) Dynamic infringement assessment mechanism: By constructing an infringement assessment model based on image similarity and introducing an infringement probability and dynamic threshold adjustment mechanism, it can flexibly adjust the infringement assessment criteria according to the actual situation, enhancing the adaptability and flexibility of the system. By introducing the Sigmoid function model and the adjustment coefficient based on industry standards, the system can accurately judge and evaluate film and television infringement behaviors according to actual legal standards and industry needs; (5) Fast template image matching: By providing users with flexible template image selection and first-order and second-order matching code generation technologies, by quickly generating matching codes and comparing the matching codes, it is determined whether the image to be evaluated matches the template image; This efficient matching method not only improves the matching speed but also enhances the accuracy of the evaluation, ensuring the matching reliability between the image to be evaluated and the template image;

[0017] (6) System design with strong adaptability: The present invention not only has high-precision film and television infringement assessment capabilities but also has good adaptability and can process film and television data from different sources, including Internet platforms, film and television databases, and custom video sources, meeting the needs for diverse data sources in practical applications. Brief Description of the Drawings

[0018] Figure 1 It is the system architecture diagram of a film and television infringement assessment system based on image matching technology proposed by the present invention; Figure 2 It is the method flow diagram of a film and television infringement assessment method based on image matching technology proposed by the present invention. Detailed Embodiments Embodiment 1, as Figure 1 shown, a film and television infringement assessment system based on image matching technology proposed by the present invention includes: an image acquisition module, an image preprocessing module, an image matching module, an infringement identification module, a result evaluation module, a user interface module, and a database. The image acquisition module acquires images in the target film and television works through a camera, screen capture software, or other devices; The image preprocessing module performs preprocessing operations on the acquired images, including but not limited to denoising, enhancement, and cropping; The image matching module uses image matching algorithms to extract features and match the images of the target film and television works, comparing the similarity between the images; The infringement identification module analyzes whether the images in the film and television works are similar or repeated to the existing content in other film and television works according to the image matching results, and judges whether there is infringement; The result evaluation module evaluates the infringement results and generates an infringement evaluation report, including but not limited to infringement probability and information on the infringed works; The user interface module provides a user interaction interface to display the evaluation results and relevant suggestions, facilitating user understanding and processing; The database is used to store the template images for matching.

[0019] Example 2, as Figure 2 shown, a film and television infringement evaluation method based on image matching technology proposed by the present invention is applied to a film and television infringement evaluation system based on image matching technology proposed in Example 1. The specific implementation steps are as follows: S1. Obtain the image data of the film and television works to be evaluated through the image acquisition module. The data sources include but are not limited to Internet platforms, film and television databases, or film and television video content obtained through screen capture tools. The specific steps include: S11. Connect to the film and television platform or database through the interface to automatically obtain the uploaded or updated film and television video data; S12. For custom video sources, a screen capture tool can be used to extract key frame images from the video player; S13. When collecting data, according to the target video content information, including but not limited to duration, chapter, and resolution, extract representative image data segments; S14. Transmit the collected image data to the image preprocessing module.

[0020] S2. The image preprocessing module performs preprocessing operations on the collected images, including but not limited to denoising, enhancement, and cropping. The specific implementation process is as follows: S21. Construct an Adaptive Multi-scale Convolutional Denoising Network (AMC-DNNet). By combining a multi-scale convolutional network and adaptive noise weight adjustment, it can effectively eliminate the noise in the image while retaining details, improving the accuracy of subsequent image feature extraction and matching; The multi-scale convolutional network consists of multiple convolutional layers and pooling layers. The network structure includes: Input image layer: The input image size is , and feature extraction is performed through an initial convolutional layer; Multi-scale convolutional layer: Use convolutional kernels of different sizes to perform convolution on the image to extract features at different scales; at each scale, after the convolution operation, an activation function (ReLU) and batch normalization (Batch Normalization) are performed to improve the nonlinear ability and stability of the model; Adaptive Weight Adjustment Layer: Each convolutional layer of the network adaptively adjusts the weights according to the noise distribution in the image. During the denoising process, regions with more noise (including but not limited to blurred regions in the image) are given higher weights to enhance the denoising effect; while in the detail regions, the weights are lower to avoid losing image details. The calculation formula for the adaptive weights is as follows: ; ; In the formula, represents the adaptive weight at position in the th scale; represents the saliency of the image at position (x,y), which is calculated through the magnitude of the image gradient ; and represent the gradients of the image in the and directions respectively; represents the adjustment coefficient, which controls the sensitivity of the noise weight; Denoising Fusion Layer: Fuses the denoising information from convolutional layers of different scales. By means of weighted averaging, the denoising results of different scales are combined to obtain a comprehensive denoising result. The fusion formula is: ; In the formula, represents the denoising result of the th scale; Output Layer: The final output image is generated through a convolutional layer to obtain the denoised image I out (x,y), with the same size as the input image; Implementation details of the multi-scale convolutional network (AMC-DNNet) are as follows: S2101. Assume that the noise in the image is additive Gaussian noise, i.e.: ; In the formula, represents the noisy image; represents the original image; represents the additive noise; S2102. To train the denoising network, the L2 loss function is used for optimization:

[0021] ; In the formula, represents the loss function; S2103. The multi-scale convolutional network performs supervised learning through a large number of pairs of noisy images and clear images, and is optimized using the Adam optimizer. The network is trained by minimizing the L2 loss function; S22. Enhancing the contrast can make the details of the image more prominent, especially the edge part of the image. The enhanced image can improve the extraction effect of feature points. Through the cumulative distribution function, histogram equalization equalizes the gray-level distribution of the image, thereby enhancing the contrast of the image, which is beneficial to subsequent image matching. The formula is: ; In the formula, represents the cumulative distribution function of the gray value r ; represents the probability density of the gray value in the image; ; S23. Since the scenes of film and television works usually contain a large number of irrelevant background elements, therefore, the region of interest (ROI) is cropped to reduce unnecessary calculations and improve the efficiency of image matching. Based on the automatic selection of the salient region, the visual saliency of the image is selected, that is, the cropping is performed by detecting the region in the image that can most attract the line of sight (including but not limited to people and objects). The formula is: ; In the formula, represents the pixel value of the output image; represents the pixel value of the image; represents the interpolation weight, which is used to determine the relationship between the pixels of the original image and the target image; N and M respectively represent the width and height of the target image; S24. To ensure the comparability and matching accuracy between different images, the images are processed for size unification:

[0022] S3. The image matching module combines a convolutional neural network (CNN) and an adaptive weighted image matching mechanism to construct a multi-modal convolutional feature extraction network, extracts multi-dimensional feature information of the image, including color features, texture features, and shape features, and uses the adaptive weighted matching algorithm to adjust the weights according to the feature matching degree, and finally realizes efficient and accurate image matching. The specific implementation process is as follows: S31. Different features in the image (including but not limited to color, texture, shape) have different importance for the evaluation of film and television infringement. Therefore, a multi-modal convolutional neural network is constructed to analyze the image from different angles, and each sub-network is specifically used to extract a specific feature. Specifically as follows: S3101. Calculate the dynamic weights of color, texture, and shape features respectively: ; ; ; ; ; ; In the formula, represents the probability of color in the image; represents the number of color categories; represents the gradient value of the image at position ; represents the total number of all pixel points in the image; represents the mean value of the gradient value; represents the second-order gradient of the image at position (x i , y i ); represents the unit step function. If represents an edge, then = 1, otherwise 0; w C , w T and w S respectively represent the weights of color, texture, and shape features; S3102. The color feature is extracted by processing the input image through a convolution operation. The formula is as follows: ; In the formula, represents the color feature of the image at position ; represents the weight of the color convolution kernel; represents the RGB or Lab color value of the image at this position; k represents the radius of the convolution kernel; S3103. The texture feature is extracted by calculating the magnitude of the image gradient. The formula is as follows: ; In the formula, represents the texture feature of the image at position ; represents the weight of the texture convolution kernel; represents the gradient value of the image at position (x, y), that is, edge or texture information; S3104. The extraction of the shape feature is completed by calculating the second-order gradient of the image (i.e., the Laplacian operator). The formula is as follows: ; In the formula, represents the shape feature of the image at position Shape features at the location; Represents the weight of the shape convolution kernel; Represents the second-order gradient value of the image at position (x, y); S3105. Weightedly fuse the color, texture, and shape features extracted by the above three sub-networks to synthesize the final feature representation of the image. The fusion method is as follows: ; In the formula, F(x, y) represents the final multi-modal feature of the image; w C , w T and w S Represent the weights of the color, texture, and shape features respectively; S32. Build an adaptive weighted image matching model to dynamically adjust the weights according to the similarity between the extracted features to improve the reliability and robustness of the matching. The specific implementation process is as follows: Based on the matching method of the image to be evaluated and the template image, select two images I1 and I2 to be evaluated (I1 is the image to be evaluated and I2 is the selected template image in the database), obtain the feature vectors F1(x, y) and F2(x, y) of the two images I1 and I2 to be matched, and calculate the similarity sim(I1, I2) of the images I1 and I2: ; In the formula, and Represent the L2 norm of the vector.

[0023] S4. The infringement recognition module is based on the infringement evaluation method of the matching similarity and legal standards. It judges whether there is infringement through the similarity of the images, and can also be flexibly adjusted according to the requirements of the industry and the law. By introducing an infringement probability model and a dynamic threshold adjustment mechanism, it ensures that the film and television infringement evaluation system has higher accuracy and adaptability. The specific implementation process is as follows: S41. Build an infringement evaluation model. According to the matching similarity and legal standards, calculate the infringement probability P infringe . When the matching similarity of the image exceeds the set infringement probability threshold, it is judged as infringement: ; ; In the formula, sim(I1, I2) represents the similarity of the images I1 and I2; P infringe Represents the probability of infringement, and the value range is [0, 1]. 0 means no infringement, and 1 means complete infringement; α' and β' represent the parameters set according to the legal standards or industry rules of the infringement evaluation, which are used to adjust the influence of the matching similarity on the infringement evaluation; σ represents the Sigmoid function; S42. Output the evaluation result: When P infringe exceeds the set threshold T infringe then the result of "infringement" is output, indicating that the two images have a high degree of similarity and may constitute an infringement; Conversely, when P infringe is less than the set threshold T infringe then the result of "no infringement" is output, indicating that the similarity between the two images is not sufficient to be determined as an infringement; S43. Transmit the evaluation result to the result evaluation module.

[0024] S5. The result evaluation module automatically generates an infringement evaluation report according to the evaluation result. The report includes but is not limited to: Infringement probability value: Evaluate the possibility of infringement according to the matched similarity; Information on the infringed work: including but not limited to the name, creator, and copyright owner of the infringing work; Infringement details: including but not limited to the time, circumstances, and similar parts of the images of the infringement; Accordingly, transmit the infringement evaluation report to the user interface module.

[0025] S6. The user interface module analyzes and displays the evaluation result and the infringement evaluation report for the user to understand and take measures. Specifically, it includes: Result visualization: Display the infringement evaluation report through the user interface, including but not limited to the name of the infringed film and television work, relevant images of the infringing content, and the specific time period when the infringement occurred; Processing suggestions: Provide processing suggestions to the copyright owner according to the infringement evaluation report, including but not limited to: If the infringement probability is high (exceeding 95%), it is recommended to issue a warning or notify the infringer; if the infringement circumstances are serious, it is recommended to handle it through legal channels; Provide specific measures for infringement correction, including but not limited to taking down pirated content and deleting infringing videos.

[0026] S7. Database update: According to the data update of the monitoring platform or the partner, the image acquisition module regularly downloads the image data of newly released film and television works from the film and television database and adds it to the database.

[0027] Embodiment 3. A method for evaluating film and television infringement based on image matching technology proposed by the present invention further includes a method for matching a to-be-evaluated image with a template image. The specific implementation steps are as follows: S1. The user custom-selects the template image I2 from the database, and automatically generates a first-order matching code C match1 =e(g,g); Among them, \(g\) is the generator of the cyclic group \(G\); \(e\) represents a bilinear mapping, \(e: G\times G\rightarrow G\). T ; \(G\) T is a cyclic group.

[0028] S2. Automatically select the image \(I1\) to be evaluated and automatically generate a second-order matching code \(C\). match2 , and its generation process is as follows: S21. Encode the image \(I1\) to be evaluated to obtain binary data Image; S22. Select a random number , and calculate the matching parameter \(C\):

[0029] ; In the formula, \(fc1\) and \(fc2\) represent matching code generation factors, \(fc1\in Z\) p , \(fc2\in Z\) p ; \(p\) is the order of the cyclic group ; is the integer ring modulo ;

[0030] S23. Calculate the second-order matching code : .

[0031] S3. If , it means that the image \(I1\) to be evaluated matches the template image \(I2\) successfully. Then, perform infringement evaluation and feedback the evaluation result to the user.

[0032] Accordingly, the user can independently select the template image for matching and, based on an efficient method for matching a to-be-evaluated image with a template image, quickly perform a preliminary matching on the selected to-be-evaluated image and template image, ensuring the accuracy of the matching between the to-be-evaluated image and the template image, thereby providing a reliable matching result for the user.

[0033] The above has described in detail the embodiments of the present invention in conjunction with the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art to which the present invention pertains.

Claims

1. A film and television infringement assessment method based on image matching technology, characterized in that: The specific implementation steps include the following: S1. Obtain image data of the film and television work to be evaluated; S2. Construct an adaptive multi-scale convolution denoising network, improve the image feature extraction and matching accuracy through multi-scale convolution, noise adaptive weight adjustment and denoising fusion, and optimize the efficiency of film and television infringement assessment by combining contrast enhancement, ROI cropping and size unification; S3. By constructing a multimodal convolutional neural network, color, texture and shape features are extracted respectively, and their weights are dynamically calculated for image analysis. After extracting features using convolution operations, different features are fused to generate the final representation of the image, and an adaptive weighted image matching model is constructed. The image to be evaluated and the template image are selected to calculate feature similarity. S4. Construct an infringement assessment model, use image matching similarity and legal standards to calculate the infringement probability, use the Sigmoid function to convert the similarity, and judge whether there is infringement based on the set threshold. When the infringement probability exceeds the threshold, the "infringement" result is output, otherwise it is "no infringement"; S5. Automatically generate an infringement assessment report, including the probability of infringement, information about the infringed work, and details of the infringement; S6. Display the infringement assessment report through the user interface, provide information on infringing works and handling suggestions. If the infringement probability is high, i.e., more than 95%, recommend warning or legal action, and provide specific corrective measures; S7. According to the data update of the monitoring platform or the partner, the newly released film and television work image data is regularly downloaded from the film and television database and added to the database.

2. According to claim 1, a method for assessing film and television infringement based on image matching technology is characterized in that: The adaptive multi-scale convolutional denoising network includes: Input image layer: The input image size is W×H, and features are extracted through an initial convolution layer; Multi-scale convolution layer: Use convolution kernels of different sizes to convolve the image to extract features at different scales; at each scale, activation function and batch normalization are performed after the convolution operation; Adaptive weight adjustment layer: Each convolution layer of the network adaptively adjusts the weight according to the noise distribution in the image. In the denoising process, areas with more noise, that is, blurred areas in the image, are given higher weights to enhance the denoising effect; while in detail areas, the weights are lower to avoid losing image details. The calculation formula for adaptive weights is as follows: ； ； Where W(x,y,i) represents the adaptive weight at position (x,y) in the i-th scale; S(x,y) represents the saliency of the image at position (x,y), which is calculated by the image gradient The amplitude of is calculated; and They represent the gradient of the image in the x and y directions respectively; α represents the adjustment coefficient, which controls the sensitivity of the noise weight; Denoising fusion layer: It fuses the denoising information from convolutional layers of different scales. It combines the denoising results of different scales by weighted average to obtain a comprehensive denoising result. The fusion formula is: ； In the formula, I i (x,y) represents the denoising result of the i-th scale; Output layer: The final output image is generated through a convolution layer to obtain the denoised image I out (x,y).

3. The method for assessing film and television infringement based on image matching technology according to claim 1, characterized in that: The denoising process is as follows: S31. Define noisy image: ; In the formula, I noise (x,y) represents a noisy image; I(x,y) represents the original image; N(x,y) represents additive noise; S32, train the adaptive multi-scale convolutional denoising network and use the L2 loss function for optimization: ； Where L represents the loss function; I out (x,y) represents the denoised image; S33, the adaptive multi-scale convolutional denoising network is supervised by a large number of noisy images and clear image pairs, optimized using the Adam optimizer, and trained by minimizing the L2 loss function.

4. The method for assessing film and television infringement based on image matching technology according to claim 1, characterized in that: The fusion process of fusing different features to generate the final representation of the image is as follows: S41. Calculate the dynamic weights of color, texture, and shape features respectively: ； ； ； ； ； ； In the formula, represents the probability of color i in the image; N1 represents the number of color categories; Indicates that the image is at position (x i ,y i ) at the gradient value; N2 represents the number of all pixels in the image; Represents the mean of the gradient values; It means that the image is at position (x i ,y i ) at the second-order gradient; δ(x) represents the unit step function, if represents the edge, then =1, otherwise 0; w C 、w T and w S Represent the weights of color, texture and shape features respectively; S42, color feature extraction processes the input image through convolution operation, the formula is as follows: ； In the formula, C(x,y) represents the color feature of the image at position (x,y); W C (i, j) represents the weight of the color convolution kernel; I(x+i, y+j) represents the RGB or Lab color value of the image at position (x, y); k represents the radius of the convolution kernel; S43, texture features are extracted by calculating the amplitude of image gradient, the formula is as follows: ； Where T(x,y) represents the texture feature of the image at position (x,y); W T (x,y) represents the weight of the texture convolution kernel; Represents the gradient value of the image at position (x, y), that is, edge or texture information; S44. The shape feature extraction is completed by calculating the second-order gradient of the image. The formula is as follows: ； In the formula, Shape(x,y) represents the shape feature of the image at position (x,y); W S (x,y) represents the weight of the shape convolution kernel; Represents the second-order gradient value of the image at position (x, y); S45, weighted fusion of the extracted color, texture and shape features to synthesize the final feature representation of the image, the fusion method is: ； Where F(x,y) represents the final multimodal features of the image; w C 、w T and w S Represent the weights of color, texture and shape features respectively.

5. The method for assessing film and television infringement based on image matching technology according to claim 1, characterized in that: The calculation process for calculating feature similarity is: S51, based on the matching method between the image to be evaluated and the template image, selecting two images I1 and I2 to be evaluated; Among them, I1 is the image to be evaluated; I2 is the template image selected in the database; S52, obtain the feature vectors F1(x, y) and F2(x, y) of the two images I1 and I2 to be matched, and calculate the similarity sim(I1, I2) between the images I1 and I2: ； Where ||F1(x,y)||2 and ||F2(x,y)||2 represent the L2 norm of the vector.

6. The method for assessing film and television infringement based on image matching technology according to claim 5, characterized in that: The implementation process of the matching method based on the image to be evaluated and the template image is as follows: S61: The user selects a template image I2, and automatically generates a first-order matching code C for the template image I2 selected by the user. match1 =e(g,g); Among them, g is the generator of the cyclic group G; e represents the bilinear mapping, e: G×G→G T ; G T is a cyclic group; S62, automatically select the image I1 to be evaluated, and automatically generate the second-order matching code C match2 , and its generation process is as follows: S6201, encode the image I1 to be evaluated to obtain binary data Image; S6202, select a random number r∈Z p , calculate the matching parameter C: ； Where fc1 and fc2 represent randomly selected matching code generation factors, fc1∈Z p , fc2∈Z p ; p is the order of the cyclic group G; Z p is the ring of integers modulo p; S6203, calculate the second-order matching code C match2 : ； S63, if C match1 =C match2 , it means that the image to be evaluated I1 matches the template image I2 successfully, and then the infringement evaluation is performed and the evaluation result is fed back to the user.

7. The method for assessing film and television infringement based on image matching technology according to claim 1, characterized in that: The calculation process of infringement probability is as follows: ； ； Where sim(I1,I2) represents the similarity between images I1 and I2; P infringe It represents the probability of infringement, with a value range of [0,1], where 0 represents no infringement and 1 represents complete infringement; α' and β' represent parameters set according to the legal standards or industry rules for infringement assessment, which are used to adjust the impact of matching similarity on infringement assessment; σ represents the Sigmoid function.

8. The method for assessing film and television infringement based on image matching technology according to claim 1, characterized in that: The ROI cropping process is as follows: Based on the automatic selection of salient regions, the selection is based on the visual saliency of the image, that is, the most eye-catching area in the image is detected and cropped. The formula is: ； In the formula, S(x,y) represents the saliency of the image at position (x,y), which is calculated by image gradient The amplitude of is calculated; and Represent the gradient of the image in the x and y directions respectively.

9. A film and television infringement assessment system based on image matching technology, which is used to execute the film and television infringement assessment method based on image matching technology as described in any one of claims 1 to 8, characterized in that: include: An image acquisition module, used to acquire images from the target film or television work through a camera, screen capture software or other devices; An image preprocessing module is used to perform preprocessing operations on the collected images; The image matching module is used to extract and match the images of the target film and television works using the image matching algorithm and compare the similarity between the images; The infringement identification module is used to analyze whether the images in the film and television works are similar or duplicated to the existing content in other film and television works based on the image matching results, and determine whether there is any infringement; The result evaluation module is used to evaluate the infringement results and generate an infringement evaluation report; The user interface module is used to provide a user interaction interface to display the evaluation results and related suggestions for easy understanding and processing by users; A database is used to store template images for matching.

Images