A short video quality evaluation method based on big data analysis

Abstract

The present invention discloses a short video quality evaluation method based on big data analysis, and relates to the technical field of video quality evaluation. In the short video quality evaluation, the present invention introduces intelligent originality analysis and homogeneity detection technology. The technology performs key frame analysis on the short video, extracts key frame images and generates data fingerprints, and then compares these data fingerprints with the short video data in the big database, so as to judge the originality and homogeneity of the short video. In addition, by calculating the number of homogeneous works of short video works and the degree of homogeneity, a more detailed originality analysis report is provided to the creator. This intelligent originality analysis and homogeneity detection method not only improves the accuracy and efficiency of short video quality evaluation, but also helps to combat the infringement of plagiarism and theft, and protect the rights and interests of the original creators.

Description

Technical Field

[0001] The present invention relates to the technical field of video quality evaluation, and in particular to a short video quality evaluation method based on big data analysis. Background Art

[0002] With the rapid development of mobile Internet and the widespread popularity of smart terminals, short video platforms have sprung up like mushrooms after a rain, and short video content has exploded, becoming an important way for people to obtain information, entertain themselves, and interact socially. Short videos, with their simplicity, intuitiveness, liveliness, fun, and ease of dissemination, have attracted the participation and attention of a large number of users, and have also spawned a large group of creators and a prosperous content ecology. However, with the massive growth of short video content, how to effectively evaluate the quality of short videos has become a problem that needs to be solved urgently.

[0003] In the prior art, in terms of originality analysis, methods for evaluating the quality of short videos rely on simple text comparison or image recognition technology, which makes it difficult to accurately judge the originality level of short videos. Especially when the content of the short videos is complex and the pictures have a high degree of similarity, the existing methods often cannot effectively distinguish between original and homogeneous content.

[0004] On the other hand, in terms of identifying illegal content, existing methods often rely on keyword filtering or manual review, which is not only inefficient but also prone to missed detections or misjudgments. In particular, for illegal content that is more concealed, existing methods are often difficult to effectively identify. Therefore, existing technologies have obvious limitations in short video quality evaluation and cannot meet the needs of the rapid development of the short video industry.

[0005] To sum up, the existing technology has obvious shortcomings in short video quality evaluation. Therefore, developing a short video quality evaluation method that can integrate feature codes and key frame data fingerprints is of great significance for improving user experience, promoting content innovation, and maintaining platform order. Summary of the invention

[0006] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide a short video quality evaluation method based on big data analysis. It can accurately judge the originality level of short videos by integrating feature codes and key frame data fingerprint comparison technology, effectively protect the rights and interests of original authors, and achieve accurate positioning and efficient screening of various illegal contents, providing a solid technical guarantee for the healthy development of the short video industry.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a short video quality evaluation method based on big data analysis, the specific steps of the method are:

[0008] S100, collecting short video data, metadata and user behavior data from the short video platform, and standardizing them, unifying the video format and resolution and storing them in a large database, wherein the large database includes historical short video data, key frame data fingerprint information and data in a violation data detection library;

[0009] S200, extracting the key frame F of the short video using the key frame extraction method k , and generate key frame data fingerprint FP k ;

[0010] S300, extracting feature codes FC from the person and object identifiers in the short video, and constructing a violation data detection library, wherein the violation data detection library includes the feature code FC of each illegal short video and its corresponding violation type R j and severity weight W j , j = 1, 2, ..., N, N is the total number of violation types and for the feature code FC of all illegal short videos, the total number of corresponding violation feature codes is N total ;

[0011] S400, comparing the key frame data fingerprint with the big database, calculating the number and degree of homogeneity of short video works, and generating originality analysis results, wherein:

[0012] In the process of comparing the key frame data fingerprint with the large database, the SSIM similarity calculation method is used to locate other videos similar to the short video to be evaluated;

[0013] According to the comparison results, count and calculate the number of homogeneous works N of short videos h and the degree of homogeneity, wherein the degree of homogeneity is quantified by counting the number of similar key frames as a proportion index H1 and the number of similar video playbacks as a proportion index H2, and the originality score O is calculated by the degree of homogeneity. And its originality score O is taken as the originality analysis result;

[0014] S500, comparing the feature code in the short video with the illegal data detection library, identifying and locating the illegal content, and generating illegal content identification results, wherein:

[0015] In the process of comparing the feature code with the illegal data detection library, an approximate nearest neighbor search algorithm based on local sensitive hashing (LSH) is used to calculate the cosine similarity between the candidate illegal feature code found and the target feature code, and the number of illegal feature codes N that match the cosine similarity is counted. m ;

[0016] The illegal content identification result is determined by the illegal content identification score V. Where Nm is the number of matches of the violation signature code, N total is the total number of illegal signature codes, W j A severity weight for each violation signature;

[0017] S600, comprehensively analyzing the originality analysis results, the illegal content identification results and the basic quality evaluation index B, calculating the final quality score Q of the short video, classifying the short video into different quality levels according to the quality score, and outputting the evaluation results in a visual manner;

[0018] The quality levels include high quality, medium quality, and low quality.

[0019] Furthermore, the key frame extraction method S200 forms a plurality of continuous video segments by dividing the short video at fixed time intervals, and sequentially selects segments of a certain length from the beginning of the video, and for each segment, calculates the pixel difference between adjacent frames within the segment. Among them, F i Represents the pixel value matrix of the i-th frame, m is the number of frames in each time segment, and the color value difference of each frame at each pixel position is compared with that of the next adjacent frame. hour, is the average pixel difference. At this time, the middle frame of the clip is determined as a key frame F. By traversing the entire short video, k key frames F are extracted in this way. k , the key frame F k Represents the content change point of the short video, quantizes and encodes each key frame, generates a binary data fingerprint with a fixed length of FP, and the data fingerprint FP can uniquely identify the image feature information of the key frame. All key frames are encoded to obtain the key frame data fingerprint FP k .

[0020] Furthermore, the S300 extracts the feature code of the person identification by first detecting and locating the person in the video screen, dividing the person's face into multiple regions by analyzing the facial contour and distribution of the facial features of the person, extracting the average color feature of each region, and extracting the area size thereof, combining the color and area features of different regions, and converting them into a fixed-length digital sequence as the feature code FC of the person through encoding. rw ;

[0021] For feature code extraction of object identification, separate the object from the background, analyze the shape of the object, extract the geometric feature parameters of the object's perimeter, area, longest axis length and shortest axis length, and extract the color features of the object. Arrange the geometric and color features in sequence and convert them into binary codes as the feature code FC of the object. wt, the character and object feature codes are combined to form the unique feature code FC of the short video.

[0022] Furthermore, the SSIM similarity calculation method of S400 defines the key frame corresponding to the i-th key frame data fingerprint of the short video to be evaluated as F ki , the key frame corresponding to the jth key frame data fingerprint of similar candidate videos in the large database is defined as F tj , divide the key frame into multiple sub-blocks of the same size, and calculate the SSIM value of each sub-block Right now Among them, x and y represent the pixel values ​​of the two sub-blocks respectively, μ x , μ y is the mean of the sub-block, σ x , σ y is the standard deviation of the sub-block, σ xy is the covariance of the sub-block, c1, c2 are constants and c1 = (0.01 × L) 2 、c2=(0.03×L) 2 , L is the dynamic range of pixel values, and the overall SSIM value SSIM of the two key frames is calculated by weighted average ij ,Right now Among them, w s is the weight of the sth sub-block, and sets the similarity threshold to SSIM T , when SSIM ij >SSIM T When the two key frames are similar, the corresponding short videos are included in the similar video set and stored in the big database. The number of SSIM values ​​matching all key frames in the similar video set is counted as the number of homogeneous works N. h .

[0023] Furthermore, the number ratio index of similar key frames in S400 Among them, M i represents the number of key frames of the i-th homogeneous work, and M is the total number of key frames of the short video to be evaluated.

[0024] Furthermore, the play volume ratio indicator of similar videos in S400 Among them, P sum is the total number of views of similar videos, P avg is the average number of video views in the large database, N t is the total number of short videos in the large database.

[0025] Furthermore, the S500 approximate nearest neighbor search algorithm maps the target feature code and the violation feature code to multiple hash buckets through multiple LSH functions with different parameters, searches for similar feature codes in each hash bucket, and calculates the cosine similarity between the candidate violation feature code found and the target feature code: in, is the i-th target feature code in the short video, is the jth feature code in the illegal data detection library. When CosSim ij >CosSim T When CosSim T is the cosine similarity threshold, which determines whether the target signature matches the violation signature.

[0026] Furthermore, the basic quality evaluation indicators in S600 include the playback volume P, the number of likes L, the number of comments C, and the completion rate W in the short video data, metadata, and user behavior data, and the corresponding weights are w p 、w l 、w c 、w w , and set the maximum values ​​of playback volume, likes, and comments to P max , L max , C max , then the basic quality evaluation index

[0027] Furthermore, the quality score of the S600 short video is Q=B×100×k, where k is a comprehensive adjustment coefficient and The final quality score When Q>Q1, the short video is of high quality, when Q2<Q<Q1, the short video is of medium quality, and when Q<Q2, the short video is of low quality, wherein Q2 and Q1 are short video quality level thresholds.

[0028] Compared with the prior art, this short video quality evaluation method based on big data analysis has the following beneficial effects:

[0029] 1. The present invention introduces intelligent originality analysis and homogeneity detection technology in the short video quality evaluation. This technology performs key frame analysis on the short video, extracts key frame images and generates data fingerprints, and then compares these data fingerprints with the short video data in the big database to judge the originality and homogeneity of the short video. In addition, it also provides creators with a more detailed originality analysis report by calculating the number of homogeneous works of short video works and the degree of homogeneity. This intelligent originality analysis and homogeneity detection method not only improves the accuracy and efficiency of short video quality evaluation, but also helps to combat plagiarism and theft and protect the rights and interests of original authors.

[0030] 2. The present invention adopts a technology based on feature code comparison in violation analysis and originality analysis. For violation analysis, accurate identification of illegal content is achieved by comparing the feature codes of people and objects in short videos with the feature codes in the violation data detection library. In originality analysis, the degree of homogeneity and originality of the video are effectively judged by comparing the key frame data fingerprints of the short video with the fingerprints in the large database. This method based on feature code comparison greatly improves the efficiency and accuracy of the analysis and reduces the subjectivity of human judgment.

[0031] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 This is an operation diagram of a short video quality evaluation method based on big data analysis;

[0034] Figure 2 The figure is a flowchart of the steps of a short video quality evaluation method based on big data analysis. DETAILED DESCRIPTION

[0035] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0036] Embodiment 1

[0037] like Figure 1 As shown, this embodiment describes in detail the application process of a short video quality evaluation method based on big data analysis. The method covers multiple steps from data collection to the output of the final quality evaluation result, through which a comprehensive evaluation of the short video quality is achieved.

[0038] First, enter the data collection and preprocessing stage (S100), collect short video data, metadata and user behavior data from the short video platform, short video data includes the video file itself, metadata covers title, tag, release time, author information, user behavior data includes likes, comments, reposts, and viewing time, and standardize the collected data to unify the video format and resolution, and store them in a large database. This large database not only stores the data collected this time, but also includes historical short video data, key frame data fingerprint information and data in the violation data detection library. Through such preprocessing, subsequent analysis can be carried out on the basis of unified data, avoiding analysis errors caused by problems such as inconsistent data formats.

[0039] Then, the key frame extraction and data fingerprint generation stage (S200) is entered, and the key frame of the short video is extracted using the key frame extraction method. The specific process is to divide the short video into a fixed time interval to form a plurality of continuous video segments. Starting from the beginning of the video, segments of a certain length are selected in turn. For each segment, the pixel difference between adjacent frames within it is calculated. Among them, F i represents the pixel value matrix of the i-th frame, m is the number of frames in each time segment, and by comparing the color value difference of each frame with its adjacent next frame at each pixel position, when D>D (D is the average pixel difference), the middle frame of the segment is determined as a key frame F, and k key frames F are extracted by traversing the entire short video. k , these key frames represent the content change points of the short video. The key frame extraction method can effectively capture the key scenes and content changes in the video, and can focus more on the dynamic changes of the video content, avoiding the problem of missing important information due to uniform sampling. For example, in short videos of fast-moving scenes and slowly changing scenes, uniform sampling cannot accurately capture the key moments in fast-moving scenes. The method can adaptively extract key frames according to content changes, so as to better reflect the overall characteristics of the video. Each key frame is quantized and encoded to generate a binary data fingerprint with a fixed length of FP. The data fingerprint FP can uniquely identify the image feature information of the key frame. All key frames are encoded to obtain the key frame data fingerprint FP. kThis data fingerprint generation method enables the key frame features to be represented in a concise and unique way, which is convenient for fast comparison and retrieval in large databases. At the same time, it is robust to common image transformations. Even if the video undergoes conventional editing operations, such as brightness adjustment, contrast adjustment, etc., it can still accurately identify the key frame features, thereby ensuring the accuracy of originality analysis and homogeneity detection.

[0040] Then, the feature code extraction and violation data detection library construction stage (S300) is entered, and feature codes FC are extracted from the person and object identification in the short video, wherein the character identification feature code extraction: when extracting feature codes from the person identification in the short video, first detect and locate the person in the video screen, divide the person's face into multiple areas by analyzing the facial contour and facial features of the person, extract the average color feature of each area, and extract its area size at the same time, combine the color and area features of different areas, and convert them into a fixed-length digital sequence through encoding as the character's feature code FC rw This character feature code extraction method can accurately capture the key features of a character, which is of great significance for identifying the identity of a character in a video and judging whether there is infringement or other violations. For example, in a short video involving celebrity portrait rights, by extracting the character feature code and comparing it with the known celebrity feature code, it is possible to quickly and accurately judge whether there is infringement. At the same time, dividing the character's face into multiple areas and extracting multiple features can improve the accuracy and uniqueness of the feature code and avoid misjudgment due to the limitations of a single feature; the object identification feature code extraction: for the feature code extraction of the object identification, the object is separated from the background, the shape of the object is analyzed, and the geometric feature parameters of the object's circumference, area, longest axis length and shortest axis length are extracted. At the same time, the color features of the object are extracted, and the geometric and color features are arranged in sequence and converted into binary codes as the feature code FC of the object. wt , combining the character and object feature codes to form a unique feature code FC of the short video. This object feature code extraction method can comprehensively describe the features of the object, which plays an important role in identifying specific objects in the video and determining whether there are any violations related to the object. At the same time, a violation data detection library is constructed, which includes the feature code FC of each illegal short video and its corresponding violation type R j and severity weight W j (j=1,2,…,N, N is the total number of violation types and for the feature code FC of all illegal short videos, the total number of violation feature codes corresponding to it is N total), the construction of this violation data detection library provides support for the identification of illegal content. By continuously updating and improving the data in the library, the ability to identify various violations can be improved. As new forms of violations emerge, relevant feature codes and violation types can be added to the library in a timely manner, so that these new violations can be identified and handled in a timely manner.

[0041] Next, we enter the originality analysis and homogeneity detection stage (S400), compare the key frame data fingerprint with the big database, calculate the number and degree of homogeneity of short video works, and generate the originality analysis results. In the comparison process, the SSIM similarity calculation method is used to define the key frame corresponding to the i-th key frame data fingerprint of the short video to be evaluated as F ki , the key frame corresponding to the jth key frame data fingerprint of similar candidate videos in the large database is defined as F tj , divide the key frame into multiple sub-blocks of the same size, and calculate the SSIM value of each sub-block Among them, x and y represent the pixel values ​​of the two sub-blocks respectively, μ x , μ y is the mean of the sub-block, σ x , σ y is the standard deviation of the sub-block, σ xy is the covariance of the sub-block, c1, c2 are constants and c1 = (0.01 × L) 2 、c2=(0.03×L) 2 , L is the dynamic range of pixel values, and the overall SSIM value of the two key frames is calculated by weighted average Among them, w s is the weight of the sth sub-block, and sets the similarity threshold to SSIM T , when SSIM ij >SSIM T When the two key frames are similar, the corresponding short videos are included in the similar video set and stored in the big database. The number of SSIM values ​​matching all key frames in the similar video set is counted as the number of homogeneous works N. h This SSIM similarity calculation method has high accuracy and effectiveness in key frame comparison. It comprehensively considers multiple factors such as the mean, standard deviation and covariance of sub-blocks, and can accurately measure the similarity between key frames, and then more accurately calculate the number and degree of homogeneous works, providing a reliable basis for originality analysis. According to the comparison results, the degree of homogeneity is calculated, and the degree of homogeneity is calculated by counting the number of similar key frames. The proportion of video views of similar videos To quantify the degree of homogeneity, M irepresents the number of key frames of the i-th homogeneous work, M is the total number of key frames of the short video to be evaluated, and P sum is the total number of views of similar videos, P avg is the average number of video views in the large database, N t is the total number of short videos in the large database. The originality score is calculated by the degree of homogeneity And take its originality score O as the result of originality analysis. This method of quantifying the degree of homogeneity comprehensively considers the two factors of the number of key frames and the volume of playback, and can more comprehensively reflect the homogeneity of short videos. By combining these two indicators, the homogeneity status of short videos in the entire short video library can be more accurately evaluated, thereby providing more reasonable results for originality analysis.

[0042] Secondly, enter the illegal content identification stage (S500), compare the feature code in the short video with the illegal data detection library, identify and locate the illegal content, and generate the illegal content identification result. In the comparison process, an approximate nearest neighbor search algorithm based on local sensitive hash LSH is adopted. Through multiple LSH functions with different parameters, the target feature code and the illegal feature code are mapped to multiple hash buckets, and similar feature codes are searched in each hash bucket. For the candidate illegal feature code found, its cosine similarity with the target feature code is calculated: Among them, FC i is the i-th target feature code in the short video, FC j is the jth feature code in the illegal data detection library. When CosSim ij >CosSim T (CosSim T is the cosine similarity threshold), determine that the target feature code matches the illegal feature code, and count the number of illegal feature codes N that match the cosine similarity. m This approximate nearest neighbor search algorithm based on local sensitive hashing (LSH) is highly efficient and accurate in identifying illegal content. By mapping feature codes to hash buckets, it can quickly find similar feature codes in a large amount of data, greatly improving search efficiency. At the same time, by calculating cosine similarity to further confirm the matching relationship, it can improve recognition accuracy and effectively avoid inefficient search and misjudgment problems caused by excessive data volume. In addition, the illegal content identification results are scored by illegal content identification. Determine, where N m is the number of matches of the violation signature code, N total is the total number of illegal signature codes, W jis the severity weight for each violation feature code. This calculation method of the violation content recognition score comprehensively considers the number of matching violation feature codes and the severity weight, and can more accurately reflect the severity of the violation content. Different violation behaviors may have different severities. By assigning different weights, the impact of the violation content on the short video quality can be more reasonably evaluated.

[0043] Finally, enter the comprehensive quality evaluation and result output stage (S600). Based on the comprehensive analysis results of originality, the recognition results of violation content, and the basic quality evaluation index B, calculate the final quality score Q of the short video. Among them, the basic quality evaluation index includes the play volume P, like count L, comment count C, and completion rate W in the short video data, metadata, and user behavior data, and their respective weights are w p 、w l 、w c 、w w . At the same time, set the maximum values of the play volume, like count, and comment count to P max 、L max 、C max , respectively. Then the basic quality evaluation index This comprehensive quality evaluation method combines the originality analysis results and the violation content recognition results with the traditional basic quality evaluation index, and can more comprehensively evaluate the quality of the short video. The basic quality evaluation index reflects the direct feedback and attention of users to the video, while the originality analysis results and the violation content recognition results reveal the quality problems of the video from a deeper perspective. By setting weights, the short video can be comprehensively evaluated according to different needs and focuses. For example, if more attention is paid to the originality and compliance of the video, the weights of the originality analysis results and the violation content recognition results can be increased; if more attention is paid to the feedback of users, the weight of the basic quality evaluation index can be increased. The comprehensive adjustment coefficient The final quality score of the short video According to the quality score, the short videos are divided into different quality levels. When Q > Q1, the short video is of high quality; when Q2 < Q < Q1, the short video is of medium quality; when Q < Q2, the short video is of low quality, where Q2 and Q1 are the thresholds of the short video quality levels. This quality level division method can intuitively reflect the quality level of the short video and provide a basis for the content management and recommendation of the short video platform. For example, high-quality short videos can be preferentially recommended to users, while low-quality short videos can remind the creators to make improvements or directly restrict their dissemination.

[0044] To sum up, this embodiment demonstrates in detail the whole process of the short video quality evaluation method based on big data analysis. By extracting key frames of short videos and generating data fingerprints, it is deeply compared with the massive short video data in the big database. At the same time, combined with the calculation of the number and degree of homogeneous works, it can accurately judge the originality level of short videos, thereby improving the accuracy and efficiency of short video quality evaluation.

[0045] Embodiment 2

[0046] like Figure 2 As shown, this embodiment demonstrates the application process of the short video quality evaluation method in infringement monitoring of short video copyright protection agencies, including each link from data collection to the output of the final quality evaluation results, reflecting the effectiveness and accuracy of the method in copyright protection.

[0047] First, enter the data collection and preprocessing stage (S100), the copyright protection agency collects a large amount of short video data and related metadata from many short video platforms. Short video data contains information such as images and audio of video files. Metadata covers title, tag, release time, author information, music work information involved in the video (such as song name, singer, copyright ownership, etc.), etc. For example, for a short video involving copyright issues, its metadata records in detail the background music used in the video as the singer's song and related copyright information. The collected data is standardized, the video format and resolution are unified, and all data are ensured to be under the same standard for subsequent analysis and processing. This process includes converting the encoding format of the video and adjusting the resolution to the set standard value. The processed data is stored in a specially constructed database, which provides a solid data foundation for subsequent analysis and avoids analysis errors caused by problems such as inconsistent data formats or differences in resolution.

[0048] Then, the key frame extraction and data fingerprint generation stage (S200) is entered, and a specific key frame extraction algorithm is used to extract key frames from the short video. The specific extraction process is shown in Example 1. By traversing the entire short video, a series of key frames are extracted. These key frames can capture the key content changes in the video. For short videos containing musical works, key frames may highlight the singer's performance movements, key changes in stage layout, and changes in the picture that match the rhythm of the music. For example, in a song video with a strong rhythm, the key frames may select the singer's passionate singing in the chorus part and the scene where the stage lights flash with the rhythm. Each key frame is quantized and encoded to generate a binary data fingerprint of a fixed length to ensure that the characteristics of the key frames can still be accurately identified after the video has undergone some routine editing operations, providing a reliable basis for subsequent comparison and analysis.

[0049] Then, the feature code extraction and violation data detection library construction phase (S300) is entered, and the character identification feature code is extracted: the feature code is extracted for the singer in the short video, and the character in the video screen is first detected and located. By analyzing the facial contour, distribution of facial features and other features of the character, the character's face is divided into multiple areas, such as the forehead, eyes, nose, mouth and other areas. For each area, the average color feature, color distribution histogram and area size information are extracted, and these features of different areas are combined and converted into a fixed-length digital sequence as the character's feature code through specific coding rules. For example, for a singer, his facial feature code can accurately reflect his unique facial features, so that in the subsequent comparison process, it is accurate to identify whether there is any infringement of unauthorized use of his image; object identification feature code extraction: for objects in the video (stage path) The feature code of the object is extracted by first using image segmentation technology to separate the object from the background, then analyzing the shape of the object, extracting the geometric feature parameters such as the circumference, area, longest axis length and shortest axis length of the object, and extracting the color features (main color components) of the object, arranging the geometric and color features in sequence, and converting them into a unique binary code as the feature code of the object. For example, for a stage prop with a unique shape and color, its feature code can accurately identify the object, so as to find out whether there is any infringement in the subsequent comparison, and the feature codes of the person and the object are combined to form a unique feature code of the short video. At the same time, a violation data detection library is constructed, which contains the feature codes of various known copyright infringements and the corresponding violation types and severity weights. For infringements involving unauthorized use of singers' images, the feature codes are recorded in the library, and the corresponding violation type is "character image infringement". The severity weight is set to a higher value to reflect its severity. For infringements involving unauthorized use of a certain brand of musical instruments, their feature codes are also recorded, and the violation type is "object infringement-musical instrument". As the forms of copyright infringement continue to change and new infringement cases emerge, the violation data detection library will be continuously updated and improved to ensure that various new infringements can be identified in a timely manner.

[0050] Next, the originality analysis and homogeneity detection stage (S400) is entered, the key frame data fingerprint is compared with the data in the database, the number and degree of homogeneous works of the short video are calculated, and the originality analysis result is generated. The SSIM similarity calculation method is used as shown in Example 1, and the number of SSIM values ​​that match all key frames in the similar video set is counted as the number of homogeneous works. According to the comparison results, the degree of homogeneity is calculated. The degree of homogeneity is quantified by counting the number of similar key frames and the number of similar video playbacks. The originality score is calculated by the degree of homogeneity. For example, for a short video containing a music work, if it is found that it has a high similarity with other videos and a high degree of homogeneity during the key frame data fingerprint comparison process, this indicates that the short video has a certain degree of homogeneity problem, and further examination is required to determine whether there is any suspicion of infringement.

[0051] Secondly, enter the illegal content identification stage (S500), compare the feature code in the short video with the illegal data detection library, identify illegal content such as copyright infringement, generate illegal content identification results, and use an approximate nearest neighbor search algorithm based on local sensitive hash LSH. As shown in Example 1, the target feature code and the illegal feature code are mapped to multiple hash buckets through LSH functions with multiple different parameters, and similar feature codes are searched in each hash bucket. For the candidate illegal feature code found, its cosine similarity with the target feature code is calculated and the number of illegal feature codes matching the cosine similarity is counted. For example, if unauthorized music works are used in the video, the infringement can be accurately identified by feature code comparison. The illegal content identification result is determined by the illegal content identification score.

[0052] Finally, the comprehensive quality evaluation and result output stage (S600) is entered. The final quality score of the short video is calculated based on the originality analysis results, the illegal content identification results and the basic quality evaluation indicators. According to the quality score, the short video is divided into different quality levels. Short videos with low scores may be judged to have serious copyright issues and legal measures need to be taken, such as sending infringement notices to the short video platform and requesting the removal of the video; short videos with medium scores need further review of their copyright status; short videos with high scores may be relatively compliant in terms of copyright.

[0053] To sum up, this embodiment demonstrates in detail the application process of the short video quality evaluation method based on big data analysis in infringement monitoring of short video copyright protection agencies. Through data collection and preprocessing, key frame extraction and data fingerprint generation, feature code extraction and illegal data detection library construction, originality analysis and homogeneity detection, illegal content identification, comprehensive quality evaluation and result output, etc., effective monitoring and quality evaluation of short video copyright infringement are achieved. This method provides copyright protection agencies with a scientific, accurate and efficient tool, which helps to safeguard the legitimate rights and interests of copyright owners.

[0054] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A short video quality evaluation method based on big data analysis, characterized in that: The specific steps of this method are: S100, collecting short video data, metadata and user behavior data from the short video platform, and standardizing them, unifying the video format and resolution and storing them in a large database, wherein the large database includes historical short video data, key frame data fingerprint information and data in a violation data detection library; S200, extracting the key frame F of the short video using the key frame extraction method k , and generate key frame data fingerprint FP k ; S300, extracting feature codes FC from the person and object identifiers in the short video, and constructing a violation data detection library, wherein the violation data detection library includes the feature code FC of each illegal short video and its corresponding violation type R j and severity weight W j , j = 1, 2, ..., N, N is the total number of violation types and for the feature code FC of all illegal short videos, the total number of corresponding violation feature codes is N total ; S400, comparing the key frame data fingerprint with the big database, calculating the number and degree of homogeneity of short video works, and generating originality analysis results, wherein: In the process of comparing the key frame data fingerprint with the large database, the SSIM similarity calculation method is used to locate other videos similar to the short video to be evaluated; According to the comparison results, count and calculate the number of homogeneous works N of short videos h and the degree of homogeneity, wherein the degree of homogeneity is quantified by counting the number of similar key frames as a proportion index H1 and the number of similar video playbacks as a proportion index H2, and the originality score O is calculated by the degree of homogeneity. And its originality score O is taken as the originality analysis result; S500, comparing the feature code in the short video with the illegal data detection library, identifying and locating the illegal content, and generating illegal content identification results, wherein: In the process of comparing the feature code with the illegal data detection library, an approximate nearest neighbor search algorithm based on local sensitive hashing (LSH) is used to calculate the cosine similarity between the candidate illegal feature code found and the target feature code, and the number of illegal feature codes N that match the cosine similarity is counted. m ; The illegal content identification result is determined by the illegal content identification score V. Where N m is the number of matches of the violation signature code, N total is the total number of illegal signature codes, W j A severity weight for each violation signature; S600, comprehensively analyzing the originality analysis results, the illegal content identification results and the basic quality evaluation index B, calculating the final quality score Q of the short video, classifying the short video into different quality levels according to the quality score, and outputting the evaluation results in a visual manner; The quality levels include high quality, medium quality, and low quality.

2. According to the method for evaluating the quality of short videos based on big data analysis in claim 1, it is characterized in that: The S200 key frame extraction method forms multiple continuous video segments by dividing the short video at fixed time intervals. Starting from the beginning of the video, segments of a certain length are selected in sequence. For each segment, the pixel difference between adjacent frames within it is calculated. where F i Represents the pixel value matrix of the i-th frame, m is the number of frames in each time segment, and the color value difference of each frame at each pixel position is compared with that of the next adjacent frame. hour, is the average pixel difference. At this time, the middle frame of the clip is determined as a key frame F. By traversing the entire short video, k key frames F are extracted in this way. k , the key frame F k Represents the content change point of the short video, quantizes and encodes each key frame, generates a binary data fingerprint with a fixed length of FP, and the data fingerprint FP can uniquely identify the image feature information of the key frame. All key frames are encoded to obtain the key frame data fingerprint FP k .

3. According to the short video quality evaluation method based on big data analysis according to claim 1, it is characterized in that: The S300 extracts the feature code of the person identification by first detecting and locating the person in the video screen, dividing the person's face into multiple regions by analyzing the facial contour and distribution of the facial features of the person, extracting the average color feature of each region, and extracting the area size of each region, combining the color and area features of different regions, and converting them into a fixed-length digital sequence as the person's feature code FC through encoding. rw ; For feature code extraction of object identification, separate the object from the background, analyze the shape of the object, extract the geometric feature parameters of the object's perimeter, area, longest axis length and shortest axis length, and extract the color features of the object. Arrange the geometric and color features in sequence and convert them into binary codes as the feature code FC of the object. wt , the character and object feature codes are combined to form the unique feature code FC of the short video.

4. The short video quality evaluation method based on big data analysis according to claim 1 is characterized in that: The SSIM similarity calculation method of S400 defines the key frame corresponding to the i-th key frame data fingerprint of the short video to be evaluated as F ki , the key frame corresponding to the jth key frame data fingerprint of similar candidate videos in the large database is defined as F tj , divide the key frame into multiple sub-blocks of the same size, and calculate the SSIM value of each sub-block Right now Among them, x and y represent the pixel values ​​of the two sub-blocks respectively, μ x , μ y is the mean of the sub-block, σ x , σ y is the standard deviation of the sub-block, σ xy is the covariance of the sub-block, c1, c2 are constants and c1 = (0.01 × L) 2 、c2=(0.03×L) 2 , L is the dynamic range of pixel values, and the overall SSIM value SSIM of the two key frames is calculated by weighted average ij ,Right now Among them, w s is the weight of the sth sub-block, and sets the similarity threshold to SSIM T , when SSIM ij >SSIM T When the two key frames are similar, the corresponding short videos are included in the similar video set and stored in the big database. The number of SSIM values ​​matching all key frames in the similar video set is counted as the number of homogeneous works N. h .

5. The short video quality evaluation method based on big data analysis according to claim 1 is characterized in that: The number ratio index of similar key frames in S400 Among them, M i represents the number of key frames of the i-th homogeneous work, and M is the total number of key frames of the short video to be evaluated.

6. The short video quality evaluation method based on big data analysis according to claim 1 is characterized in that: The play volume ratio indicator of similar videos in S400 Among them, P sum is the total number of views of similar videos, P avg is the average number of video views in the large database, N t is the total number of short videos in the large database.

7. The short video quality evaluation method based on big data analysis according to claim 1 is characterized in that: The S500 approximate nearest neighbor search algorithm maps the target feature code and the violation feature code to multiple hash buckets through multiple LSH functions with different parameters, searches for similar feature codes in each hash bucket, and calculates the cosine similarity between the candidate violation feature code found and the target feature code: in, is the i-th target feature code in the short video, is the jth feature code in the illegal data detection library. When CosSim ij >CosSim T When CosSim T is the cosine similarity threshold, which determines whether the target signature matches the violation signature.

8. The short video quality evaluation method based on big data analysis according to claim 1 is characterized in that: The basic quality evaluation indicators in S600 include the playback volume P, the number of likes L, the number of comments C, and the completion rate W in the short video data, metadata, and user behavior data, and the corresponding weights are w p 、w l 、w c 、w w , and set the maximum values ​​of playback volume, likes, and comments to P max , L max , C max , then the basic quality evaluation index 9. The short video quality evaluation method based on big data analysis according to claim 1 is characterized in that: The quality score of the S600 short video is Q=B×100×k, where k is a comprehensive adjustment coefficient and The final quality score When Q>Q1, the short video is of high quality, when Q2<Q<Q1, the short video is of medium quality, and when Q<Q2, the short video is of low quality, wherein Q2 and Q1 are short video quality level thresholds.

Images