Video quality determination model training method, video recommendation method and device

Abstract

The present disclosure relates to the field of computer technology, and specifically to a training method for a video quality determination model, a video recommendation method, a device, an electronic device medium, and a computer program product, wherein the method comprises: obtaining a video sample; when the video sample does not have a quality grade label, generating a corresponding first enhanced sample and a second enhanced sample according to the video sample; obtaining a predicted quality grade corresponding to the first enhanced sample through a neural network model to be trained, and determining a reference quality grade label of the second enhanced sample according to the predicted quality grade corresponding to the first enhanced sample; obtaining a predicted quality grade corresponding to the second enhanced sample through a neural network model to be trained; determining a second loss function; and training the neural network model to be trained according to the second loss function to obtain a video quality determination model. The technical solution of the embodiments of the present disclosure can solve the problem of relying on user behavior when recommending videos in the prior art.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a video quality determination model training method, a video recommendation method, a video quality determination model training device, a video recommendation device, an electronic device, a computer-readable storage medium, and a computer program product. Background Art

[0002] With the rapid development of the Internet, various video platforms have emerged, and video platforms can recommend videos to users. In related technologies, videos can be recommended based on the number of times they are played, or based on the number of likes, or based on the number of favorites. For example, videos with a large number of plays can be recommended.

[0003] However, for some videos that have been uploaded for a short time, their user behaviors (play, like, favorite, etc.) are relatively few, and since the recommendation traffic for newly uploaded videos is limited, it is impossible to determine the quality grade label of the video, which makes it impossible to determine the quality grade of the new video, and it is difficult to recommend these new videos based on their quality grade. Among these videos, there are many videos with good quality. How to recommend newly uploaded videos based on their quality has become an urgent problem to be solved. Summary of the invention

[0004] The present disclosure provides a video quality determination model training, a video quality determination model training device, an electronic device, a computer-readable storage medium, and a computer program product, to at least solve the problem of relying on user behavior when recommending videos in the related art. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a method for training a video quality determination model is provided, comprising obtaining a video sample; when the video sample does not have a quality grade label, generating a corresponding first enhancement sample and a second enhancement sample according to the video sample; wherein the degree of change of the first enhancement sample relative to the video sample is less than the degree of change of the second enhancement sample relative to the video sample; obtaining a predicted quality grade corresponding to the first enhancement sample through a neural network model to be trained, and determining a reference quality grade label of the second enhancement sample according to the predicted quality grade corresponding to the first enhancement sample; obtaining a predicted quality grade corresponding to the second enhancement sample through the neural network model to be trained, and determining a second loss function according to the predicted quality grade corresponding to the second enhancement sample and the reference quality grade label of the second enhancement sample; and training the neural network model to be trained according to the second loss function to obtain a video quality determination model.

[0006] Optionally, the training method also includes: when the video sample has a quality level label, obtaining the predicted quality level corresponding to the video sample through the neural network model to be trained, generating a first loss function of the neural network model to be trained according to the quality level label and the predicted quality level, and training the neural network model to be trained according to the first loss function.

[0007] Optionally, the step of obtaining the predicted quality level corresponding to the video sample through the neural network model to be trained includes: obtaining text features, audio features and image features corresponding to the video sample through the neural network model to be trained; and obtaining the predicted quality level corresponding to the video sample based on the text features, audio features and image features corresponding to the video sample.

[0008] Optionally, the step of generating corresponding first enhanced samples and second enhanced samples according to video samples includes: obtaining image samples corresponding to the video samples; and performing one or more image enhancement processing operations on the image samples to obtain first enhanced samples and second enhanced samples corresponding to the video samples.

[0009] Optionally, the step of obtaining the predicted quality level corresponding to the first enhanced sample through the neural network model to be trained, and determining the reference quality level label of the second enhanced sample according to the predicted quality level includes: obtaining the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level through the neural network model to be trained; when the probability corresponding to the predicted quality level is greater than the confidence probability value, determining the reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample.

[0010] Optionally, the predicted quality level corresponding to the first enhanced sample includes at least two predicted sub-quality levels. When the probability corresponding to the predicted quality level is greater than the confidence probability value, the step of determining the predicted quality level label for the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample includes: obtaining at least two predicted sub-quality levels and the probability corresponding to each predicted sub-quality level, and determining the predicted sub-quality level with the largest probability as the target predicted sub-quality level; when the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, determining the target predicted sub-quality level as the reference quality level label for the second enhanced sample.

[0011] Optionally, the training method also includes: when the probability corresponding to the predicted quality level is less than the confidence probability value, obtaining the predicted quality level corresponding to the second enhanced sample through the neural network model to be trained; and determining the second loss function according to the predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample.

[0012] Optionally, the step of obtaining a predicted quality grade corresponding to the video sample through the neural network model to be trained includes: obtaining a predicted quality grade score corresponding to the video sample through the neural network model to be trained; wherein the predicted quality grade score has a score range.

[0013] Optionally, the video sample has an exposure number label, and the training method also includes: obtaining the predicted exposure number corresponding to the first enhanced sample through the neural network model to be trained, generating a fourth loss function of the neural network model to be trained according to the exposure number label corresponding to the first enhanced sample and the predicted exposure number corresponding to the first enhanced sample, and training the neural network model to be trained according to the fourth loss function and the second loss function.

[0014] Optionally, the video sample has an exposure number label, and the training method also includes: obtaining the predicted exposure number corresponding to the video sample through the neural network model to be trained, generating a third loss function of the neural network model to be trained according to the exposure number label and the predicted exposure number, and training the neural network model to be trained according to the third loss function and the first loss function.

[0015] Optionally, the step of obtaining the predicted exposure times corresponding to the video sample through the neural network model to be trained includes: obtaining the predicted exposure times score corresponding to the video sample through the neural network model to be trained; wherein the predicted exposure times score has a score range.

[0016] According to a second aspect of an embodiment of the present disclosure, a video recommendation method is provided, comprising obtaining a video to be recommended, inputting the video to be recommended into a video quality determination model to obtain a quality level of the video to be recommended; wherein the video quality determination model is obtained by a training method such as the above-mentioned video quality determination model; and recommending the video to be recommended according to the quality level of the video to be recommended.

[0017] Optionally, the step of recommending the video to be recommended according to the quality level of the video to be recommended includes: obtaining multiple videos to be recommended and the quality levels corresponding to the videos to be recommended, sorting the videos to be recommended according to the quality levels corresponding to the multiple videos to be recommended to obtain a first sorting result; and recommending the video to be recommended according to the first sorting result.

[0018] Optionally, the videos to be recommended have quality grade scores, and the step of sorting the videos to be recommended according to the quality grades corresponding to the multiple videos to be recommended to obtain a first sorting result includes: sorting the videos to be recommended according to the quality grade scores corresponding to the multiple videos to be recommended to obtain the first sorting result.

[0019] Optionally, the step of recommending the video to be recommended according to the quality level of the video to be recommended includes: determining the number of exposures of the video to be recommended through a video quality determination model; and recommending the video to be recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended.

[0020] Optionally, the step of recommending the video to be recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended includes: obtaining multiple videos to be recommended and the quality levels and exposure times corresponding to the videos to be recommended, sorting the videos to be recommended according to the quality levels and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result; and recommending the video to be recommended according to the second sorting result.

[0021] Optionally, the videos to be recommended have quality grade scores and exposure times scores, and the step of sorting the videos to be recommended according to the quality grades and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result includes: sorting the videos to be recommended according to the quality grade scores and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result.

[0022] Optionally, the step of sorting the videos to be recommended according to the quality grade scores and exposure scores corresponding to multiple videos to be recommended to obtain a second sorting result includes: obtaining a basic video recommendation score; obtaining a quality grade score threshold and an adjustment weight corresponding to the quality grade range distinguished by the quality grade score threshold, obtaining an exposure score threshold and an adjustment weight corresponding to the quality grade range distinguished by the exposure score threshold; determining a recommendation score for the video to be recommended according to the basic video recommendation score, the adjustment weight corresponding to the quality grade score threshold, and the adjustment weight corresponding to the exposure score threshold; and sorting the multiple videos to be recommended according to the recommendation scores to obtain a second sorting result.

[0023] According to a third aspect of an embodiment of the present disclosure, a training device for a video quality determination model is provided, comprising a sample acquisition unit, configured to execute acquisition of video samples; an enhancement sample generation unit, configured to execute, when the video samples do not have quality grade labels, generation of corresponding first enhancement samples and second enhancement samples according to the video samples; wherein the degree of change of the first enhancement sample relative to the video sample is less than the degree of change of the second enhancement sample relative to the video sample; a reference quality grade label generation unit, configured to execute obtaining a predicted quality grade corresponding to the first enhancement sample through a neural network model to be trained, and determining a reference quality grade label of the second enhancement sample according to the predicted quality grade corresponding to the first enhancement sample; a second loss function determination unit, configured to execute obtaining a predicted quality grade corresponding to the second enhancement sample through the neural network model to be trained, and determining a second loss function according to the predicted quality grade corresponding to the second enhancement sample and the reference quality grade label of the second enhancement sample; and a neural network model training unit, configured to execute training of the neural network model to be trained according to the second loss function to obtain a video quality determination model.

[0024] Optionally, the training device for the video quality determination model also includes: a first loss function determination unit, configured to execute when the video sample has a quality level label, obtain the predicted quality level corresponding to the video sample through the neural network model to be trained, generate a first loss function of the neural network model to be trained according to the quality level label and the predicted quality level, and train the neural network model to be trained according to the first loss function.

[0025] Optionally, the predicted quality level corresponding to the video sample is obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: a feature acquisition unit, configured to execute the acquisition of text features, audio features and image features corresponding to the video sample through the neural network model to be trained; and a quality prediction unit, configured to execute the predicted quality level corresponding to the video sample based on the text features, audio features and image features corresponding to the video sample.

[0026] Optionally, a corresponding first enhanced sample and a second enhanced sample are generated based on the video sample, and the training device of the video quality determination model also includes: an image sample acquisition unit, configured to execute acquisition of image samples corresponding to the video sample; an enhancement sample acquisition unit, configured to execute one or more image enhancement processing operations on the image sample to obtain the first enhanced sample and the second enhanced sample corresponding to the video sample.

[0027] Optionally, the predicted quality level corresponding to the first enhanced sample is obtained through the neural network model to be trained, and the reference quality level label of the second enhanced sample is determined according to the predicted quality level. The training device of the video quality determination model also includes: a predicted quality probability acquisition unit, configured to execute the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level through the neural network model to be trained; and a first reference label determination unit, configured to determine the reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample when the probability corresponding to the predicted quality level is greater than the confidence probability value.

[0028] Optionally, the predicted quality level corresponding to the first enhanced sample includes at least two predicted sub-quality levels. When the probability corresponding to the predicted quality level is greater than the confidence probability value, the predicted quality level corresponding to the first enhanced sample is determined as the reference quality level label of the second enhanced sample. The training device of the video quality determination model also includes: a target sub-quality level determination unit, configured to execute acquisition of at least two predicted sub-quality levels and the probability corresponding to each predicted sub-quality level, and determine the predicted sub-quality level with the largest probability as the target predicted sub-quality level; and a second reference label determination unit, configured to execute when the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, determine the target predicted sub-quality level as the reference quality level label of the second enhanced sample.

[0029] Optionally, the training device of the video quality determination model also includes: an enhanced sample prediction quality determination unit, configured to execute, when the probability corresponding to the prediction quality level is less than the confidence probability value, obtaining the prediction quality level corresponding to the second enhanced sample through the neural network model to be trained; and a second loss function determination subunit, configured to execute a second loss function based on the prediction quality level corresponding to the first enhanced sample and the prediction quality level corresponding to the second enhanced sample.

[0030] Optionally, the predicted quality corresponding to the video sample is obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: a predicted quality score acquisition unit, configured to execute the predicted quality grade score corresponding to the video sample obtained by the neural network model to be trained; wherein the predicted quality grade score has a score range.

[0031] Optionally, the video sample has an exposure number label, and the training device of the video quality determination model also includes: a quality level label training unit, configured to execute the predicted exposure number corresponding to the first enhanced sample obtained by the neural network model to be trained, generate a fourth loss function of the neural network model to be trained according to the exposure number label corresponding to the first enhanced sample and the predicted exposure number corresponding to the first enhanced sample, and train the neural network model to be trained according to the fourth loss function and the second loss function.

[0032] Optionally, the video sample has an exposure number label, and the training device of the video quality determination model also includes: a training unit without a quality grade label, configured to execute the predicted exposure number corresponding to the video sample through the neural network model to be trained, generate a third loss function of the neural network model to be trained according to the exposure number label and the predicted exposure number, and train the neural network model to be trained according to the third loss function and the first loss function.

[0033] Optionally, the predicted exposure times corresponding to the video samples are obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: a predicted exposure times score acquisition unit, which is configured to execute the predicted exposure times score corresponding to the video samples obtained by the neural network model to be trained; wherein the predicted exposure times score has a score range.

[0034] According to a fourth aspect of an embodiment of the present disclosure, a video recommendation device is provided, comprising a video acquisition unit for recommendation, configured to acquire the video to be recommended, input the video to be recommended into a video quality determination model to obtain a quality level of the video to be recommended; wherein the video quality determination model is obtained by a training method such as the above-mentioned video quality determination model; and the video recommendation unit is configured to recommend the video to be recommended according to the quality level of the video to be recommended.

[0035] Optionally, the video to be recommended is recommended according to its quality level, and the video recommendation device further includes: a quality level sorting unit, configured to execute acquisition of multiple videos to be recommended and quality levels corresponding to the videos to be recommended, and to sort the videos to be recommended according to the quality levels corresponding to the multiple videos to be recommended to obtain a first sorting result; and a first sorting unit, configured to execute recommendation of the video to be recommended according to the first sorting result.

[0036] Optionally, the video to be recommended has a quality grade score, and the videos to be recommended are sorted according to the quality grades corresponding to the multiple videos to be recommended to obtain a first sorting result. The video recommendation device also includes: a quality grade score sorting unit, which is configured to execute sorting of the videos to be recommended according to the quality grade scores corresponding to the multiple videos to be recommended to obtain a first sorting result.

[0037] Optionally, the training device of the video quality determination model also includes: recommending the video to be recommended according to the quality level of the video to be recommended, and the video recommendation device also includes: an exposure times acquisition unit, configured to execute the exposure times of the video to be recommended determined by the video quality level determination model; a comprehensive recommendation unit, configured to execute the recommendation of the video to be recommended according to the quality level of the video to be recommended and the exposure times of the video to be recommended.

[0038] Optionally, the video to be recommended is recommended according to its quality level and exposure times, and the video recommendation device further includes: a second sorting result acquisition unit, configured to execute acquisition of multiple videos to be recommended and quality levels and exposure times corresponding to the videos to be recommended, and to sort the videos to be recommended according to the quality levels and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result; and a second sorting recommendation unit, configured to execute recommendation of the video to be recommended according to the second sorting result.

[0039] Optionally, the videos to be recommended have quality grade scores and exposure times scores, and the videos to be recommended are sorted according to the quality grades and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result. The video recommendation device also includes: a second sorting result acquisition subunit, which is configured to execute sorting of the videos to be recommended according to the quality grade scores and exposure times scores corresponding to the multiple videos to be recommended to obtain a second sorting result.

[0040] Optionally, the videos to be recommended are sorted according to the quality grade scores and exposure number scores corresponding to multiple videos to be recommended to obtain a second sorting result, and the video recommendation device also includes: a video recommendation basic score acquisition unit, configured to execute acquisition of the video recommendation basic score; an adjustment weight acquisition unit, configured to execute acquisition of the quality grade score threshold and the adjustment weight corresponding to the quality grade range distinguished by the quality grade score threshold, and acquisition of the exposure number score threshold and the adjustment weight corresponding to the quality grade range distinguished by the exposure number score threshold; a recommendation score acquisition unit, configured to execute determination of the recommendation score of the video to be recommended according to the video recommendation basic score, the adjustment weight corresponding to the quality grade score threshold, and the adjustment weight corresponding to the exposure number score threshold; and a recommendation score sorting unit, configured to execute sorting of the multiple videos to be recommended according to the recommendation score to obtain the second sorting result.

[0041] According to a fifth aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0042] processor;

[0043] a memory for storing processor-executable instructions;

[0044] The processor is configured to execute executable instructions to implement the training of a video quality determination model or a video recommendation method as described above.

[0045] According to a sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform training of a video quality determination model or a video recommendation method as described above.

[0046] According to a seventh aspect of an embodiment of the present disclosure, a computer program product, a computer program / instruction, is provided, characterized in that when the computer program / instruction is executed by a processor, any of the above-mentioned video quality determination model training or video recommendation methods is implemented.

[0047] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0048] In a video quality determination model training method and a video recommendation method provided by an embodiment of the present disclosure, the video quality determination model can be trained when the video does not have a quality grade label, and the video can be recommended based on the trained video quality model. On the one hand, the video quality model can be trained without relying on the quality grade label of the video, avoiding the problem that the video quality determination model cannot be trained due to the lack of quality grade labels, thereby improving the accuracy of the trained video quality determination model in determining the quality grade of new videos; on the other hand, recommending new videos based on the trained video quality determination model can highlight new videos with better quality among multiple newly uploaded videos, thereby more reasonably allocating recommended traffic, and thus enabling users to see new videos with better quality; on the other hand, recommending new videos based on the trained video quality determination model can greatly enhance the creative enthusiasm of video creators, thereby improving the quality and quantity of creation, and thus enhancing the influence of the video platform.

[0049] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0051] Figure 1 is a flowchart of a method for training a video quality determination model according to an exemplary embodiment;

[0052] Figure 2 is a flowchart showing a method of performing one or more image enhancement processing operations on an image sample to obtain a first enhanced sample and a second enhanced sample corresponding to a video sample according to an exemplary embodiment;

[0053] Figure 3 is a flowchart showing a method of determining a reference quality level label of a second enhanced sample according to a predicted quality level corresponding to a first enhanced sample according to an exemplary embodiment;

[0054] Figure 4is a flowchart showing a method of determining a target predicted sub-quality level as a reference quality level label of a second enhanced sample when a probability corresponding to the target predicted sub-quality level is greater than a confidence probability value according to an exemplary embodiment;

[0055] Figure 5 is a flowchart showing a method of determining a second loss function according to a prediction quality level corresponding to a first enhanced sample and a prediction quality level corresponding to a second enhanced sample according to an exemplary embodiment;

[0056] Figure 6 is a flow chart showing a method of obtaining a predicted quality level corresponding to a video sample according to text features, audio features and image features corresponding to the video sample according to an exemplary embodiment;

[0057] Figure 7 is a flow chart showing a method of recommending a video to be recommended according to the quality of the video to be recommended according to an exemplary embodiment;

[0058] Figure 8 is a flowchart showing a method of sorting recommended videos according to quality levels corresponding to a plurality of recommended videos to obtain a first sorting result, and recommending the recommended videos according to the first sorting result according to an exemplary embodiment;

[0059] Fig. 9 is a schematic diagram showing a method of allocating startup traffic according to a recommendation score according to an exemplary embodiment;

[0060] Fig.10 is a schematic diagram of a method for training a video quality determination model according to an exemplary embodiment;

[0061] Fig.11 is a flowchart showing a method of recommending a video to be recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended according to an exemplary embodiment;

[0062] Fig.12 is a flowchart showing a method of recommending a video to be recommended according to a second ranking result according to an exemplary embodiment;

[0063] Fig.13 is a flow chart showing a method of sorting recommended videos according to recommendation scores to obtain a second sorting result according to an exemplary embodiment;

[0064] Fig.14 is a schematic diagram showing the composition of a training device for a video quality determination model according to an exemplary embodiment;

[0065] Fig.15 is a schematic diagram showing the composition of a video recommendation device according to an exemplary embodiment;

[0066] Fig.16 It is a schematic diagram showing the structure of an electronic device suitable for implementing the exemplary embodiments of the present disclosure according to an exemplary embodiment. DETAILED DESCRIPTION

[0067] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0068] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0069] With the rapid development of the Internet, various video platforms have emerged, and video platforms can recommend videos to users. In related technologies, videos can be recommended based on the number of times they are played, or based on the number of likes, or based on the number of favorites. For example, videos with a large number of plays can be recommended.

[0070] However, for some videos that have been uploaded for a short time, their user behaviors (play, like, favorite, etc.) are relatively few, and since the recommendation traffic for newly uploaded videos is limited, it is impossible to determine the quality grade label of the video, which makes it impossible to determine the quality grade of the new video, and it is difficult to recommend these new videos based on their quality grade. Among these videos, there are many videos with good quality. How to recommend newly uploaded videos based on their quality has become an urgent problem to be solved.

[0071] Figure 1 is a flowchart of training a video quality determination model according to an exemplary embodiment. Figure 1 As shown, the training of the video quality determination model includes the following steps.

[0072] In step S110, a video sample is obtained;

[0073] In step S120, when the video sample does not have a quality level label, a corresponding first enhanced sample and a second enhanced sample are generated according to the video sample; wherein the degree of change of the first enhanced sample relative to the video sample is less than the degree of change of the second enhanced sample relative to the video sample;

[0074] In step S130, the predicted quality level corresponding to the first enhanced sample is obtained by the neural network model to be trained, and the reference quality level label of the second enhanced sample is determined according to the predicted quality level corresponding to the first enhanced sample;

[0075] In step S140, the predicted quality level corresponding to the second enhanced sample is obtained by the neural network model to be trained, and the second loss function is determined according to the predicted quality level corresponding to the second enhanced sample and the reference quality level label of the second enhanced sample;

[0076] In step S150, the neural network model to be trained is trained according to the second loss function to obtain a video quality determination model.

[0077] In a video quality determination model training method and a video recommendation method provided by an embodiment of the present disclosure, the video quality determination model can be trained when the video does not have a quality grade label, and the video can be recommended based on the trained video quality model. On the one hand, the video quality model can be trained without relying on the quality grade label of the video, avoiding the problem that the video quality determination model cannot be trained due to the lack of quality grade labels, thereby improving the accuracy of the trained video quality determination model in determining the quality grade of new videos; on the other hand, recommending new videos based on the trained video quality determination model can highlight new videos with better quality among multiple newly uploaded videos, thereby more reasonably allocating recommended traffic, and thus enabling users to see new videos with better quality; on the other hand, recommending new videos based on the trained video quality determination model can greatly enhance the creative enthusiasm of video creators, thereby improving the quality and quantity of creation, and thus enhancing the influence of the video platform.

[0078] Next, we will combine Figure 1 The embodiments describe in more detail steps S110 to S150 of training the video quality determination model in the present exemplary embodiment.

[0079] Step S110, obtaining a video sample, and determining whether the video sample has a quality level label;

[0080] In an example embodiment of the present disclosure, video samples may be obtained. Specifically, video samples may include videos uploaded by users on a video platform (such as Kuaishou video platform), and these video samples may include videos with user behaviors (such as likes, favorites, plays, etc.). After the user uploads the video, the video may be stored in a server or storage, and when the video needs to be used, the video may be called in the server or storage. It should be noted that the present disclosure does not specifically limit the method for obtaining video samples.

[0081] In an exemplary embodiment of the present disclosure, after obtaining the video sample through the above steps, it can be determined whether the video sample has a quality grade label. Specifically, the quality grade label can be used to evaluate the quality grade of the video sample. The quality grade of the video sample can be divided into multiple levels, and each level represents a different quality grade. For example, the quality grade can be divided into excellent, ordinary, grayscale, poor quality, etc., among which the video sample with excellent quality grade has the highest quality grade, and the video sample with poor quality grade has the lowest quality grade. It should be noted that the present disclosure does not specifically limit the classification method of quality grades.

[0082] Specifically, the quality grade labels can be manually labeled by the staff of the video platform. For example, the staff of the video platform can review the video content and label the video according to their own perception; they can also be automatically generated when the video sample meets certain conditions. For example, when the number of collections of the video is greater than 5,000, the video can be labeled as ordinary, and when the number of collections of the video is greater than 20,000, the video can be labeled as excellent. It should be noted that the present disclosure does not specifically limit the method for generating quality grade labels.

[0083] In an example embodiment of the present disclosure, after obtaining the above-mentioned video sample, it can be determined whether the video sample has a quality grade label. Specifically, due to missing data or failure to meet the quality grade label generation conditions, the video sample does not have a quality grade label. For example, when manually annotating videos, only videos with a playback volume greater than 10,000 are manually annotated, and videos with a playback volume less than 10,000 do not have corresponding quality grade labels; or, when automatically generating annotations for videos, only videos with more than 3,000 likes are automatically annotated, and videos with less than 3,000 likes do not have corresponding quality grade labels. It should be noted that the present disclosure does not specifically limit the specific types of video samples that do not have quality grade labels.

[0084] In an exemplary embodiment of the present disclosure, the quality level label corresponding to the video sample can be stored in an associated form. After obtaining the video sample, it can be determined whether the video sample has the quality level label according to the association relationship between the video sample and the quality level label corresponding to the video sample. It should be noted that the present disclosure does not specifically limit the specific method for determining whether a video sample has a quality level label.

[0085] Step S120, when the video sample does not have a quality level label, generating a corresponding first enhanced sample and a second enhanced sample according to the video sample;

[0086] In an exemplary embodiment of the present disclosure, when it is determined through the above steps that the video sample does not have a quality level label, a corresponding first enhanced sample and a second enhanced sample can be generated according to the video sample. Specifically, the first enhanced sample and the second enhanced sample refer to video samples obtained after processing the video sample. Among them, the degree of change of the first enhanced sample relative to the video sample is less than the degree of change of the second enhanced sample relative to the video sample. For example, it can include random image cropping processing, rotation processing, normalization processing, contrast processing, brightness processing, color processing, translation processing, etc. It should be noted that the present disclosure does not specifically limit the specific method of video processing.

[0087] Specifically, the change degree of the first enhanced sample relative to the video sample is smaller than the change degree of the second enhanced sample relative to the video sample, that is, the change of the first enhanced sample relative to the video sample is smaller, and the change of the second enhanced sample relative to the video sample is larger.

[0088] In an exemplary embodiment of the present disclosure, after obtaining the first enhanced sample corresponding to the video sample and the second enhanced sample corresponding to the video sample through the above steps, the predicted quality level corresponding to the first enhanced sample can be obtained through the neural network model to be trained, and the reference quality level label of the second enhanced sample can be determined according to the predicted quality level corresponding to the first enhanced sample. Specifically, the first enhanced sample can be input into the neural network model to be trained to obtain the predicted quality level corresponding to the first enhanced sample, and then the reference quality level label of the second enhanced sample can be determined according to the predicted quality level corresponding to the first enhanced sample.

[0089] In an exemplary embodiment of the present disclosure, an image sample corresponding to a video sample may be obtained, and one or more image enhancement processing operations may be performed on the image sample to obtain a first enhanced sample and a second enhanced sample corresponding to the video sample. Figure 2 As shown, performing one or more image enhancement processing operations on the image sample to obtain the first enhanced sample and the second enhanced sample corresponding to the video sample may include the following steps S210 to S220:

[0090] Step S210, obtaining an image sample corresponding to the video sample;

[0091] Step S220: performing one or more image enhancement processing operations on the image sample to obtain a first enhanced sample and a second enhanced sample corresponding to the video sample.

[0092] In an example embodiment of the present disclosure, an image sample corresponding to a video sample may be obtained. Specifically, the image sample corresponding to the video sample may include an image of a certain frame in the video sample. For example, a certain frame may be randomly selected from the video sample as the image sample corresponding to the video sample; or, the image sample corresponding to the video sample may include an image that can summarize the video sample. For example, multiple frames of images of the video sample may be obtained, and then the multiple frames of images may be averaged to obtain the image sample corresponding to the video sample. It should be noted that the present disclosure does not specifically limit the specific method for obtaining the image sample corresponding to the video sample.

[0093] In an exemplary embodiment of the present disclosure, after obtaining the image sample corresponding to the video sample through the above steps, one or more image enhancement processing operations may be performed on the image sample to obtain the first enhanced sample and the second enhanced sample corresponding to the video sample. Specifically, when creating the first enhanced sample and the second enhanced sample corresponding to the video sample for the video sample, image processing may be performed on the image sample corresponding to the video sample to obtain the first enhanced image sample and the second enhanced image sample, wherein the degree of change of the first enhanced image sample relative to the image information corresponding to the video sample is less than the degree of change of the second enhanced sample relative to the image information corresponding to the video sample. Specifically, the image processing may include random cropping, rotation, normalization, contrast, brightness, color, translation, etc. of the image. It should be noted that the present disclosure does not specifically limit the specific type of image enhancement processing operation.

[0094] Through the above steps S210 to S220, image samples corresponding to the video samples can be obtained, and one or more image enhancement processing operations are performed on the image samples to obtain first enhanced samples and second enhanced samples corresponding to the video samples.

[0095] Step S130, obtaining a predicted quality level corresponding to the first enhanced sample through the neural network model to be trained, and determining a reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample;

[0096] Step S140, obtaining a predicted quality level corresponding to the second enhanced sample through the neural network model to be trained, and determining a second loss function according to the predicted quality level corresponding to the second enhanced sample and a reference quality level label of the second enhanced sample;

[0097] In an exemplary embodiment of the present disclosure, after obtaining the predicted quality level corresponding to the first enhanced sample and the reference quality level label of the second enhanced sample through the above steps, the predicted quality level corresponding to the second enhanced sample can be obtained through the neural network model to be trained, and the second loss function is determined according to the predicted quality level corresponding to the second enhanced sample and the predicted quality level corresponding to the first enhanced sample. It should be noted that the present disclosure does not specifically limit the specific method of determining the second loss function according to the predicted quality level corresponding to the second enhanced sample and the reference quality level label of the second enhanced sample.

[0098] Specifically, the second loss function includes the difference between the predicted quality level of the first enhanced sample and the predicted quality level of the second enhanced sample. The second loss function can represent the difference between the result of the neural network model to be trained in predicting the quality level of the first enhanced sample and the predicted quality level of the second enhanced sample predicted by the neural network model to be trained. Furthermore, the second loss function may include an absolute value loss function, a logarithmic loss function, a mean square error loss function, an exponential loss function, a Hinge loss function, a perceptual loss function, a cross entropy loss function, etc. It should be noted that the present disclosure does not specifically limit the specific form of the second loss function.

[0099] Step S150: training the neural network model to be trained according to the second loss function to obtain a video quality determination model.

[0100] In an exemplary embodiment of the present disclosure, after obtaining the second loss function through the above steps, the neural network parameters of the neural network model to be trained can be iteratively updated according to the second loss function to train the neural network model to be trained. Specifically, the neural network parameters of the convolutional neural network of the neural network model to be trained can be initialized first, and the output value can be obtained through multiple layers in the convolutional neural network (such as convolutional layer, deconvolution layer, fully connected layer, etc.), and the neural network parameters are adjusted according to the overall loss function, and the neural network parameters are iteratively updated. When the overall loss function converges, the training ends. Specifically, the entire model can be back-propagated according to the second loss function to obtain the gradient of each neural network parameter in the neural network model to be trained, and the neural network model to be trained is optimized during the iterative update process to make the value of the second loss function smaller and smaller.

[0101] It should be noted that the present disclosure does not specifically limit the structure of the neural network model to be trained and the method of training the neural network model to be trained according to the second loss function.

[0102] In an exemplary embodiment of the present disclosure, the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level can be obtained through the neural network model to be trained. When the probability corresponding to the predicted quality level is greater than the confidence probability value, the reference quality level label of the second enhanced sample is determined according to the predicted quality level corresponding to the first enhanced sample. Figure 3 As shown, determining the reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample may include the following steps S310-S320:

[0103] Step S310, obtaining a predicted quality level corresponding to the first enhanced sample and a probability corresponding to the predicted quality level through the neural network model to be trained;

[0104] In an example embodiment of the present disclosure, the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level can be obtained by the neural network model to be trained. Specifically, when the neural network model to be trained predicts the first enhanced sample, it can output the predicted quality level of the first enhanced sample and the probability corresponding to the predicted quality level. Among them, the probability corresponding to the predicted quality level can be used to indicate the confidence level of the model in the predicted quality level obtained by the prediction, that is, the greater the probability corresponding to the predicted quality level, the greater the possibility that the first enhanced sample is of the predicted quality level. For example, the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level obtained by the neural network model to be trained are normal-80%. It should be noted that the present disclosure does not specifically limit the specific method of obtaining the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level by the neural network model to be trained.

[0105] Step S320, when the probability corresponding to the predicted quality level is greater than the confidence probability value, determining a reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample;

[0106] In an exemplary embodiment of the present disclosure, after obtaining the predicted quality level corresponding to the first enhanced sample through the above steps, the relationship between the predicted quality level corresponding to the first enhanced sample and the confidence probability value can be determined, and when the probability corresponding to the predicted quality level is greater than the confidence probability value, the predicted quality level corresponding to the first enhanced sample is determined as the reference quality level label of the second enhanced sample. Specifically, the first enhanced sample is created based on the video sample, and the first enhanced sample refers to an enhanced sample whose degree of change relative to the video sample is less than the degree of change of the second enhanced sample relative to the video sample. When the probability corresponding to the predicted quality level of the first enhanced sample is greater than the confidence probability value, it means that the predicted quality level can be used as the reference quality level label of the first enhanced sample, and the second enhanced sample is also created based on the video sample. Therefore, the predicted quality level corresponding to the first enhanced sample can be determined as the reference quality level label of the second enhanced sample.

[0107] For example, the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level are normal-80%, and the confidence probability value is 70%, that is, the predicted quality level corresponding to the first enhanced sample can be determined as the reference quality level label of the second enhanced sample.

[0108] It should be noted that the present disclosure does not impose any special limitation on the specific numerical value of the confidence probability value.

[0109] In an exemplary embodiment of the present disclosure, the second enhanced sample can be input into the neural network model to be trained, the predicted quality level corresponding to the second enhanced sample is obtained through the neural network model to be trained, and the second loss function is obtained according to the predicted quality level corresponding to the second enhanced sample and the reference quality level label of the second enhanced sample. It should be noted that the present disclosure does not specifically limit the specific method of determining the second loss function according to the predicted quality level corresponding to the second enhanced sample and the reference quality level label of the second enhanced sample.

[0110] Specifically, the second loss function includes the difference between the predicted quality level of the second enhanced sample and the reference quality level label of the second enhanced sample. The second loss function can represent the predicted difference between the result of the neural network model to be trained in predicting the quality level of the second enhanced sample and the reference quality level label of the second enhanced sample.

[0111] Through the above steps S310 to S320, the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level can be obtained through the neural network model to be trained. When the probability corresponding to the predicted quality level is greater than the confidence probability value, the reference quality level label of the second enhanced sample is determined according to the predicted quality level corresponding to the first enhanced sample.

[0112] In an exemplary embodiment of the present disclosure, at least two predicted sub-quality levels and the probabilities corresponding to each predicted sub-quality level may be obtained, and the predicted sub-quality level with the largest probability may be determined as the target predicted sub-quality level. When the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, the target predicted sub-quality level may be determined as the reference quality level label of the second enhanced sample. Figure 4 As shown, when the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, determining the target predicted sub-quality level as the reference quality level label of the second enhanced sample may include the following steps S410-S420:

[0113] Step S410, obtaining at least two predicted sub-quality levels and the probability corresponding to each predicted sub-quality level, and determining the predicted sub-quality level with the greatest probability as the target predicted sub-quality level;

[0114] In an exemplary embodiment of the present disclosure, at least two predicted sub-quality levels and the probabilities corresponding to each predicted sub-quality level can be obtained. Specifically, after the first enhanced sample is input into the neural network model to be trained, at least two predicted sub-quality levels corresponding to the first enhanced sample and the probabilities corresponding to the predicted sub-quality levels can be output, and the predicted sub-quality level with the largest probability corresponding to the predicted sub-quality level is determined as the target predicted sub-quality level. It should be noted that the present disclosure does not specifically limit the specific method for obtaining the predicted sub-quality level and the probabilities corresponding to the predicted sub-quality level.

[0115] Step S420: When the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, the target predicted sub-quality level is determined as the reference quality level label of the second enhanced sample.

[0116] In an exemplary embodiment of the present disclosure, when the target predicted sub-quality level corresponding to the first enhanced sample and the probability corresponding to the target predicted sub-quality level are obtained through the above steps, the relationship between the probability corresponding to the target predicted sub-quality level and the confidence probability value can be determined. When the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, it indicates that the target predicted sub-quality level is most likely the predicted quality level corresponding to the first enhanced sample, that is, at this time, the target predicted sub-quality level can be determined as the reference quality level label corresponding to the first enhanced sample, and since the second enhanced sample is also obtained by processing the video sample, the target predicted sub-quality level can be determined as the reference quality level label of the second enhanced sample.

[0117] For example, the predicted sub-quality level corresponding to the first enhanced sample may include excellent, ordinary, grayscale, and poor quality. At this time, the first enhanced sample is input into the neural network model to be trained, and the output of the predicted sub-quality level corresponding to the first enhanced sample and the predicted sub-quality level corresponding to the probability of excellent 5%, ordinary 80%, grayscale 10%, and poor quality 5% are output. At this time, ordinary can be used as the reference quality level label of the first enhanced sample. Since the second enhanced sample is also obtained by processing the video sample, ordinary can be determined as the reference quality level label of the second enhanced sample.

[0118] In an exemplary embodiment of the present disclosure, Y_s_c_pred is used as the predicted quality level corresponding to the second enhanced sample, and Y_w_c_pred is used as the reference quality level label of the second enhanced sample. Then, the expression of the second loss function loss2 is as follows:

[0119] loss2=log(Y_s_c_pred)*argmax(Y_w_c_p red)

[0120] Through the above steps S410 to S420, the predicted sub-quality level and the probability corresponding to the predicted sub-quality level can be obtained, and the predicted sub-quality level corresponding to the predicted sub-quality level with the largest probability is determined as the target predicted sub-quality level. When the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, the target predicted sub-quality level is determined as the reference quality level label of the second enhanced sample, which greatly improves the generalization of the neural network model to be trained.

[0121] In an exemplary embodiment of the present disclosure, when the probability corresponding to the predicted quality level is less than the confidence probability value, the predicted quality level corresponding to the second enhanced sample can be obtained by the neural network model to be trained, and the second loss function can be determined according to the predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample. Figure 5 As shown, determining the second loss function according to the prediction quality level corresponding to the first enhanced sample and the prediction quality level corresponding to the second enhanced sample may include the following steps S510-S520:

[0122] Step S510, when the probability corresponding to the predicted quality level is less than the confidence probability value, obtaining the predicted quality level corresponding to the second enhanced sample by using the neural network model to be trained;

[0123] In an example embodiment of the present disclosure, when the probability corresponding to the predicted quality level is less than the confidence probability value, the predicted quality level corresponding to the first enhanced sample can be obtained through the neural network model to be trained. Specifically, the probability corresponding to the predicted quality level can be used to indicate the model's confidence level in the predicted quality level, that is, the greater the probability corresponding to the predicted quality level, the greater the possibility that the first enhanced sample is of the predicted quality level. When the probability corresponding to the predicted quality level is less than the confidence probability value, it means that the possibility that the quality level label of the first enhanced sample is the above-mentioned predicted quality level is small, and the predicted quality level cannot be used as the quality level label of the first enhanced sample. At this time, the second enhanced sample needs to be input into the neural network model to be trained, and the predicted quality level corresponding to the second enhanced sample is obtained through the neural network model to be trained.

[0124] Step S520: determining a second loss function according to the prediction quality level corresponding to the first enhanced sample and the prediction quality level corresponding to the second enhanced sample.

[0125] In an exemplary embodiment of the present disclosure, after obtaining the predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample through the above steps, the second loss function can be determined according to the predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample. Specifically, the second loss function includes the difference between the predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample. The second loss function can represent the predicted difference between the quality level of the first enhanced sample and the quality level of the second enhanced sample predicted by the neural network model to be trained. The predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample can be brought closer by training the neural network model to be trained according to the second loss function.

[0126] In an exemplary embodiment of the present disclosure, Y_w_c_p red is used as the predicted quality level corresponding to the first enhanced sample, and Y_s_c_pred is used as the predicted quality level corresponding to the second enhanced sample. The expression of the second loss function loss3 is as follows:

[0127] loss3=Y_w_c_p red*(log(Y_w_c_pred)-log(Y_s_c_pred))

[0128] Through the above steps S510 to S520, when the probability corresponding to the predicted quality level is less than the confidence probability value, the predicted quality level corresponding to the second enhanced sample can be obtained through the neural network model to be trained, and the second loss function can be determined according to the predicted quality level corresponding to the first enhanced sample and the predicted quality level corresponding to the second enhanced sample.

[0129] In an exemplary embodiment of the present disclosure, when it is determined through the above steps that a video sample has a quality level label, the video sample can be input into the neural network model to be trained, and the neural network model to be trained can output the predicted quality level of the video sample. The predicted quality level of the video sample is then compared with the quality level label corresponding to the video sample to obtain a first loss function of the neural network model to be trained, and the neural network model to be trained is trained according to the first loss function.

[0130] In an exemplary embodiment of the present disclosure, the predicted quality level corresponding to the video sample can be obtained by the neural network model to be trained. Specifically, the neural network model to be trained refers to the video quality determination model to be trained in the present disclosure, through which the quality level of the video sample can be predicted to obtain the predicted quality level. Furthermore, the video quality determination model may include a neural network, and the neural network may include a convolution layer, a pooling layer, an activation function layer, a normalization layer, a fully connected layer, etc. The video sample can be input into the neural network model to be trained, and the input low-quality video sample can be adjusted according to the neural network parameters in the neural network model to be trained to obtain the predicted quality level corresponding to the video sample.

[0131] In an exemplary embodiment of the present disclosure, the first loss function of the neural network model to be trained can be obtained according to the quality grade label and the predicted quality grade. Wherein, the loss function is a function that maps the value of a random event or its related random variable to a non-negative real number to represent the "loss" of the random event, and is used to evaluate the degree to which the predicted value of the model (the predicted quality grade of the video sample) and the true value (the quality grade label of the video sample) are different. The lower the loss function, the better the performance of the model. Specifically, the first loss function includes the difference between the predicted quality grade of the video sample and the quality grade label of the video sample. The first loss function can represent the predicted difference of the neural network model to be trained in predicting the quality grade of the video sample. Furthermore, the first loss function can include an absolute value loss function, a log loss function, a mean square error loss function, an exponential loss function, a Hinge loss function, a perceptual loss function, a cross entropy loss function, etc. It should be noted that the present disclosure does not specifically limit the specific form of the first loss function.

[0132] In an exemplary embodiment of the present disclosure, Y_c_p red is used as the predicted quality level corresponding to the video sample, and Y_c is used as the quality level label corresponding to the video sample. The expression of the first loss function loss1 is as follows:

[0133] loss1=log(Y_c_pred)*Y_c

[0134] In an exemplary embodiment of the present disclosure, after obtaining the first loss function through the above steps, the neural network parameters of the neural network model to be trained can be iteratively updated according to the first loss function to train the neural network model to be trained. Specifically, the neural network parameters of the convolutional neural network of the neural network model to be trained can be initialized first, and the output value can be obtained through multiple layers in the convolutional neural network (such as convolutional layer, deconvolution layer, fully connected layer, etc.), and the neural network parameters are adjusted according to the first loss function, and the neural network parameters are iteratively updated. When the first loss function converges, the training ends. Specifically, the entire model can be back-propagated according to the first loss function to obtain the gradient of each neural network parameter in the neural network model to be trained, and the neural network model to be trained is optimized during the iterative update process to make the value of the first loss function smaller and smaller.

[0135] Furthermore, the convolution layer can upsample the video samples, and the deconvolution layer can downsample the video samples, wherein the convolution layer and the deconvolution layer can correspond one to one and be interconnected to facilitate signal transmission between the convolution layer and the deconvolution layer. When training the neural network model to be trained, the signal can be transmitted from the convolution layer to the deconvolution layer, or the signal can be transmitted from the deconvolution layer to the convolution layer, which can speed up the training process of the neural network model to be trained.

[0136] It should be noted that the present disclosure does not specifically limit the structure of the neural network model to be trained and the method of training the neural network model to be trained according to the first loss function.

[0137] Furthermore, when the video sample has a quality grade label, the predicted quality grade score corresponding to the video sample can be obtained through the neural network model to be trained, and the predicted quality grade score can be used to indicate the quality grade of the video, wherein the predicted quality grade score has a score range. For example, the score range of the predicted quality grade score is 0 to 1, and the predicted quality grade score is 0.7. It should be noted that the present disclosure does not specifically limit the specific numerical value of the score range corresponding to the predicted quality grade score.

[0138] In an exemplary embodiment of the present disclosure, the text features, audio features, and image features corresponding to the video sample can be obtained through the neural network model to be trained, and the predicted quality level corresponding to the video sample can be obtained according to the text features, audio features, and image features corresponding to the video sample. Figure 6 As shown, obtaining the predicted quality level corresponding to the video sample according to the text features, audio features and image features corresponding to the video sample may include the following steps S610-S620:

[0139] Step S610, obtaining text features, audio features, and image features corresponding to the video sample through the neural network model to be trained;

[0140] In an exemplary embodiment of the present disclosure, after obtaining the video sample through the above steps, the text features, audio features and image features corresponding to the video sample may be obtained.

[0141] Specifically, the text information corresponding to the video sample can be obtained through a text extraction operation, and the text information corresponding to the video sample can be converted into text features. For example, the text information related to the video sample can be obtained according to the video title, OCR text recognition and speech recognition of the video image frame, and the obtained text information can be converted into text features that can be received by the neural network model to be trained, for example, the obtained text information can be converted into text vector features.

[0142] Secondly, the audio information corresponding to the video sample can be obtained through an audio extraction operation, and the audio information corresponding to the video sample can be converted into audio features. For example, the audio information corresponding to the video sample can be obtained through an audio capture operation, and the obtained audio information can be converted into audio features that can be received by the neural network model to be trained, for example, the obtained audio information can be converted into audio vector features.

[0143] In addition, the image information corresponding to the video sample can be obtained through the image acquisition operation, and the image information corresponding to the video sample can be converted into image features. For example, 8 frames of images can be randomly selected from the video sample to obtain the image information corresponding to the video sample, and the obtained image information can be converted into image features that can be received by the neural network model to be trained. For example, the obtained image information can be converted into image vector features, and then the obtained 8 groups of image vector features are added and averaged to obtain average image vector features, and the average image vector features are used as the image features corresponding to the video sample.

[0144] It should be noted that the present disclosure does not specifically limit the specific method of obtaining text features, audio features, and image features corresponding to video samples.

[0145] Step S620, obtaining a predicted quality level corresponding to the video sample according to the text features, audio features and image features corresponding to the video sample.

[0146] In an exemplary embodiment of the present disclosure, after obtaining the text features, audio features and image features corresponding to the video sample through the above steps, the predicted quality level corresponding to the video sample can be obtained based on the text features, audio features and image features corresponding to the video sample. Specifically, the text features, audio features and image features corresponding to the video sample can be used as inputs to the neural network model to be trained to predict the quality level corresponding to the video sample. It should be noted that the present disclosure does not specifically limit the specific method of obtaining the predicted quality level corresponding to the video sample based on the text features, audio features and image features corresponding to the video sample.

[0147] Through the above steps S610 to S620, the text features, audio features and image features corresponding to the video sample can be obtained through the neural network model to be trained, and the predicted quality level corresponding to the video sample can be obtained based on the text features, audio features and image features corresponding to the video sample.

[0148] In an exemplary embodiment of the present disclosure, when it is determined through the above steps that the video sample does not have a quality level label, the first enhanced sample can be input into the neural network model to be trained, and the predicted exposure times corresponding to the first enhanced sample can be obtained through the neural network model to be trained.

[0149] In an exemplary embodiment of the present disclosure, since the first enhanced sample is obtained by processing the video sample, the exposure times label of the first enhanced sample is the same as the exposure times label of the video sample, so the fourth loss function of the neural network model to be trained can be generated according to the exposure times label and the predicted exposure times corresponding to the first enhanced sample. Among them, the fourth loss function is a function that maps the value of a random event or its related random variable to a non-negative real number to represent the "loss" of the random event, which is used to evaluate the degree to which the predicted value of the model (the predicted exposure times corresponding to the first enhanced sample) and the true value (exposure times label) are different. The lower the loss function, the better the performance of the model. Specifically, the fourth loss function includes the difference between the predicted exposure times corresponding to the first enhanced sample and the exposure times label. The fourth loss function can represent the difference between the predicted result of the neural network model to be trained in predicting the exposure times of the first enhanced sample and the predicted exposure times label corresponding to the first enhanced sample. Further, the fourth loss function can include an absolute value loss function, a logarithmic loss function, a mean square error loss function, an exponential loss function, a Hinge loss function, a perceptual loss function, a cross entropy loss function, etc. It should be noted that the present disclosure does not specifically limit the specific form of the fourth loss function.

[0150] In two example embodiments of the present disclosure, after obtaining the second loss function and the fourth loss function through the above steps, the neural network model to be trained can be trained according to the fourth loss function and the second loss function. Specifically, the second overall loss function can be obtained according to the fourth loss function and the second loss function, and the neural network model to be trained can be trained according to the second overall loss function. Specifically, the neural network parameters can be adjusted according to the second overall loss function, and the neural network parameters can be iteratively updated. When the second overall loss function converges, the training ends. Specifically, the entire model can be back-propagated according to the second overall loss function to obtain the gradient of each neural network parameter in the neural network model to be trained, and the neural network model to be trained can be optimized during the iterative update process so that the value of the second overall loss function becomes smaller and smaller. It should be noted that the present disclosure does not specifically limit the specific method of training the neural network model to be trained according to the fourth loss function and the second loss function.

[0151] In an exemplary embodiment of the present disclosure, when it is determined through the above steps that the video sample has a quality level label, the video sample can be input into the neural network model to be trained, and the predicted exposure times corresponding to the video sample can be obtained through the neural network model to be trained. Among them, the video sample has an exposure times label. Specifically, the exposure times label can be used to indicate the exposure times of the video sample, and the exposure times can include the exposure times within a certain period of time. For example, the exposure times label can be the exposure times within 3 days of uploading the video sample. Furthermore, the exposure times label can directly use a numerical value to represent the exposure times, or it can also represent the exposure times in other forms. For example, the exposure times can be divided into multiple levels, such as high exposure, normal exposure, low exposure, etc. It should be noted that the present disclosure does not specifically limit the specific form of the exposure times label.

[0152] In an exemplary embodiment of the present disclosure, a third loss function of the neural network model to be trained can be generated based on the exposure times label and the predicted exposure times. Among them, the third loss function is a function that maps the value of a random event or its related random variable to a non-negative real number to represent the "loss" of the random event, which is used to evaluate the degree to which the predicted value (predicted exposure times) of the model is different from the true value (exposure times label). The lower the loss function, the better the performance of the model. Specifically, the third loss function includes the difference between the predicted exposure times of the video sample and the exposure times label of the video sample. The third loss function can represent the predicted difference between the result of the neural network model to be trained in predicting the exposure times of the video sample and the exposure times label corresponding to the video sample. Further, the third loss function can include an absolute value loss function, a logarithmic loss function, a mean square error loss function, an exponential loss function, a Hinge loss function, a perceptual loss function, a cross entropy loss function, etc. It should be noted that the present disclosure does not specifically limit the specific form of the third loss function.

[0153] Furthermore, the first loss function can be assigned a first weight p1, the third loss function can be assigned a second weight p2, Y_c_p red is used as the predicted quality level corresponding to the video sample, Y_c is used as the quality level label corresponding to the video sample, Y_r_p red is used as the predicted exposure number corresponding to the video sample, and Y_r is used as the exposure number label corresponding to the video sample. Then, the expression of the first overall loss function loss4 is as follows:

[0154] loss4=p1(log(Y_c_pred)*Y_c)+p2|Y_r-Y_r_p red|

[0155] In an exemplary embodiment of the present disclosure, after obtaining the first loss function and the third loss function through the above steps, the neural network model to be trained can be trained according to the third loss function and the first loss function. Specifically, the first overall loss function can be obtained according to the third loss function and the first loss function, and the neural network model to be trained can be trained according to the first overall loss function. Specifically, the neural network parameters can be adjusted according to the first overall loss function, and the neural network parameters can be iteratively updated. When the first overall loss function converges, the training ends. Specifically, the entire model can be back-propagated according to the first overall loss function to obtain the gradient of each neural network parameter in the neural network model to be trained, and the neural network model to be trained can be optimized during the iterative update process so that the value of the first overall loss function becomes smaller and smaller. It should be noted that the present disclosure does not specifically limit the specific method of training the neural network model to be trained according to the third loss function and the first loss function.

[0156] Furthermore, when the video sample has a quality level label, the predicted exposure score corresponding to the video sample can be obtained through the neural network model to be trained, and the predicted exposure score can be used to indicate the number of exposures of the video, wherein the predicted exposure score has a score range. For example, the score range of the predicted exposure score is 0 to 1, and the predicted exposure score is 0.7. It should be noted that the present disclosure does not specifically limit the specific numerical value of the score range corresponding to the predicted exposure score.

[0157] In an exemplary embodiment of the present disclosure, Y_s_c_pred may be used as the predicted quality level corresponding to the second enhanced sample, Y_w_c_pred may be used as the reference quality level label of the second enhanced sample, Y_r may be used as the exposure number label corresponding to the first enhanced sample, and Y_r_p red may be used as the predicted exposure number corresponding to the first enhanced sample. Then:

[0158] When the probability of the reference quality level label Y_w_c_pred (the target predicted sub-quality level corresponding to the first enhanced sample) of the second enhanced sample is greater than the confidence probability value, the third weight p3 is assigned to the second loss function, and the fourth weight p4 is assigned to the fourth loss function. The expression of the second overall loss function loss2 is as follows:

[0159] loss5=p3(log(Y_s_c_p red)*argmax(Y_w_c_pred))+p4|Y_r-Y_r_p red|

[0160] When the probability of the reference quality level label Y_w_c_pred (the target predicted sub-quality level corresponding to the first enhanced sample) of the second enhanced sample is less than the confidence probability value, the fifth weight p5 is assigned to the second loss function, and the sixth weight p6 is assigned to the fourth loss function. The expression of the second overall loss function loss6 is as follows:

[0161] loss6=p5(log(Y_s_c_pred)*argmax(Y_w_c_p red))+p6|Y_r-Y_r_p red|

[0162] In an exemplary embodiment of the present disclosure, a video to be recommended may be obtained, and the video to be recommended may be input into a video quality determination model to obtain a quality level of the video to be recommended, and the video to be recommended may be recommended according to the quality level of the video to be recommended. Figure 7 The embodiments describe steps S710 to S720 of training the video quality determination model in the present exemplary embodiment in more detail.

[0163] Step S710, obtaining a video to be recommended, and inputting the video to be recommended into a video quality determination model to obtain a quality level of the video to be recommended;

[0164] Step S720: recommend the video to be recommended according to the quality level of the video to be recommended.

[0165] In an exemplary embodiment of the present disclosure, a video to be recommended may be obtained and input into a video quality determination model to obtain a quality level of the video to be recommended. The video quality determination model is obtained by the training method of the above-mentioned video quality determination model. Specifically, the quality level of the video to be recommended may be used to indicate whether the content of the video to be recommended is rich, interesting, creative, etc., and may be used to indicate the audience's acceptance of the recommended video.

[0166] It should be noted that the present disclosure does not specifically limit the evaluation method of quality grade. For example, the quality grade can be divided into high quality grade, general quality grade, and low quality grade; or the quality grade can be divided into excellent, ordinary, gray, and poor quality.

[0167] After obtaining the quality level of the video to be recommended through the above steps, the video to be recommended can be recommended according to the quality level of the video to be recommended. For example, the quality levels of multiple videos to be recommended can be obtained through the above steps, and the multiple videos to be recommended can be sorted according to the quality level, and the videos to be recommended with higher quality levels can be recommended first; for another example, only the videos to be recommended with excellent quality levels can be recommended. It should be noted that the present disclosure does not specifically limit the specific method of recommending videos to be recommended according to the quality level of the video to be recommended.

[0168] Through the above steps S710 to S720, the video to be recommended can be obtained, and the video to be recommended is input into the video quality determination model to obtain the quality level of the video to be recommended, and the video to be recommended is recommended according to the quality level of the video to be recommended.

[0169] In an exemplary embodiment of the present disclosure, multiple videos to be recommended and their corresponding quality levels may be obtained, the videos to be recommended may be sorted according to their corresponding quality levels to obtain a first sorting result, and the videos to be recommended may be recommended according to the first sorting result. Figure 8 As shown, sorting the recommended videos according to the quality levels corresponding to the plurality of recommended videos to obtain a first sorting result, and recommending the recommended videos according to the first sorting result may include the following steps S810-S820:

[0170] Step S810, obtaining a plurality of videos to be recommended and quality levels corresponding to the videos to be recommended, and sorting the videos to be recommended according to the quality levels corresponding to the plurality of videos to be recommended to obtain a first sorting result;

[0171] Step S820: recommend the video to be recommended according to the first sorting result.

[0172] In an exemplary embodiment of the present disclosure, multiple videos to be recommended and quality levels corresponding to the videos to be recommended may be obtained. Specifically, the multiple videos to be recommended may be input into a video quality determination model to obtain quality levels corresponding to the multiple videos to be recommended, and then the multiple videos to be recommended may be sorted according to the quality levels corresponding to the multiple videos to be recommended obtained in the above steps to obtain a first sorting result, and then the videos to be recommended may be recommended according to the first sorting result.

[0173] For example, input the video A to be recommended, the video B to be recommended, the video C to be recommended, and the video D to be recommended into the video quality determination model to obtain the video A to be recommended-ordinary, the video B to be recommended-excellent, the video C to be recommended-poor, and the video D to be recommended-grayscale. At this time, the recommended videos A, B to be recommended, C to be recommended, and D to be recommended can be sorted according to the quality level to obtain the first sorting result: video B to be recommended, video A to be recommended, video D to be recommended, and video C to be recommended. At this time, the video B to be recommended with a better quality level can be recommended first. It should be noted that the present disclosure does not specifically limit the specific method of recommending the recommended videos according to the first sorting result.

[0174] In an exemplary embodiment of the present disclosure, the quality grade scores corresponding to multiple videos to be recommended can be obtained through a video quality determination model, and the multiple videos to be recommended can be sorted according to the quality grade scores corresponding to the videos to be recommended to obtain a first sorting result. For example, the video to be recommended A, the video to be recommended B, the video to be recommended C, and the video to be recommended D are input into the video quality determination model to obtain the video to be recommended A-0.6, the video to be recommended B-0.9, the video to be recommended C-0.3, and the video to be recommended D-0.5. At this time, the video to be recommended A, the video to be recommended B, the video to be recommended C, and the video to be recommended D can be sorted according to the quality grade scores to obtain the first sorting result: the video to be recommended B, the video to be recommended A, the video to be recommended D, and the video to be recommended C. It should be noted that the present disclosure does not specifically limit the specific method of sorting the recommended videos according to the quality grade scores corresponding to the multiple videos to be recommended to obtain the first sorting result.

[0175] Through the above steps S810 to S820, multiple videos to be recommended and their corresponding quality levels can be obtained, the videos to be recommended are sorted according to their corresponding quality levels to obtain a first sorting result, and the videos to be recommended are recommended according to the first sorting result.

[0176] In an exemplary embodiment of the present disclosure, Fig. 9 As shown, the text information, audio information and image information corresponding to the video sample can be obtained, and the text information, audio information and image information can be input into the video quality determination model to obtain the predicted quality level and predicted exposure times corresponding to the video sample, and then the recommendation score is calculated according to the predicted quality level and predicted exposure times corresponding to the video sample, and the cold start traffic is allocated according to the recommendation score.

[0177] In an exemplary embodiment of the present disclosure, Fig.10 As shown, if the video sample does not have a quality level label, the image information corresponding to the video sample can be obtained, and a first enhanced sample and a second enhanced sample can be created according to the image information. The first enhanced sample and the second enhanced sample are respectively input into the video quality determination model to obtain the predicted exposure times of the first enhanced sample, and the predicted quality level of the first enhanced sample and the probability of the predicted quality level are obtained. When the probability of the predicted quality level is greater than the confidence probability value, the predicted quality level is determined as the reference quality level label of the second enhanced sample (the video quality determination model is trained by cross entropy). When the probability corresponding to the predicted quality level is less than the confidence probability value, the distance between the first enhanced sample and the second enhanced sample is shortened by the KL divergence objective function (the video quality determination model is trained by relative entropy).

[0178] In an exemplary embodiment of the present disclosure, the exposure times of the video to be recommended can be determined by a video quality determination model, and the video to be recommended can be recommended according to the quality level of the video to be recommended and the exposure times of the video to be recommended. Fig.11 As shown, recommending a video to be recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended may include the following steps S1110 to S1120:

[0179] Step S1110, determining the exposure times of the video to be recommended by using a video quality determination model;

[0180] Step S1120: recommending the video to be recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended.

[0181] In an exemplary embodiment of the present disclosure, after the quality grade of the video to be recommended is obtained through the above steps, the number of exposures of the video to be recommended can be obtained, and then the video to be recommended can be recommended based on the quality grade of the video to be recommended and the number of exposures of the video to be recommended. Specifically, an exposure number recommendation threshold and a quality grade recommendation threshold can be set, and the video to be recommended is recommended only when the quality grade of the video to be recommended and the number of exposures of the video to be recommended are greater than the quality grade recommendation threshold and the exposure number recommendation threshold, respectively; or, the total recommendation score can be calculated based on the quality grade of the video to be recommended and the number of exposures of the video to be recommended, and the video to be recommended is recommended when the total recommendation score is greater than the recommendation threshold. It should be noted that the present disclosure does not specifically limit the specific method of recommending the video to be recommended based on the quality grade of the video to be recommended and the number of exposures of the video to be recommended.

[0182] Through the above steps S1110 to S1120, the exposure times of the video to be recommended can be determined by the video quality determination model, and the video to be recommended can be recommended according to the quality level of the video to be recommended and the exposure times of the video to be recommended.

[0183] In an exemplary embodiment of the present disclosure, multiple videos to be recommended and their corresponding quality levels and exposure times may be obtained, the videos to be recommended may be sorted according to their corresponding quality levels and exposure times to obtain a second sorting result, and the videos to be recommended may be recommended according to the second sorting result. Fig.12 As shown, recommending the video to be recommended according to the second sorting result may include the following steps S1210-S1220:

[0184] Step S1210, obtaining a plurality of videos to be recommended and quality levels and exposure times corresponding to the videos to be recommended, and sorting the videos to be recommended according to the quality levels and exposure times corresponding to the plurality of videos to be recommended to obtain a second sorting result;

[0185] Step S1220: recommend the video to be recommended according to the second sorting result.

[0186] In an exemplary embodiment of the present disclosure, after obtaining multiple videos to be recommended and the quality levels and exposure times corresponding to the videos to be recommended through the above steps, the videos to be recommended can be sorted according to the quality levels and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result, and the videos to be recommended can be recommended according to the second sorting result. Specifically, the multiple videos to be recommended can be sorted first based on the quality level of the videos to be recommended, and when there are videos to be recommended with the same quality level, the videos to be recommended with the same quality level can be sorted based on the exposure times of the videos to be recommended to obtain the second sorting result.

[0187] For example, through the above steps, multiple videos to be recommended and their corresponding quality levels and exposure times are obtained: video A to be recommended-high quality level-2000 times, video B to be recommended-high quality level-3000 times, video C to be recommended-medium quality level-200 times. The recommended video A, video B, and video C to be recommended can be sorted according to their quality levels to obtain video A to be recommended, video B to be recommended, and video C to be recommended, and then the videos to be recommended with the same quality level can be sorted according to their exposure times to obtain video B to be recommended, video A to be recommended, and video C to be recommended.

[0188] It should be noted that the present disclosure does not specifically limit the specific manner of recommending the recommended videos according to the second ranking result.

[0189] In an exemplary embodiment of the present disclosure, the quality grade scores and exposure number scores corresponding to multiple videos to be recommended can be obtained through a video quality determination model, and the multiple videos to be recommended can be sorted according to the quality grade scores and exposure number scores corresponding to the videos to be recommended to obtain a second sorting result. For example, input the video A to be recommended, the video B to be recommended, the video C to be recommended, and the video D to be recommended into the video quality determination model to obtain the video A to be recommended-0.6 (quality grade score)-0.5 (exposure score), the video B to be recommended-0.9-0.8, the video C to be recommended-0.3-0.5, and the video D to be recommended-0.5-0.4. At this time, the video A to be recommended, the video B to be recommended, the video C to be recommended, and the video D to be recommended can be sorted according to the quality grade score and the exposure score to obtain a second sorting result. For example, the quality grade score and the exposure score of the video to be recommended can be added to obtain the total score of the video to be recommended: video A to be recommended-1.1, video B to be recommended-1.7, video C to be recommended-0.8, video D to be recommended-0.9. According to the total score of the video to be recommended, the videos to be recommended are sorted to obtain the second sorting result: video B to be recommended, video A to be recommended, video D to be recommended, and video C to be recommended. It should be noted that the present disclosure does not specifically limit the specific method of sorting the recommended videos according to the quality grade scores and exposure number scores corresponding to the multiple recommended videos to obtain the second sorting result.

[0190] Through the above steps S1210 to S1220, multiple videos to be recommended and the quality levels and exposure times corresponding to the videos to be recommended can be obtained, the videos to be recommended are sorted according to the quality levels and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result, and the videos to be recommended are recommended according to the second sorting result.

[0191] In an exemplary embodiment of the present disclosure, a video recommendation basic score can be obtained, a quality grade score threshold and an adjustment weight corresponding to a quality grade range distinguished by the quality grade score threshold can be obtained, an exposure number score threshold and an adjustment weight corresponding to a quality grade range distinguished by the exposure number score threshold can be obtained, and a recommendation score of the video to be recommended can be determined based on the video recommendation basic score, the adjustment weight corresponding to the quality grade score threshold, and the adjustment weight corresponding to the exposure number score threshold, and the recommendation score of the video to be recommended can be sorted based on the recommendation score to obtain a second sorting result. Fig.13 As shown, sorting the recommended videos according to the recommendation scores to obtain a second sorting result may include the following steps S1310 to S1320:

[0192] Step S1310, obtaining a basic score for video recommendation;

[0193] In an exemplary embodiment of the present disclosure, the video recommendation basic score refers to a basic score that can be used when recommending a video, and the recommendation score of the video to be recommended is adjusted by superimposing different coefficients on the video recommendation basic score. It should be noted that the present disclosure does not specifically limit the specific value of the video recommendation basic score.

[0194] Step S1320, obtaining a quality grade score threshold and an adjustment weight corresponding to a quality grade range differentiated by the quality grade score threshold, obtaining an exposure count score threshold and an adjustment weight corresponding to a quality grade range differentiated by the exposure count score threshold;

[0195] Step S1330, determining the recommendation score of the video to be recommended according to the video recommendation basic score, the adjustment weight corresponding to the quality level score threshold, and the adjustment weight corresponding to the exposure count score threshold;

[0196] In an exemplary embodiment of the present disclosure, the adjustment weight corresponding to the quality grade score threshold and the quality grade range distinguished by the quality grade score threshold means that within the quality grade score range, when the quality grade score is lower than the quality grade score threshold, an adjustment weight can be set to adjust the video recommendation basic score to obtain an intermediate recommendation score, and when the quality grade score is higher than the quality grade score threshold, an adjustment weight can be set to adjust the video recommendation basic score to obtain an intermediate recommendation score. The adjustment weight corresponding to the exposure count score threshold and the exposure count range distinguished by the exposure count score threshold means that within the exposure count score range, when the exposure count score is lower than the exposure count score threshold, an adjustment weight can be set to adjust the intermediate recommendation score to obtain a recommendation score, and when the exposure count score is higher than the exposure count score threshold, an adjustment weight can be set to adjust the intermediate recommendation score to obtain a recommendation score.

[0197] Step S1340: sort the multiple videos to be recommended according to the recommendation scores to obtain a second sorting result.

[0198] In an exemplary embodiment of the present disclosure, after obtaining the recommendation scores of the multiple videos to be recommended through the above steps, the multiple videos to be recommended may be sorted according to the recommendation scores to obtain a second sorting result.

[0199] For example, the multiple recommended videos-quality grade scores-exposure times scores obtained through the above steps are: recommended video A-0.5-0.6, recommended video B-0.8-0.8, recommended video C-0.5-0.8, the quality grade score threshold is set to 0.6, the adjustment weight corresponding to the quality grade score below the quality grade score threshold is 0.5, the adjustment weight corresponding to the quality grade score below the quality grade score threshold is 1.5, the exposure times score is 0.7, the adjustment weight corresponding to the exposure times score below the exposure times score threshold is 0.5, and the adjustment weight corresponding to the exposure times score below the exposure times score threshold is 1.5. The adjustment weight corresponding to the exposure score of the score threshold is 1.5, and the basic score for video recommendation is 300. For video A to be recommended, its quality grade score is 0.5, which is less than the quality grade score threshold of 0.6, that is, the adjustment weight 0.5 is multiplied by the basic score for video recommendation to obtain an intermediate recommendation score of 150, and the exposure score of video A to be recommended, 0.6, is less than the exposure score threshold of 0.7, then the adjustment weight 0.5 is multiplied by the intermediate recommendation score 150 to obtain a recommendation score of 75, and similarly, the recommendation score of video B to be recommended is 675 and the recommendation score of video C to be recommended is 225. At this time, video A to be recommended, video B to be recommended, and video C to be recommended can be sorted according to the recommendation score to obtain a second sorting result: video B to be recommended, video C to be recommended, and video A to be recommended. It should be noted that the present disclosure does not specifically limit the specific method of sorting multiple videos to be recommended according to the recommendation score to obtain the second sorting result.

[0200] Through the above steps S1310 to S1340, the basic score for video recommendation can be obtained, the quality grade score threshold and the adjustment weight corresponding to the quality grade range distinguished by the quality grade score threshold can be obtained, the exposure number score threshold and the adjustment weight corresponding to the quality grade range distinguished by the exposure number score threshold can be obtained, the recommendation score of the video to be recommended is determined according to the basic score for video recommendation, the adjustment weight corresponding to the quality grade score threshold and the adjustment weight corresponding to the exposure number score threshold, and the videos to be recommended are sorted according to the recommendation scores to obtain a second sorting result.

[0201] In a video quality determination model training method and a video recommendation method provided by an embodiment of the present disclosure, the video quality determination model can be trained when the video does not have a quality grade label, and the video can be recommended based on the trained video quality model. On the one hand, the video quality model can be trained without relying on the quality grade label of the video, avoiding the problem that the video quality determination model cannot be trained due to the lack of quality grade labels, thereby improving the accuracy of the trained video quality determination model in determining the quality grade of new videos; on the other hand, recommending new videos based on the trained video quality determination model can highlight new videos with better quality among multiple newly uploaded videos, thereby more reasonably allocating recommended traffic, and thus enabling users to see new videos with better quality; on the other hand, recommending new videos based on the trained video quality determination model can greatly enhance the creative enthusiasm of video creators, thereby improving the quality and quantity of creation, and thus enhancing the influence of the video platform.

[0202] Fig.14 FIG. 1 is a block diagram of a training device for a video quality determination model according to an exemplary embodiment. Fig.14 The video quality determination model training device 1400 includes a sample acquisition unit 1410, an enhanced sample generation unit 1420, a reference quality level label generation unit 1430, a second loss function determination unit 1440, and a neural network model training unit 1450.

[0203] Among them, the sample acquisition unit is configured to execute acquisition of video samples; the enhancement sample generation unit is configured to execute, when the video sample does not have a quality level label, generating corresponding first enhancement samples and second enhancement samples according to the video sample; wherein the degree of change of the first enhancement sample relative to the video sample is less than the degree of change of the second enhancement sample relative to the video sample; the reference quality level label generation unit is configured to execute obtaining the predicted quality level corresponding to the first enhancement sample through the neural network model to be trained, and determining the reference quality level label of the second enhancement sample according to the predicted quality level corresponding to the first enhancement sample; the second loss function determination unit is configured to execute obtaining the predicted quality level corresponding to the second enhancement sample through the neural network model to be trained, and determining the second loss function according to the predicted quality level corresponding to the second enhancement sample and the reference quality level label of the second enhancement sample; the neural network model training unit is configured to execute training of the neural network model to be trained according to the second loss function to obtain a video quality determination model.

[0204] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the training device of the video quality determination model also includes: a first loss function determination unit, which is configured to execute when the video sample has a quality level label, obtain the predicted quality level corresponding to the video sample through the neural network model to be trained, generate a first loss function of the neural network model to be trained according to the quality level label and the predicted quality level, and train the neural network model to be trained according to the first loss function.

[0205] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the predicted quality level corresponding to the video sample is obtained through the neural network model to be trained, and the training device of the video quality determination model also includes: a feature acquisition unit, configured to execute the acquisition of text features, audio features and image features corresponding to the video sample through the neural network model to be trained; a quality prediction unit, configured to execute the prediction quality level corresponding to the video sample according to the text features, audio features and image features corresponding to the video sample.

[0206] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, corresponding first enhanced samples and second enhanced samples are generated according to video samples, and the training device of the video quality determination model also includes: an image sample acquisition unit, configured to execute acquisition of image samples corresponding to the video samples; an enhanced sample acquisition unit, configured to execute one or more image enhancement processing operations on the image samples to obtain the first enhanced samples and the second enhanced samples corresponding to the video samples.

[0207] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the predicted quality level corresponding to the first enhanced sample is obtained by the neural network model to be trained, and the reference quality level label of the second enhanced sample is determined according to the predicted quality level. The training device of the video quality determination model also includes: a predicted quality probability acquisition unit, which is configured to execute the predicted quality level corresponding to the first enhanced sample and the probability corresponding to the predicted quality level by the neural network model to be trained; and a first reference label determination unit, which is configured to determine the reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample when the probability corresponding to the predicted quality level is greater than the confidence probability value.

[0208] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the predicted quality level corresponding to the first enhanced sample includes at least two predicted sub-quality levels, and when the probability corresponding to the predicted quality level is greater than the confidence probability value, the predicted quality level corresponding to the first enhanced sample is determined as the reference quality level label of the second enhanced sample, and the training device of the video quality determination model also includes: a target sub-quality level determination unit, configured to execute acquisition of at least two predicted sub-quality levels and the probability corresponding to each predicted sub-quality level, and determine the predicted sub-quality level with the largest probability as the target predicted sub-quality level; a second reference label determination unit, configured to execute when the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value, determine the target predicted sub-quality level as the reference quality level label of the second enhanced sample.

[0209] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the training device of the video quality determination model also includes: an enhanced sample prediction quality determination unit, configured to execute when the probability corresponding to the prediction quality level is less than the confidence probability value, to obtain the prediction quality level corresponding to the second enhanced sample through the neural network model to be trained; a second loss function determination subunit, configured to execute the second loss function according to the prediction quality level corresponding to the first enhanced sample and the prediction quality level corresponding to the second enhanced sample.

[0210] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the predicted quality corresponding to the video sample is obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: a predicted quality score acquisition unit, which is configured to execute the predicted quality grade score corresponding to the video sample obtained by the neural network model to be trained; wherein the predicted quality grade score has a score range.

[0211] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the video sample has an exposure number label, and the training device of the video quality determination model also includes: a quality level label training unit, configured to execute the predicted exposure number corresponding to the first enhanced sample obtained by the neural network model to be trained, generate a fourth loss function of the neural network model to be trained according to the exposure number label corresponding to the first enhanced sample and the predicted exposure number corresponding to the first enhanced sample, and train the neural network model to be trained according to the fourth loss function and the second loss function.

[0212] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the video sample has an exposure number label, and the training device of the video quality determination model also includes: a training unit without a quality level label, configured to execute the predicted exposure number corresponding to the video sample through the neural network model to be trained, generate a third loss function of the neural network model to be trained according to the exposure number label and the predicted exposure number, and train the neural network model to be trained according to the third loss function and the first loss function.

[0213] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the predicted exposure times corresponding to the video sample are obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: a predicted exposure times score acquisition unit, which is configured to execute the predicted exposure times score corresponding to the video sample by the neural network model to be trained; wherein the predicted exposure times score has a score range.

[0214] Fig.15 FIG. 1 is a block diagram of a video recommendation device according to an exemplary embodiment. Fig.15 The video recommendation device 1500 includes a to-be-recommended video acquisition unit 1510 and a video recommendation unit 1520 .

[0215] Among them, the video acquisition unit to be recommended is configured to execute the video acquisition unit to be recommended, and is configured to execute the acquisition of the video to be recommended, and input the video to be recommended into the video quality determination model to obtain the quality level of the video to be recommended; wherein the video quality determination model is obtained by the training method of the video quality determination model as mentioned above; the video recommendation unit is configured to execute the recommendation of the video to be recommended according to the quality level of the video to be recommended.

[0216] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the video to be recommended is recommended according to the quality level of the video to be recommended, and the video recommendation device also includes: a quality level sorting unit, configured to execute acquisition of multiple videos to be recommended and the quality levels corresponding to the videos to be recommended, and to sort the videos to be recommended according to the quality levels corresponding to the multiple videos to be recommended to obtain a first sorting result; a first sorting unit, configured to execute recommendation of the video to be recommended according to the first sorting result.

[0217] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the video to be recommended has a quality grade score, and the videos to be recommended are sorted according to the quality grades corresponding to the multiple videos to be recommended to obtain a first sorting result. The video recommendation device also includes: a quality grade score sorting unit, which is configured to execute sorting of the videos to be recommended according to the quality grade scores corresponding to the multiple videos to be recommended to obtain a first sorting result.

[0218] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the video to be recommended is recommended according to the quality level of the video to be recommended, and the video recommendation device also includes: an exposure times acquisition unit, configured to determine the exposure times of the video to be recommended through a video quality determination model; a comprehensive recommendation unit, configured to recommend the video to be recommended according to the quality level of the video to be recommended and the exposure times of the video to be recommended.

[0219] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the video to be recommended is recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended, and the video recommendation device also includes: a second sorting result acquisition unit, configured to execute acquisition of multiple videos to be recommended and the quality levels and exposure times corresponding to the videos to be recommended, and to sort the videos to be recommended according to the quality levels and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result; and a second sorting recommendation unit, configured to execute recommendation of the video to be recommended according to the second sorting result.

[0220] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the videos to be recommended have quality grade scores and exposure times scores, and the videos to be recommended are sorted according to the quality grades and exposure times corresponding to the multiple videos to be recommended to obtain a second sorting result. The video recommendation device also includes: a second sorting result acquisition subunit, which is configured to execute sorting of the videos to be recommended according to the quality grade scores and exposure times scores corresponding to the multiple videos to be recommended to obtain a second sorting result.

[0221] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the videos to be recommended are sorted according to the quality grade scores and exposure times scores corresponding to multiple videos to be recommended to obtain a second sorting result, and the video recommendation device also includes: a video recommendation basic score acquisition unit, configured to execute acquisition of the video recommendation basic score; an adjustment weight acquisition unit, configured to execute acquisition of the quality grade score threshold and the adjustment weight corresponding to the quality grade range distinguished by the quality grade score threshold, and acquisition of the exposure times score threshold and the adjustment weight corresponding to the quality grade range distinguished by the exposure times score threshold; a recommendation score acquisition unit, configured to execute determination of the recommendation score of the video to be recommended according to the video recommendation basic score, the adjustment weight corresponding to the quality grade score threshold, and the adjustment weight corresponding to the exposure times score threshold; and a recommendation score sorting unit, configured to execute sorting of the multiple videos to be recommended according to the recommendation score to obtain the second sorting result.

[0222] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0223] Refer to the following Fig.1616 is a diagram to describe an electronic device 1600 according to such an embodiment of the present disclosure. Fig.16 The electronic device 1600 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0224] like Fig.16 As shown, the electronic device 1600 is in the form of a general computing device. The components of the electronic device 1600 may include, but are not limited to: the at least one processing unit 1610, the at least one storage unit 1620, a bus 1630 connecting different system components (including the storage unit 1620 and the processing unit 1610), and a display unit 1640.

[0225] The storage unit stores program codes, which can be executed by the processing unit 1610, so that the processing unit 1610 performs the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processing unit 1610 can perform the following steps: Figure 1 As shown in step S110, a video sample is obtained; in step S120, when the video sample does not have a quality level label, a corresponding first enhanced sample and a second enhanced sample are generated according to the video sample; wherein the degree of change of the first enhanced sample relative to the video sample is less than the degree of change of the second enhanced sample relative to the video sample; in step S130, a predicted quality level corresponding to the first enhanced sample is obtained by the neural network model to be trained, and a reference quality level label of the second enhanced sample is determined according to the predicted quality level corresponding to the first enhanced sample; in step S140, a predicted quality level corresponding to the second enhanced sample is obtained by the neural network model to be trained, and a second loss function is determined according to the predicted quality level corresponding to the second enhanced sample and the reference quality level label of the second enhanced sample; in step S150, the neural network model to be trained is trained according to the second loss function to obtain a video quality determination model.

[0226] Alternatively, the processing unit 1610 may also execute the following steps: Figure 7 In step S710 shown in the figure, the video to be recommended is obtained, and the video to be recommended is input into the video quality determination model to obtain the quality level of the video to be recommended; in step S720, the video to be recommended is recommended according to the quality level of the video to be recommended.

[0227] For example, electronic devices can achieve Figure 1 or Figure 7 The steps shown.

[0228] The storage unit 1620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 1621 and / or a cache memory unit 1622 , and may further include a read-only memory unit (ROM) 1623 .

[0229] The storage unit 1620 may also include a program / utility 1624 having a set (at least one) of program modules 1625, such program modules 1625 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0230] Bus 1630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0231] The electronic device 1600 may also communicate with one or more external devices 1670 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 1600, and / or communicate with any device that enables the electronic device 1600 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 1650. Furthermore, the electronic device 1600 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 1660. As shown, the network adapter 1660 communicates with other modules of the electronic device 1600 via a bus 1630. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 1600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0232] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiment of the present disclosure.

[0233] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory including instructions, and the above instructions can be executed by a processor of the device to complete the above method. Optionally, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0234] In an exemplary embodiment, a computer program product is also provided, including a computer program / instruction, which implements the video quality determination model training method or video recommendation method in the above embodiments when the computer program / instruction is executed by a processor.

[0235] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0236] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for training a video quality determination model, characterized in that: include: Get video samples; When the video sample does not have a quality level label, generating a corresponding first enhanced sample and a second enhanced sample according to the video sample; wherein the degree of change of the first enhanced sample relative to the video sample is less than the degree of change of the second enhanced sample relative to the video sample; Obtaining a predicted quality level corresponding to the first enhanced sample by using a neural network model to be trained, and determining a reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample; Obtaining a predicted quality level corresponding to the second enhanced sample by using the neural network model to be trained, and determining a second loss function according to the predicted quality level corresponding to the second enhanced sample and a reference quality level label of the second enhanced sample; Training the neural network model to be trained according to the second loss function to obtain a video quality determination model; The second loss function is further determined by the following steps: when the probability corresponding to the predicted quality level of the first enhanced sample is less than the confidence probability value, obtaining the predicted quality level corresponding to the second enhanced sample by the neural network model to be trained; The second loss function is determined according to the prediction quality level corresponding to the first enhanced sample and the prediction quality level corresponding to the second enhanced sample.

2. The video quality determination model training method according to claim 1, characterized in that: The training method further comprises: When the video sample has the quality level label, the predicted quality level corresponding to the video sample is obtained through the neural network model to be trained, and a first loss function of the neural network model to be trained is generated according to the quality level label and the predicted quality level, and the neural network model to be trained is trained according to the first loss function.

3. The video quality determination model training method according to claim 2, characterized in that: The step of obtaining the predicted quality level corresponding to the video sample through the neural network model to be trained comprises: Acquire text features, audio features, and image features corresponding to the video sample through the neural network model to be trained; The predicted quality level corresponding to the video sample is obtained according to the text features, audio features and image features corresponding to the video sample.

4. The video quality determination model training method according to claim 1, characterized in that: The step of generating the corresponding first enhanced sample and the corresponding second enhanced sample according to the video sample comprises: Obtaining an image sample corresponding to the video sample; One or more image enhancement processing operations are performed on the image sample to obtain a first enhanced sample and a second enhanced sample corresponding to the video sample.

5. The method for training a video quality determination model according to claim 1, characterized in that: The step of obtaining the predicted quality level corresponding to the first enhanced sample by using the neural network model to be trained, and determining the reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample comprises: Obtaining, by means of the neural network model to be trained, a predicted quality level corresponding to the first enhanced sample and a probability corresponding to the predicted quality level of the first enhanced sample; When the probability corresponding to the predicted quality level of the first enhanced sample is greater than the confidence probability value, the reference quality level label of the second enhanced sample is determined according to the predicted quality level corresponding to the first enhanced sample.

6. The video quality determination model training method according to claim 5, characterized in that: The predicted quality level corresponding to the first enhanced sample includes at least two predicted sub-quality levels, and when the probability corresponding to the predicted quality level is greater than the confidence probability value, the step of determining the predicted quality level corresponding to the first enhanced sample as the reference quality level label of the second enhanced sample comprises: Acquire the at least two predicted sub-quality levels and the probabilities corresponding to each of the predicted sub-quality levels, and determine the predicted sub-quality level with the greatest probability as the target predicted sub-quality level; When the probability corresponding to the target prediction sub-quality level is greater than the confidence probability value, the target prediction sub-quality level is determined as a reference quality level label of the second enhanced sample.

7. The video quality determination model training method according to claim 1, characterized in that: The step of obtaining the predicted quality level corresponding to the video sample through the neural network model to be trained comprises: The predicted quality grade score corresponding to the video sample is obtained through the neural network model to be trained; wherein the predicted quality grade score has a score range.

8. The video quality determination model training method according to claim 1, characterized in that: The video sample has an exposure number label, and the training method further includes: The predicted exposure times corresponding to the first enhanced sample are obtained through the neural network model to be trained, and the fourth loss function of the neural network model to be trained is generated according to the exposure time label corresponding to the first enhanced sample and the predicted exposure time corresponding to the first enhanced sample, and the neural network model to be trained is trained according to the fourth loss function and the second loss function.

9. The video quality determination model training method according to claim 2, characterized in that: The video sample has an exposure number label, and the training method further includes: The predicted exposure times corresponding to the video sample are obtained through the neural network model to be trained, a third loss function of the neural network model to be trained is generated according to the exposure time label and the predicted exposure time, and the neural network model to be trained is trained according to the third loss function and the first loss function.

10. The video quality determination model training method according to claim 9, characterized in that: The step of obtaining the predicted exposure times corresponding to the video sample through the neural network model to be trained comprises: The predicted exposure times score corresponding to the video sample is obtained through the neural network model to be trained; wherein the predicted exposure times score has a score range.

11. A video recommendation method, characterized in that: include: Obtain a video to be recommended, and input the video to be recommended into a video quality determination model to obtain a quality level of the video to be recommended; wherein the video quality determination model is obtained by the training method of the video quality determination model according to any one of claims 1 to 10; The video to be recommended is recommended according to the quality level of the video to be recommended.

12. The video recommendation method according to claim 11, characterized in that: The step of recommending the video to be recommended according to the quality level of the video to be recommended comprises: Acquire a plurality of videos to be recommended and quality levels corresponding to the videos to be recommended, and sort the videos to be recommended according to the quality levels corresponding to the plurality of videos to be recommended to obtain a first sorting result; The video to be recommended is recommended according to the first sorting result.

13. The video recommendation method according to claim 12, characterized in that: The videos to be recommended have quality grade scores, and the step of sorting the videos to be recommended according to the quality grades corresponding to the multiple videos to be recommended to obtain a first sorting result comprises: The videos to be recommended are sorted according to the quality level scores corresponding to the multiple videos to be recommended to obtain a first sorting result.

14. The video recommendation method according to claim 11, characterized in that: The step of recommending the video to be recommended according to the quality level of the video to be recommended comprises: Determining the exposure times of the video to be recommended by the video quality determination model; The video to be recommended is recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended.

15. The video recommendation method according to claim 14, characterized in that: The step of recommending the video to be recommended according to the quality level of the video to be recommended and the exposure times of the video to be recommended comprises: Acquire a plurality of the videos to be recommended and the quality levels and exposure times corresponding to the videos to be recommended, and sort the videos to be recommended according to the quality levels and exposure times corresponding to the plurality of videos to be recommended to obtain a second sorting result; The video to be recommended is recommended according to the second sorting result.

16. The video recommendation method according to claim 15, characterized in that: The videos to be recommended have a quality grade score and an exposure count score, and the step of sorting the videos to be recommended according to the quality grades and exposure counts corresponding to the multiple videos to be recommended to obtain a second sorting result comprises: The videos to be recommended are sorted according to the quality level scores and exposure number scores corresponding to the multiple videos to be recommended to obtain a second sorting result.

17. The video recommendation method according to claim 16, characterized in that: The step of sorting the videos to be recommended according to the quality level scores and exposure number scores corresponding to the multiple videos to be recommended to obtain a second sorting result comprises: Get the basic score for video recommendation; Obtain a quality grade score threshold and an adjustment weight corresponding to a quality grade range distinguished by the quality grade score threshold, and obtain an exposure count score threshold and an adjustment weight corresponding to a quality grade range distinguished by the exposure count score threshold; Determining a recommendation score of the video to be recommended according to the video recommendation basic score, the adjustment weight corresponding to the quality grade score threshold, and the adjustment weight corresponding to the exposure count score threshold; The multiple videos to be recommended are sorted according to the recommendation scores to obtain a second sorting result.

18. A training device for a video quality determination model, characterized in that: include: A sample acquisition unit, configured to acquire video samples; An enhanced sample generating unit is configured to generate a corresponding first enhanced sample and a corresponding second enhanced sample according to the video sample when the video sample does not have a quality level label; wherein the degree of change of the first enhanced sample relative to the video sample is less than the degree of change of the second enhanced sample relative to the video sample; A reference quality level label generating unit is configured to obtain a predicted quality level corresponding to the first enhanced sample through a neural network model to be trained, and determine a reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample; A second loss function determining unit is configured to obtain the predicted quality level corresponding to the second enhanced sample through the neural network model to be trained, and determine the second loss function according to the predicted quality level corresponding to the second enhanced sample and the reference quality level label of the second enhanced sample; A neural network model training unit is configured to train the neural network model to be trained according to the second loss function to obtain a video quality determination model; an enhanced sample prediction quality determination unit, configured to obtain the prediction quality level corresponding to the second enhanced sample by using the neural network model to be trained when the probability corresponding to the prediction quality level of the first enhanced sample is less than the confidence probability value; The second loss function determination subunit is configured to determine the second loss function according to the prediction quality level corresponding to the first enhanced sample and the prediction quality level corresponding to the second enhanced sample.

19. The video quality determination model training device according to claim 18, characterized in that: The training device of the video quality determination model also includes: The first loss function determination unit is configured to execute, when the video sample has the quality level label, obtaining the predicted quality level corresponding to the video sample through the neural network model to be trained, generating a first loss function of the neural network model to be trained according to the quality level label and the predicted quality level, and training the neural network model to be trained according to the first loss function.

20. The video quality determination model training device according to claim 19, characterized in that: The predicted quality level corresponding to the video sample is obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: A feature acquisition unit is configured to acquire text features, audio features, and image features corresponding to the video sample through the neural network model to be trained; The quality prediction unit is configured to obtain a predicted quality level corresponding to the video sample according to text features, audio features and image features corresponding to the video sample.

21. The video quality determination model training device according to claim 18, characterized in that: The first enhanced sample and the second enhanced sample are generated according to the video sample, and the training device of the video quality determination model further includes: An image sample acquisition unit, configured to acquire an image sample corresponding to the video sample; The enhanced sample acquisition unit is configured to perform one or more image enhancement processing operations on the image sample to obtain a first enhanced sample and a second enhanced sample corresponding to the video sample.

22. The video quality determination model training device according to claim 18, characterized in that: The predicted quality level corresponding to the first enhanced sample is obtained by the neural network model to be trained, and the reference quality level label of the second enhanced sample is determined according to the predicted quality level corresponding to the first enhanced sample. The training device of the video quality determination model also includes: A prediction quality probability acquisition unit, configured to obtain the prediction quality level corresponding to the first enhanced sample and the probability corresponding to the prediction quality level of the first enhanced sample through the neural network model to be trained; The first reference label determination unit is configured to determine the reference quality level label of the second enhanced sample according to the predicted quality level corresponding to the first enhanced sample when the probability corresponding to the predicted quality level of the first enhanced sample is greater than the confidence probability value.

23. The video quality determination model training device according to claim 22, characterized in that: The predicted quality level corresponding to the first enhanced sample includes at least two predicted sub-quality levels, and when the probability corresponding to the predicted quality level is greater than the confidence probability value, the predicted quality level corresponding to the first enhanced sample is determined as the reference quality level label of the second enhanced sample, and the training device of the video quality determination model further includes: a target sub-quality level determining unit, configured to obtain the at least two predicted sub-quality levels and the probability corresponding to each of the predicted sub-quality levels, and determine the predicted sub-quality level with the largest probability as the target predicted sub-quality level; The second reference label determining unit is configured to determine the target predicted sub-quality level as a reference quality level label of the second enhanced sample when the probability corresponding to the target predicted sub-quality level is greater than the confidence probability value.

24. The video quality determination model training device according to claim 18, characterized in that: The predicted quality corresponding to the video sample is obtained by the neural network model to be trained, and the training device of the video quality determination model also includes: The predicted quality score acquisition unit is configured to execute the predicted quality grade score corresponding to the video sample through the neural network model to be trained; wherein the predicted quality grade score has a score range.

25. The video quality determination model training device according to claim 18, characterized in that: The video sample has an exposure number label, and the training device of the video quality determination model further includes: A quality grade label training unit is provided, and is configured to execute the predicted exposure times corresponding to the first enhanced sample obtained through the neural network model to be trained, generate a fourth loss function of the neural network model to be trained according to the exposure times label corresponding to the first enhanced sample and the predicted exposure times corresponding to the first enhanced sample, and train the neural network model to be trained according to the fourth loss function and the second loss function.

26. The video quality determination model training device according to claim 19, characterized in that: The video sample has an exposure number label, and the training device of the video quality determination model further includes: A training unit without a quality grade label is configured to execute the predicted exposure times corresponding to the video sample obtained by the neural network model to be trained, generate a third loss function of the neural network model to be trained according to the exposure times label and the predicted exposure times, and train the neural network model to be trained according to the third loss function and the first loss function.

27. The video quality determination model training device according to claim 26, characterized in that: The predicted exposure times corresponding to the video sample are obtained by the neural network model to be trained, and the training device of the video quality determination model further includes: The predicted exposure times score acquisition unit is configured to execute the predicted exposure times score corresponding to the video sample through the neural network model to be trained; wherein the predicted exposure times score has a score range.

28. A video recommendation device, characterized in that: include: A video acquisition unit to be recommended, configured to acquire the video to be recommended, and input the video to be recommended into a video quality determination model to obtain a quality level of the video to be recommended; wherein the video quality determination model is obtained by the training method of the video quality determination model according to any one of claims 1 to 10; The video recommendation unit is configured to recommend the video to be recommended according to the quality level of the video to be recommended.

29. The video recommendation device according to claim 28, characterized in that: The video to be recommended is recommended according to the quality level of the video to be recommended, and the video recommendation device further comprises: A quality ranking unit is configured to obtain a plurality of videos to be recommended and quality rankings corresponding to the videos to be recommended, and to sort the videos to be recommended according to the quality rankings corresponding to the plurality of videos to be recommended to obtain a first ranking result; The first sorting unit is configured to recommend the to-be-recommended videos according to the first sorting result.

30. The video recommendation device according to claim 29, characterized in that: The videos to be recommended have quality grade scores, and the videos to be recommended are sorted according to the quality grades corresponding to the multiple videos to be recommended to obtain a first sorting result, and the video recommendation device further includes: The quality grade score sorting unit is configured to sort the videos to be recommended according to the quality grade scores corresponding to the multiple videos to be recommended to obtain a first sorting result.

31. The video recommendation device according to claim 29, characterized in that: The video to be recommended is recommended according to the quality level of the video to be recommended, and the video recommendation device further comprises: An exposure number acquisition unit, configured to determine the exposure number of the video to be recommended by using the video quality determination model; The comprehensive recommendation unit is configured to recommend the video to be recommended according to the quality level of the video to be recommended and the number of exposures of the video to be recommended.

32. The video recommendation device according to claim 31, characterized in that: The video to be recommended is recommended according to the quality level of the video to be recommended and the exposure times of the video to be recommended, and the video recommendation device further comprises: A second ranking result obtaining unit is configured to obtain a plurality of videos to be recommended and quality levels and exposure times corresponding to the videos to be recommended, and to sort the videos to be recommended according to the quality levels and exposure times corresponding to the plurality of videos to be recommended to obtain a second ranking result; The second sorting and recommending unit is configured to recommend the video to be recommended according to the second sorting result.

33. The video recommendation device according to claim 32, characterized in that: The videos to be recommended have a quality grade score and an exposure number score, and the videos to be recommended are sorted according to the quality grades and exposure numbers corresponding to the multiple videos to be recommended to obtain a second sorting result, and the video recommendation device further includes: The second sorting result obtaining subunit is configured to sort the videos to be recommended according to the quality grade scores and exposure number scores corresponding to the multiple videos to be recommended to obtain a second sorting result.

34. The video recommendation device according to claim 33, characterized in that: The step of sorting the videos to be recommended according to the quality grade scores and exposure number scores corresponding to the multiple videos to be recommended to obtain a second sorting result, the video recommendation device further comprises: A video recommendation basic score acquisition unit is configured to execute acquisition of a video recommendation basic score; An adjustment weight acquisition unit is configured to acquire a quality grade score threshold and an adjustment weight corresponding to a quality grade range distinguished by the quality grade score threshold, and acquire an exposure count score threshold and an adjustment weight corresponding to a quality grade range distinguished by the exposure count score threshold; A recommendation score acquisition unit, configured to determine the recommendation score of the video to be recommended according to the video recommendation basic score, the adjustment weight corresponding to the quality level score threshold, and the adjustment weight corresponding to the exposure number score threshold; The recommendation score sorting unit is configured to sort the multiple videos to be recommended according to the recommendation scores to obtain a second sorting result.

35. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the executable instructions to implement the video quality determination model training method or video recommendation method as described in any one of claims 1 to 17.

36. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the video quality determination model training method or video recommendation method as described in any one of claims 1 to 17.

37. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the video quality determination model training method or video recommendation method described in any one of claims 1 to 17 is implemented.

Images