A live broadcast data analysis and management system

Abstract

The present invention belongs to the field of data analysis management, and discloses a live broadcast data analysis management system, including an acquisition module, a style analysis module and an output module; the acquisition module is used to acquire a set of video images for analysis in a live broadcast video; the style analysis module is used to input a set of video screenshots into a pre-trained style analysis model for analysis, and obtain the live broadcast style corresponding to the live broadcast video; the output module is used to store the live broadcast video number and the live broadcast style corresponding to the live broadcast video. By performing style analysis of the live broadcast video through a trained style analysis model, the consistency of the analysis standard can be maintained, so that the analysis process can be more objective, thereby ensuring the accuracy of the obtained live broadcast style results. In addition, the present invention uses a style analysis model to perform style analysis, and compared with the manual analysis method, because the computer can run for a long time, it can significantly improve the efficiency of the analysis.

Description

Technical Field

[0001] The present invention relates to the field of data analysis management, and in particular to a live broadcast data analysis management system. Background Art

[0002] In order to improve the sales effect of live broadcasts, it is usually necessary to analyze the live broadcast data to find out the factors that have a greater impact on sales during the live broadcast. Live broadcast data includes live broadcast recordings, barrage during the live broadcast, and sales of live products. The live broadcast style will have a greater impact on the effect of the live broadcast. In the prior art, when analyzing live broadcast recordings to determine the live broadcast style, the live broadcast style is usually determined manually. However, the manual determination method will inevitably have subjective opinions during the analysis, and it is difficult to make a completely objective judgment. This also leads to an inaccurate determination of the live broadcast style, which will affect the analysis effect of data analysis based on the live broadcast style. Summary of the invention

[0003] The purpose of the present invention is to disclose a live broadcast data analysis management system to solve the problem of how to improve the accuracy of determining the live broadcast style when analyzing the live broadcast video.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] The present invention provides a live broadcast data analysis management system, including an acquisition module, a style analysis module and an output module;

[0006] The acquisition module is used to acquire a set of video images for analysis in the live video;

[0007] The style analysis module is used to input a collection of video screenshots into a pre-trained style analysis model for analysis to obtain the live broadcast style corresponding to the live broadcast video;

[0008] The output module is used to store the live video number and the live style corresponding to the live video.

[0009] Preferably, the acquisition module includes an extraction unit and a storage unit;

[0010] The extraction unit is used to obtain video images for analysis from the live video according to the set acquisition rules;

[0011] The storage unit is used to save the video image acquired by the extraction unit into a set of video images.

[0012] Preferably, obtaining a video image for analysis from a live video according to a set acquisition rule includes:

[0013] The first step is to split the live video into D video images, where D represents the total number of frames contained in the live video.

[0014] The second step is to number the video images in sequence from the earliest to the latest according to the shooting time of the video images, with the numbering starting from 1 and being a positive integer;

[0015] Step 3: The video images numbered 1 and 1+K are stored in the set FraU as the video images for analysis, where K is a preset positive integer;

[0016] Step 4: Use the video image with the latest acquisition time in FraU as the comparison image;

[0017] Step 5, calculating the number of the next video image for analysis based on the comparison image;

[0018] Step 6: Determine whether the number calculated in step 5 is greater than D. If so, stop the calculation and output the set FraU. If not, proceed to step 7.

[0019] Step 7: Store the video image corresponding to the number calculated in step 5 into the set FraU and proceed to step 4.

[0020] Preferably, calculating the serial number of the next video image for analysis based on the comparison image comprises:

[0021] Use FraCmp to represent the comparison image;

[0022] The video images in FraU are sorted in the order of shooting time from early to late. After sorting, c is used to represent the sequence number of the comparison image;

[0023] Use fra c-1 Indicates the video image with sequence number c-1 in FraU;

[0024] Based on FraCmp and fra c-1 Calculates the number of the next video image to be analyzed.

[0025] Preferably, based on FraCmp and fra c-1 Calculate the number of the next video image to be analyzed, including:

[0026] Calculate FraCmp and fra c-1 The difference coefficient between chgidx:

[0027]

[0028] numdif represents the total number of changed pixels in FraCmp, numfracmp represents the total number of pixels in FraCmp, α represents the quantity weight, and distcmp represents the center and fra of the judgment area of ​​FraCmp. c-1 The distance between the centers of the judgment areas, fralentg represents the number of rows of pixels of FraCmp, and β represents the distance weight;

[0029] Calculate the number of the next video image to be analyzed:

[0030]

[0031] nextnum indicates the number of the next video image to be analyzed, and d indicates the number of FraCmp.

[0032] Preferably, the process of obtaining the total number of changed pixels in FraCmp includes:

[0033] FraCmp and fra c-1 Grayscale is performed to obtain grayscale images G1 and G2;

[0034] The following judgments are made for each pixel in G1:

[0035] Using pix G1 Represents a pixel in G1, and compares the pixel in G2 with pix G1 The pixel points with the same coordinates are represented by pix G2 ;

[0036] Calculating pix G1 and pix G2 The difference in grayscale values ​​between them is graydif:

[0037] graydif=graypix G1 -graypix G2

[0038] graypix G1 and graypix G2 Respectively represent pix G1 and pix G2 Gray value of

[0039] If |graydif|≥graythre, then pix G1 is the changing pixel point, and graythre represents the preset grayscale value.

[0040] Preferably, for the image randimg, randimg∈{FraCmp,fra c-1}, the process of obtaining the judgment area in randimg is as follows:

[0041] Use the preset detection algorithm to detect the pixels in randimg and obtain the pixels that meet the detection requirements;

[0042] Obtain the pixels to be connected from all the pixels that meet the detection requirements;

[0043] The preset connection algorithm is used to process the pixels to be connected to obtain the judgment area.

[0044] Preferably, a preset detection algorithm is used to detect pixels in randimg to obtain pixels that meet the detection requirements, including:

[0045] The following judgments are made for each pixel in randimg:

[0046] For the pixel rdpix in randimg, the Cr component value and Cb component value of rdpix in the YCrCb color space are represented as Cr rdpix and Cb rdpix ;

[0047] The values ​​of the H component, S component, and V component of randpix in the HSV color space are represented as H rdpix , S rdpix and V rdpix ;

[0048] If rdpix meets the following judgment rules, it means that rdpix is ​​a pixel that meets the detection requirements:

[0049]

[0050] crmin and crmax respectively represent the lower limit and upper limit of the preset Cr component, cbmin and cbmax respectively represent the lower limit and upper limit of the preset Cb component, Hmin and Hmax respectively represent the lower limit and upper limit of the preset H component, Sthre represents the preset S component threshold, and Vthre represents the preset V component threshold.

[0051] Preferably, the pixels to be connected are obtained from all the pixels that meet the detection requirements, including:

[0052] For a pixel point cmpix that meets the detection requirements, if all the pixels in the 8-neighborhood of cmpix are pixels that meet the detection requirements, cmpix does not belong to the pixel point to be connected; if the number of pixels in the 8-neighborhood of cmpix that do not meet the detection requirements is greater than or equal to 1, the connection judgment value of cmpix is ​​calculated;

[0053] If the connection judgment value is less than the set connection judgment value threshold, cmpix is ​​a pixel to be connected; if the connection judgment value is greater than or equal to the set connection judgment value threshold, cmpix does not belong to the pixel to be connected.

[0054] Preferably, a preset connection algorithm is used to process the pixels to be connected to obtain a judgment area, including:

[0055] Use a preset connection algorithm to obtain the connection set corresponding to each pixel to be connected;

[0056] The connected domain with the largest area, which is composed of all pixels in the connection set and all pixels that meet the detection requirements, is taken as the middle area;

[0057] Get the minimum bounding rectangle of the middle area, and use the area where the minimum bounding rectangle is located as the judgment area.

[0058] Beneficial effects:

[0059] The present invention performs style analysis of live video through a trained style analysis model, which can maintain the consistency of analysis standards and make the analysis process more objective, thereby ensuring that the accuracy of the obtained live style results is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0061] Figure 1 A schematic diagram of a live broadcast data analysis and management system of the present invention.

[0062] Figure 2 Another schematic diagram of a live broadcast data analysis and management system of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] like Figure 1In one embodiment shown, the present invention provides a live broadcast data analysis management system, including an acquisition module, a style analysis module and an output module;

[0065] The acquisition module is used to acquire a set of video images for analysis in the live video;

[0066] The style analysis module is used to input a collection of video screenshots into a pre-trained style analysis model for analysis to obtain the live broadcast style corresponding to the live broadcast video;

[0067] The output module is used to store the live video number and the live style corresponding to the live video.

[0068] In the above process, the style analysis of the live video is performed by using a trained style analysis model, which can maintain the consistency of the analysis standard, making the analysis process more objective, thereby improving the accuracy of the obtained live style results. In addition, compared with manual analysis, the style analysis is performed by using a style analysis model, which can significantly improve the efficiency of the analysis because the computer can run for a long time.

[0069] Preferably, if Figure 2 As shown, the acquisition module includes an extraction unit and a storage unit;

[0070] The extraction unit is used to obtain video images for analysis from the live video according to the set acquisition rules;

[0071] The storage unit is used to save the video image acquired by the extraction unit into a set of video images.

[0072] In the present invention, instead of directly inputting each video image in the live video as analysis data into the style analysis model for analysis, the extraction is performed through an extraction unit, thereby avoiding the style analysis model from analyzing video images that do not contain useful information, thereby improving the computational efficiency of the style analysis model.

[0073] Preferably, obtaining a video image for analysis from a live video according to a set acquisition rule includes:

[0074] The first step is to split the live video into D video images, where D represents the total number of frames contained in the live video.

[0075] The second step is to number the video images in sequence from the earliest to the latest according to the shooting time of the video images, with the numbering starting from 1 and being a positive integer;

[0076] Step 3: The video images numbered 1 and 1+K are stored in the set FraU as the video images for analysis, where K is a preset positive integer;

[0077] Step 4: Use the video image with the latest acquisition time in FraU as the comparison image;

[0078] Step 5, calculating the number of the next video image for analysis based on the comparison image;

[0079] Step 6: Determine whether the number calculated in step 5 is greater than D. If so, stop the calculation and output the set FraU. If not, proceed to step 7.

[0080] Step 7: Store the video image corresponding to the number calculated in step 5 into the set FraU and proceed to step 4.

[0081] In the present invention, a fixed numbering interval is not used to obtain video images for analysis. This is because in live video, if a fixed numbering interval is used, when the effective content of the picture in the live video (i.e., the foreground part) changes relatively quickly, it is easy to lead to the inability to accurately obtain video images with analytical significance (i.e., very different from the video image in FraU); when the effective content of the picture in the live video changes relatively slowly, it is easy to lead to the acquisition of multiple video images with very similar effective content, resulting in the style analysis module needing to perform too many useless calculations, affecting the efficiency of the output live style results.

[0082] In the present invention, D can be calculated as follows:

[0083] D=framrat×dura

[0084] framrat indicates the frame rate of the live video in fps, and dura indicates the duration of the live video in seconds.

[0085] Preferably, when deframing the live video, software such as Adobe After Effects, VLC MediaPlayer, and FFmpeg can be used.

[0086] Preferably, the preset positive integer is 20.

[0087] The positive integer here can be set according to actual conditions. If the live video does not involve situations where the foreground content changes greatly, such as dancing, the value can be set to be relatively large. Otherwise, the value can be set to be relatively small.

[0088] Preferably, calculating the serial number of the next video image for analysis based on the comparison image comprises:

[0089] Use FraCmp to represent the comparison image;

[0090] The video images in FraU are sorted in the order of shooting time from early to late. After sorting, c is used to represent the sequence number of the comparison image;

[0091] Use fra c-1 Indicates the video image with sequence number c-1 in FraU;

[0092] Based on FraCmp and fra c-1 Calculates the number of the next video image to be analyzed.

[0093] In the present invention, the number of the next video image for analysis is obtained based on the analysis of the two video images with the latest shooting time in FraU. In this way, the numbering interval between adjacent video images for analysis can change with the change of the effective content of the live video screen, which is beneficial to improving the effectiveness of the video images in FraU.

[0094] Preferably, based on FraCmp and fra c-1 Calculate the number of the next video image to be analyzed, including:

[0095] Calculate FraCmp and fra c-1 The difference coefficient between chgidx:

[0096]

[0097] numdif represents the total number of changed pixels in FraCmp, numfracmp represents the total number of pixels in FraCmp, α represents the quantity weight, and distcmp represents the center and fra of the judgment area of ​​FraCmp. c-1 The distance between the centers of the judgment areas, fralentg represents the number of rows of pixels of FraCmp, and β represents the distance weight;

[0098] Calculate the number of the next video image to be analyzed:

[0099]

[0100] nextnum indicates the number of the next video image to be analyzed, and d indicates the number of FraCmp.

[0101] In the present invention, the numbering interval between nextnum and d is related to the discrimination coefficient. The smaller the discrimination coefficient, the larger the numbering interval; the larger the discrimination coefficient, the smaller the numbering interval. Thus, the numbering interval can be adaptively changed with the change of the effective content in the live video, so as to timely obtain video images with analytical significance while avoiding obtaining too many videos with similar screen contents. In addition, when calculating the discrimination coefficient, the present invention does not simply compare the difference of each pixel point, because this comparison method will obtain too large a discrimination coefficient when the background changes, thereby causing the numbering interval of adjacent video images for analysis to be too small, and obtaining too many video images with high effective analysis content similarity. Therefore, when calculating the discrimination coefficient, the present invention also adds the change of the center of the judgment area, so that even if the background changes, and the local content of the foreground does not change significantly, the discrimination coefficient will not be obviously too large, thereby effectively reducing the probability of obtaining too many video images with high content similarity.

[0102] Preferably, the quantity weight is 0.3 and the distance weight is 0.7.

[0103] Preferably, the process of obtaining the total number of changed pixels in FraCmp includes:

[0104] FraCmp and fra c-1 Grayscale is performed to obtain grayscale images G1 and G2;

[0105] The following judgments are made for each pixel in G1:

[0106] Using pix G1 Represents a pixel in G1, and compares the pixel in G2 with pix G1 The pixel points with the same coordinates are represented by pix G2 ;

[0107] Calculating pix G1 and pix G2 The difference in grayscale values ​​between them is graydif:

[0108] graydif=graypix G1 -graypix G2

[0109] graypix G1 and graypix G2 Respectively represent pix G1 and pix G2 Gray value of

[0110] If |graydif|≥graythre, then pix G1is the changing pixel point, and graythre represents the preset grayscale value.

[0111] In the present invention, the changed pixel points are the pixel points whose grayscale values ​​have changed greatly. The more the changed pixel points are, the greater the change in the content of the picture is.

[0112] Preferably, the preset grayscale value is 10.

[0113] Preferably, for the image randimg, randimg∈{FraCmp,fra c-1}, the process of obtaining the judgment area in randimg is as follows:

[0114] Use the preset detection algorithm to detect the pixels in randimg and obtain the pixels that meet the detection requirements;

[0115] Obtain the pixels to be connected from all the pixels that meet the detection requirements;

[0116] The preset connection algorithm is used to process the pixels to be connected to obtain the judgment area.

[0117] During detection, the area composed of pixels that meet the detection requirements is prone to breakage, holes, etc., so the area composed of pixels that meet the detection requirements cannot be directly used as the judgment area. Therefore, after obtaining the pixels that meet the detection requirements, the present invention further processes them through a connection algorithm, which is conducive to obtaining a more accurate judgment area.

[0118] Preferably, a preset detection algorithm is used to detect pixels in randimg to obtain pixels that meet the detection requirements, including:

[0119] The following judgments are made for each pixel in randimg:

[0120] For the pixel rdpix in randimg, the Cr component value and Cb component value of rdpix in the YCrCb color space are represented as Cr rdpix and Cb rdpix ;

[0121] The values ​​of the H component, S component, and V component of randpix in the HSV color space are represented as H rdpix , S rdpix and V rdpix ;

[0122] If rdpix meets the following judgment rules, it means that rdpix is ​​a pixel that meets the detection requirements:

[0123]

[0124] crmin and crmax respectively represent the lower limit and upper limit of the preset Cr component, cbmin and cbmax respectively represent the lower limit and upper limit of the preset Cb component, Hmin and Hmax respectively represent the lower limit and upper limit of the preset H component, Sthre represents the preset S component threshold, and Vthre represents the preset V component threshold.

[0125] In the process of detecting pixels that meet the requirements, the present invention does not only perform the detection in one color space, but also judges whether the detection requirements are met based on the judgment rules of two color spaces at the same time, which can improve the probability of correctly identifying the pixels belonging to the skin area of ​​the person in the live broadcast. Moreover, since the present invention will be further processed by the connection algorithm later, even if a judgment rule stricter than the requirements of one color space is used, it will not affect the integrity of the judgment area finally obtained by the present invention.

[0126] Preferably, the preset lower limit value and upper limit value of the Cr component are 132 and 172 respectively.

[0127] Preferably, the preset lower limit value and upper limit value of the Cb component are 76 and 126 respectively.

[0128] Preferably, the lower limit value and the upper limit value of the H component are 1 and 19, respectively.

[0129] Preferably, the preset S component threshold is 49.

[0130] Preferably, the preset V component threshold is 51.

[0131] Preferably, the pixels to be connected are obtained from all the pixels that meet the detection requirements, including:

[0132] For a pixel point cmpix that meets the detection requirements, if all the pixels in the 8-neighborhood of cmpix are pixels that meet the detection requirements, cmpix does not belong to the pixel point to be connected; if the number of pixels in the 8-neighborhood of cmpix that do not meet the detection requirements is greater than or equal to 1, the connection judgment value of cmpix is ​​calculated;

[0133] If the connection judgment value is less than the set connection judgment value threshold, cmpix is ​​a pixel to be connected; if the connection judgment value is greater than or equal to the set connection judgment value threshold, cmpix does not belong to the pixel to be connected.

[0134] In the present invention, the pixel points that need to be connected are the pixel points at the edge of the area composed of pixel points that meet the detection requirements, so that the present invention does not need to use a connection algorithm to process all pixel points that meet the detection requirements, thereby effectively reducing the amount of calculation while obtaining an accurate judgment area.

[0135] Preferably, the calculation function of the connection judgment value is:

[0136]

[0137] conval cmpix represents the connection judgment value of cmpix, distmin represents the minimum distance between cmpix and the connected domain in the set bkU, bkU is the set of connected domains composed of all pixels that meet the detection requirements except the connected domain where cmpix is ​​located; arebk cmpix Indicates the area of ​​the connected domain where cmpix is ​​located, arebk max represents the maximum area of ​​the connected domain in bkU, μ represents the first weight, Represents the first weight.

[0138] In the present invention, since the area composed of pixels that meet the detection requirements is prone to fractures, holes, etc., the pixels that meet the detection requirements may not all be in the same connected domain, and there is a possibility of being divided into multiple connected domains. Therefore, the present invention can identify the pixels to be connected that have a good repair effect on the entire judgment area after connection by calculating the connection judgment value. The defects existing in the area composed of pixels that meet the detection requirements are effectively repaired, and a more accurate judgment area is obtained. It is conducive to obtaining more accurate numbering.

[0139] The connection judgment value of the present invention is related to the distance between the pixel point and other connected domains and the area of ​​the connected domain. The larger the distance and the smaller the area, the larger the connection judgment value and the smaller the probability of being judged as a pixel point to be connected. It can effectively avoid invalid connection operations on pixels with too small an area and too far away from other connected domains, because such pixels to be connected may belong to pixels in the background, thereby improving the effectiveness of the connection operation.

[0140] Preferably, the first weight is 0.6 and the second weight is 0.4.

[0141] Preferably, the connection judgment value threshold is set to 0.2.

[0142] Preferably, a preset connection algorithm is used to process the pixels to be connected to obtain a judgment area, including:

[0143] Use a preset connection algorithm to obtain the connection set corresponding to each pixel to be connected;

[0144] The connected domain with the largest area, which is composed of all pixels in the connection set and all pixels that meet the detection requirements, is taken as the middle area;

[0145] Get the minimum bounding rectangle of the middle area, and use the area where the minimum bounding rectangle is located as the judgment area.

[0146] The judgment area of ​​the present invention is not the same as the foreground area in the traditional sense, because if only the foreground area is obtained, when there are multiple anchors broadcasting live together, the detection algorithm according to the present invention can still obtain an accurate foreground area, and based on the foreground area acquisition algorithm in the traditional sense, such as a threshold-based segmentation algorithm to acquire the foreground area, multiple foreground areas will be detected, and when the midpoints of the multiple foreground areas are obtained, there is a possibility that they remain unchanged or change very little when the anchor's action or position changes, resulting in an inability to accurately detect changes in the foreground, resulting in inaccurate numbering. Therefore, the present invention obtains the judgment area based on the connection set and all pixel points that meet the detection requirements after obtaining the connection set, which is conducive to accurately detecting changes in the foreground.

[0147] Preferably, a preset connection algorithm is used to obtain a connection set corresponding to each pixel to be connected, including:

[0148] S1, store all pixels that meet the detection requirements into the set accU; store all pixels to be connected into the set wtcU;

[0149] S2, randomly select a pixel from wtcU as the connection pixel;

[0150] S3, storing the connected pixel points into the intermediate set, and calculating the connection judgment value of each pixel point adjacent to the connected pixel point;

[0151] S4, determining whether the minimum value of the connection judgment value obtained in S4 is less than the set connection judgment value threshold, if so, proceeding to S5, if not, proceeding to S6;

[0152] S5, taking the pixel point corresponding to the maximum value of the connection judgment value obtained in S3 as the next connection pixel point, and entering S3;

[0153] S6, taking the pixel points in the intermediate set as the pixel points of a new connected set, resetting the intermediate set to an empty set, and entering S7;

[0154] S7, determine whether wtcU is an empty set, if so, enter S8, if not, enter S2;

[0155] S8, output all obtained connection sets.

[0156] In the above process of the present invention, after each connection set is obtained, the connection pixel points corresponding to the next connection set are obtained from wtcU again, thereby achieving the acquisition of the corresponding connection set for each pixel point to be connected. In addition, the present invention realizes the connection between the connected domain where the pixel point to be connected is located and other connected domains composed of pixel points that meet the detection requirements based on the connection judgment value, realizes the repair of boundary breaks, improves the accuracy of the shape of the middle area, and thus obtains a judgment area with a more accurate position.

[0157] Preferably, the calculation process of the connection judgment value is as follows:

[0158] For pixel s, store all the pixels adjacent to s into the set nerU s ;

[0159] will nerU s Delete the pixels belonging to accU and get the set nerU s,1 ;

[0160] will nerU s,1 Delete the pixels belonging to the existing connection set in the set nerU s,2 ;

[0161] For nerU s,2 The connection judgment value between pixel u, s and u is conpro s,u The calculation formula is:

[0162] conpro s,u =|L s -L u |

[0163] L s and L u Respectively represent the values ​​of s and u on the brightness component in the Lab color space;

[0164] Preferably, the connection judgment value threshold is 5.

[0165] Preferably, the style analysis model is a convolutional neural network model.

[0166] Preferably, the training process of the style analysis model includes:

[0167] 1. Determine the network type of the style analysis model. The style analysis model can be CNN, RNN, GAN, LSTM and other network types;

[0168] 2. Data preparation:

[0169] Acquire images marked with live broadcast styles; live broadcast styles include modern style, ancient style, mature style, young style, etc.;

[0170] Modern style refers to the host wearing non-ancient costumes during live broadcast, ancient costume style refers to the host wearing ancient costumes during live broadcast, mature style means the host is older, for example, over 40 years old, and young style means the host is younger, for example, under 40 years old.

[0171] Performing standardization processing on the acquired image to obtain a standardized image;

[0172] The obtained standardized images are subjected to data enhancement processing. By rotating, flipping, scaling, cropping and other processing on the standardized images, more standardized images are obtained, which increases data diversity and improves the generalization ability of the model.

[0173] 3. Model training:

[0174] Split the standardized images into training and testing sets;

[0175] Use the training set to train the style analysis model to obtain a trained style analysis model;

[0176] Use the test set to test the trained style analysis model and evaluate the difference between the model's output live broadcast style and the actual live broadcast style;

[0177] The weights of the model are updated according to the test results and the loss function, and the weights are optimized using the gradient descent algorithm.

[0178] Preferably, it also includes a comprehensive analysis module;

[0179] The comprehensive analysis module is used to comprehensively analyze the live broadcast process based on live broadcast data such as live broadcast style, barrage during the live broadcast, product sales, etc., to obtain factors that have a significant impact on the live broadcast effect.

[0180] For example, by analyzing the sales of goods under different live broadcast styles of the same anchor, we can determine the live broadcast style that has a significant improvement in the sales effect.

[0181] Then, another influencing factor is analyzed for different anchors with the same live broadcast style to determine whether the other factor has a significant impact.

[0182] By controlling a single variable for analysis, we can identify the factors that have a significant impact on the live broadcast effect. Then, we can improve the live broadcast plan based on these factors to achieve better live broadcast effects.

[0183] It should be noted that the above embodiments illustrate the present invention rather than limit it, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising a number of different elements and by means of a suitably programmed computer. In a unit claim enumerating a number of devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc., does not indicate any order. These words may be interpreted as names.

[0184] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A live broadcast data analysis and management system, characterized in that: Includes acquisition module, style analysis module and output module; The acquisition module is used to acquire a set of video images for analysis in the live video; The style analysis module is used to input a collection of video screenshots into a pre-trained style analysis model for analysis to obtain the live broadcast style corresponding to the live broadcast video; The output module is used to store the live video number and the live style corresponding to the live video; The acquisition module includes an extraction unit and a storage unit; The extraction unit is used to obtain video images for analysis from the live video according to the set acquisition rules; The storage unit is used to save the video image acquired by the extraction unit into a set of video images; According to the set acquisition rules, video images for analysis are acquired from the live video, including: The first step is to split the live video into D video images, where D represents the total number of frames contained in the live video. The second step is to number the video images in sequence from the earliest to the latest according to the shooting time of the video images, with the numbering starting from 1 and being a positive integer; Step 3: The video images numbered 1 and 1+K are stored in the set FraU as the video images for analysis, where K is a preset positive integer; Step 4: Use the video image with the latest acquisition time in FraU as the comparison image; Step 5, calculating the number of the next video image for analysis based on the comparison image; Step 6: Determine whether the number calculated in step 5 is greater than D. If so, stop the calculation and output the set FraU. If not, proceed to step 7. Step 7, store the video image corresponding to the number calculated in step 5 into the set FraU, and go to step 4; Calculate the number of the next video image for analysis based on the comparison image, including: Use FraCmp to represent the comparison image; The video images in FraU are sorted in the order of shooting time from early to late. After sorting, c is used to represent the sequence number of the comparison image; Use fra c-1 Indicates the video image with sequence number c-1 in FraU; Based on FraCmp and fra c-1 Calculate the number of the next video image to be analyzed; Based on FraCmp and fra c-1 Calculate the number of the next video image to be analyzed, including: Calculate FraCmp and fra c-1 The difference coefficient between chgidx: numdif represents the total number of changed pixels in FraCmp, numfracmp represents the total number of pixels in FraCmp, α represents the quantity weight, and distcmp represents the center and fra of the judgment area of ​​FraCmp. c-1 The distance between the centers of the judgment areas, fralentg represents the number of rows of pixels of FraCmp, and β represents the distance weight; Calculate the number of the next video image to be analyzed: nextnum indicates the number of the next video image to be analyzed, and d indicates the number of FraCmp; For the image randimg, randimg∈{FraCmp,fra c-1 }, the process of obtaining the judgment area in randimg is as follows: Use the preset detection algorithm to detect the pixels in randimg and obtain the pixels that meet the detection requirements; Obtain the pixels to be connected from all the pixels that meet the detection requirements; Use the preset connection algorithm to process the pixels to be connected to obtain the judgment area; Use the preset connection algorithm to process the pixels to be connected to obtain the judgment area, including: The connected domain with the largest area, which is composed of all pixels in the connection set and all pixels that meet the detection requirements, is taken as the middle area; Get the minimum bounding rectangle of the middle area, and use the area where the minimum bounding rectangle is located as the judgment area.

2. A live broadcast data analysis and management system according to claim 1, characterized in that: The process of obtaining the total number of changed pixels in FraCmp includes: FraCmp and fra c-1 Grayscale is performed to obtain grayscale images G1 and G2; The following judgments are made for each pixel in G1: Using pix G1 Represents a pixel in G1, and compares the pixel in G2 with pix G1 The pixel points with the same coordinates are represented by pix G2 ; Calculating pix G1 and pix G2 The difference in grayscale values ​​between them is graydif: graydif=graypix G1 -graypix G2 graypix G1 and graypix G2 Respectively represent pix G1 and pix G2 Gray value of If |graydif|≥graythre, then pix G1 is the changing pixel point, and graythre represents the preset grayscale value.

3. A live broadcast data analysis and management system according to claim 1, characterized in that: Use the preset detection algorithm to detect the pixels in randimg and obtain the pixels that meet the detection requirements, including: The following judgments are made for each pixel in randimg: For the pixel point rdpix in randimg, the value of the Cr component and the value of the Cb component of rdpix in the YCrCb color space are represented as Cr rdpix and Cb rdpix ; The values ​​of the H component, S component, and V component of randpix in the HSV color space are represented as H rdpix , S rdpix and V rdpix ; If rdpix meets the following judgment rules, it means that rdpix is ​​a pixel that meets the detection requirements: crmin and crmax respectively represent the lower limit and upper limit of the preset Cr component, cbmin and cbmax respectively represent the lower limit and upper limit of the preset Cb component, Hmin and Hmax respectively represent the lower limit and upper limit of the preset H component, Sthre represents the preset S component threshold, and Vthre represents the preset V component threshold.

4. A live broadcast data analysis and management system according to claim 1, characterized in that: Get the pixels to be connected from all the pixels that meet the detection requirements, including: For a pixel point cmpix that meets the detection requirements, if all of the eight neighboring pixels of cmpix are pixels that meet the detection requirements, cmpix does not belong to the pixel point to be connected; if the number of pixels that do not meet the detection requirements among the eight neighboring pixels of cmpix is ​​greater than or equal to 1, the connection judgment value of cmpix is ​​calculated; If the connection judgment value is less than the set connection judgment value threshold, cmpix is ​​a pixel to be connected; if the connection judgment value is greater than or equal to the set connection judgment value threshold, cmpix does not belong to the pixel to be connected.

Images