Accurate targeted advertisement delivery method based on movie and television scenes and articles

Abstract

The invention discloses a precise targeted advertisement delivery method based on movie and television scenes and articles, and belongs to the technical field of digital media advertisements. The method comprises the following steps: intercepting a played picture at regular time or as required in a video playing process, carrying out recognition through a machine learning algorithm in an object recognition model, and combining an interest label of a user, thereby matching advertisement materials and displaying the matched advertisement materials through a video playing system. According to the method, a more comprehensive and accurate user portrait is constructed by fusing the interest label, the watching history, the scene article identification result and other multi-modal data of the user. Therefore, the advertisement putting system can better understand and meet the requirements of the user, and the click rate and conversion rate of the advertisement are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital media advertising, and specifically is a method for accurately targeting advertisements based on film and television scenes and objects. Background Art

[0002] In the IPTV and OTT industries, the advertising platform system includes the advertising delivery system and the resource media system, which are responsible for advertising delivery and the management of advertising resources. Advertising delivery is mainly targeted based on the region and product target group labels, which can improve.

[0003] For this kind of scenario, one method is to deliver ads based on the scene to improve the user's viewing experience. The other method is to deliver ads based on scene object recognition to improve conversion and viewing experience, allowing users to be more immersed in the advertising scene based on their interest in the plot.

[0004] Currently in the IPTV and OTT industries, existing technologies for advertising delivery usually push ads based on user regions, user interests, age, browsing history, search keywords, and click behaviors. The process of data collection and analysis in advertising delivery. For example, based on historical browsing data, the platform will record the user's viewing history, including the type, duration, and frequency of programs watched. These data are important bases for analyzing user interest preferences. Secondly, the user's click behavior during video viewing, such as clicking on ads, links, and recommended content, can also provide valuable references for advertising delivery. However, in the scenarios of IPTV and OTT advertising delivery, based on tags or user behavior, they are all operations from family members, and the age and interest distribution of family members are different, with large differences, and the resulting advertising promotion will also be inaccurate. Summary of the invention

[0005] The purpose of the present invention is to address the problem of inaccurate advertising push based on tags or user behavior in the current existing advertising push scenarios, and propose a method for accurate and targeted advertising delivery based on film and television scenes and objects. The present invention captures the playback screen regularly or on demand during video playback, and identifies it through the machine learning algorithm in the object recognition model, while combining the user's interest tags, so as to match the advertising materials for display through the video playback system. The present invention constructs a more comprehensive and accurate user portrait by integrating multimodal data such as user interest tags, viewing history, scene object recognition results, etc. This enables the advertising delivery system to better understand and meet user needs and improve the click-through rate and conversion rate of advertisements.

[0006] In order to solve the above technical problems, the present invention provides a method for accurate and targeted advertising based on film and television scenes and objects, comprising the following steps:

[0007] S1. The user initiates a video request. The video playback system receives the user's request and starts playing the video.

[0008] S2. During video playback, capture the currently playing image regularly or on demand and send it to the object recognition model;

[0009] S3, the object recognition model receives the image data, identifies the objects and their categories in the picture through a machine learning algorithm, and obtains a recognition result;

[0010] S4. The advertising delivery system receives the recognition result and searches for relevant advertising materials from the advertising library in combination with the user's interest tags;

[0011] S5. The advertising delivery system conducts bidding and displays the selected advertising material through the video playback system;

[0012] S6. Behavior data and feedback data are collected through the video playback system, and the data management platform generates new user interest tags and behavior models based on the behavior data to optimize the delivery strategy.

[0013] Preferably, in S2, the object recognition model comprises the following steps:

[0014] S21, collect film and television scene pictures as training sets, and mark the categories of objects in the scenes in the pictures;

[0015] S22, preprocessing the film and television scene images, including resizing, color space conversion, and noise removal;

[0016] S23. Select a machine learning algorithm as the basic model, use the labeled training set to train the basic model, and continuously adjust the network weights through the back propagation algorithm to obtain the object recognition model.

[0017] Preferably, in S4, the interest tags include classification tags pre-stored by the user in the system and historical classification tags recorded by the system during the historical identification process.

[0018] Preferably, the interest tags also include classification tags selected by the user when using the system.

[0019] Preferably, the bidding process of the advertisement delivery system comprises the following steps:

[0020] Sa. The advertising delivery system sends the advertising request to the demand-side platform, which participates in the bidding for advertising space based on the advertiser's budget and bidding strategy;

[0021] Sb, the advertising exchange platform receives bidding requests from multiple demand-side platforms and determines the ownership of advertising space through a real-time bidding mechanism;

[0022] Sc. The advertising exchange platform feeds the bidding results back to the advertising delivery system, which selects the advertising material with the highest bid;

[0023] S4. The advertisement delivery system sends the selected advertisement material to the video playback system to display the corresponding advertisement content.

[0024] Preferably, the determining the ownership of the advertising space by a real-time bidding mechanism includes the real-time bidding mechanism taking into account bids, budgets, target audiences, and historical performances of advertisers.

[0025] As a priority, dynamically adjust the bid amount based on the quality of the ad space, the value of the user and the advertiser's budget; analyze historical bidding and ad display data to optimize bidding strategies; consider the ad's click-through rate, conversion rate, and user retention rate, and balance the advertiser's short-term benefits and long-term effects through a multi-objective optimization algorithm; use machine learning models to predict ad performance and guide bidding strategies.

[0026] Preferably, the behavior data includes users clicking on ads, skipping ads, or continuing to watch videos after watching the ads; the feedback data includes click-through rate and viewing time, and the feedback data is transmitted to the advertising delivery system through an API interface.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This solution introduces scene and object recognition technology, and uses advanced image recognition algorithms such as convolutional neural networks (CNN) to identify scenes and objects in real time when users are watching videos. This enables the advertising delivery system to push ads more accurately based on the user's immediate interests and current viewing content; at the same time, this solution can not only span different media and terminals, but also optimize advertising delivery strategies through real-time data analysis to improve advertising interactivity and conversion rates. Such a system will greatly improve the efficiency of advertising delivery, better serve the needs of advertisers, and provide users with a more relevant and attractive advertising experience.

[0029] 2. This solution introduces the real-time bidding (RTB) mechanism and dynamic bidding strategy. The system dynamically adjusts the bid amount based on the quality of the ad space, the value of the user and the advertiser's budget to ensure that advertisers can obtain high-quality ad display opportunities at the right time.

[0030] 3. This solution collects and analyzes user feedback data in real time through the data management platform (DMP) to generate new user interest tags and behavior models. These optimized data are immediately fed back to the advertising delivery system to achieve continuous optimization of advertising delivery strategies and improve advertising conversion rates and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic diagram of object recognition performed by the object recognition model of the present invention;

[0032] Figure 2 The scene of advertising for the identified scene objects of the present invention Figure 1 ;

[0033] Figure 3 The scene of advertising for the identified scene objects of the present invention Figure 2 . DETAILED DESCRIPTION

[0034] Example 1: Figure 1 - Figure 3 As shown, a method for accurate and targeted advertising based on film and television scenes and objects includes:

[0035] S1. The user initiates a video request, and the video playback system receives the user's request and starts playing the video.

[0036] S2. During video playback, the currently playing image is captured periodically or as needed, and sent to the object recognition model.

[0037] Among them, in S2, the object recognition model includes the following steps:

[0038] S21, collect film and television scene pictures as training sets, and mark the categories of objects in the scenes in the pictures;

[0039] S22, preprocessing the film and television scene images, including resizing, color space conversion, and noise removal;

[0040] S23. Select a machine learning algorithm as the basic model, use the labeled training set to train the basic model, and continuously adjust the network weights through the back propagation algorithm to obtain the object recognition model.

[0041] S3. The object recognition model receives image data, identifies objects and their categories in the image through a machine learning algorithm, and obtains a recognition result.

[0042] Specifically, the process of object recognition based on film and television scenes is as follows:

[0043] 1. Data preparation: First, a large number of film and television scene images need to be collected as training sets, and these images need to be labeled with the object categories in the scene. For example, if an advertiser wants to target a car brand for advertising, it needs to prepare a large number of film and television scene images containing different brands of cars and accurately mark the brand and model of the car.

[0044] 2. Preprocessing: Preprocess the image data, including resizing, color space conversion, noise removal, etc., to ensure the data quality of the neural network input.

[0045] 3. Model selection: You can choose convolutional neural network (CNN) as the basic model because CNN performs well in image recognition tasks. Commonly used CNN architectures include ResNet, Inception, VGG, etc.

[0046] 4. Training process: Use the labeled data set to train the model, and continuously adjust the network weights through the back-propagation algorithm so that the model can accurately predict the objects in the image. Appropriate hyperparameters need to be set during the training process, such as learning rate, batch size, number of iterations, etc.

[0047] 5. Evaluation and optimization: After training is completed, use the validation set to evaluate the performance of the model, and adjust the model structure or hyperparameters based on the evaluation results until satisfactory recognition accuracy is achieved.

[0048] For example, suppose we want to train a model to recognize coffee cups. First, we need to prepare thousands of pictures containing coffee cups, and each picture needs to be marked with the location and category of the coffee cup. Then, use this data to train a CNN model based on ResNet50. During the training process, the generalization ability of the model can be improved by adjusting the learning rate, increasing data enhancement, etc. Finally, use the validation set to evaluate the recognition accuracy of the model. If the recognition rate is lower than expected, you can try to use a deeper network structure or a larger data set for training.

[0049] Furthermore, the specific contents of the algorithm and code for optimizing the image recognition model (object recognition model) are supplemented:

[0050] Algorithm optimization:

[0051] 1. Data enhancement: By rotating, translating, scaling, and cropping the original images, the diversity of training data is increased, which helps improve the generalization ability of the model.

[0052] 2. Regularization: Use L1 / L2 regularization, Dropout and other techniques to prevent model overfitting.

[0053] 3. Transfer learning: Using pre-trained models (such as those on ImageNet) and fine-tuning them on new tasks can significantly reduce training time and improve model performance.

[0054] 4. Model fusion: By integrating multiple models (such as voting, weighted average, etc.), the recognition accuracy can be further improved.

[0055] It should be noted that the following code can be used as an example:

[0056] import tensorflow as tf

[0057] from tensorflow.keras.applications import ResNet50

[0058] from tensorflow.keras.layers import Dense,GlobalAveragePooling2D

[0059] from tensorflow.keras.models import Model

[0060] # Load the pre-trained model

[0061] base_model=ResNet50(weights='imagenet',include_top=False,input_shape=(224,224,3))

[0062] #Add custom layer

[0063] x = base_model.output

[0064] x=GlobalAveragePooling2D()(x)

[0065] x=Dense(1024,activation='relu')(x)

[0066] predictions=Dense(num_classes,activation='softmax')(x)

[0067] #Build a new model

[0068] model=Model(inputs=base_model.input,outputs=predictions)

[0069] #Freeze the layers of the pre-trained model

[0070] for layer in base_model.layers:

[0071] layer.trainable = False

[0072] # Compile the model

[0073] model.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['accuracy'])

[0074] #Data Augmentation

[0075] data_gen=tf.keras.preprocessing.image.ImageDataGenerator(

[0076] rotation_range=20,

[0077] width_shift_range=0.2,

[0078] height_shift_range=0.2,

[0079] horizontal_flip=True,

[0080] vertical_flip=True )

[0082] #Train model

[0083] train_generator=data_gen.flow_from_directory(train_dir,target_size=(224,224),batch_size=32)

[0084] model.fit(train_generator,epochs=10)

[0085] In the prior art, traditional advertising delivery technology mainly relies on basic information such as the user's region, age, interest tags, as well as the user's browsing history and search keywords. Although these data can provide a certain user portrait, they are limited in real time and accuracy. This application introduces scene and object recognition technology, which uses advanced image recognition algorithms such as convolutional neural networks (CNN) to identify scenes and objects when users watch videos in real time. This enables the advertising delivery system to push advertisements more accurately based on the user's immediate interests and current viewing content.

[0086] S4. The advertisement delivery system receives the recognition result and searches for relevant advertisement materials from the advertisement library in combination with the user's interest tags.

[0087] Specifically, in S4, the interest tags include the classification tags pre-stored by the user in the system and the historical classification tags recorded by the system during the historical recognition process. The interest tags also include the classification tags selected by the user when using the system.

[0088] S5. The advertising delivery system conducts bidding and displays the selected advertising material through the video playback system.

[0089] Specifically, Figure 1 As shown, the interaction process between the S2-S5 object recognition model and the advertising delivery system is as follows:

[0090] 1. Real-time recognition: When users watch film and television content, the system captures the currently playing screen in real time and sends it to the object recognition model for recognition.

[0091] 2. Data transmission: The recognition results (i.e. the objects in the scene and their categories) are transmitted to the advertising delivery system through the API interface.

[0092] 3. Advertisement matching: The ad delivery system selects appropriate ads from the ad library based on the recognition results and the user’s interest tags.

[0093] 4. Feedback mechanism: After the advertisement is released, the system will collect user feedback data (such as click-through rate, viewing time, etc.) to optimize future advertising strategies.

[0094] For example: Suppose a user is watching a movie and a high-end watch appears on the screen. The object recognition model identifies the watch and sends the result (brand and model of the watch) to the ad delivery system. The ad delivery system selects an ad related to the watch for delivery based on the user's interest tag (such as liking high-end luxury goods) and the recognition result. The user sees the ad and clicks it. The system records this click and uses it as a reference for future ad delivery.

[0095] S6. Behavior data and feedback data are collected through the video playback system, and the data management platform generates new user interest tags and behavior models based on the behavior data to optimize the delivery strategy.

[0096] Specifically, the bidding process of the advertising delivery system includes the following steps:

[0097] Sa. The advertising delivery system sends the advertising request to the demand-side platform, which participates in the bidding for advertising space based on the advertiser's budget and bidding strategy;

[0098] Sb, the advertising exchange platform receives bidding requests from multiple demand-side platforms and determines the ownership of advertising space through a real-time bidding mechanism;

[0099] Sc. The advertising exchange platform feeds the bidding results back to the advertising delivery system, which selects the advertising material with the highest bid;

[0100] S4. The advertisement delivery system sends the selected advertisement material to the video playback system to display the corresponding advertisement content.

[0101] Among them, the real-time bidding mechanism determines the ownership of the advertising space, including the real-time bidding mechanism considering the bid, budget, target audience, and historical performance of the advertiser. The bid amount is adjusted dynamically according to the quality of the advertising space, the value of the user, and the budget of the advertiser; the historical bidding and ad display data are analyzed to optimize the bidding strategy; the click-through rate, conversion rate, and user retention rate of the advertisement are considered, and the advertiser's short-term benefits and long-term effects are balanced through a multi-objective optimization algorithm; the machine learning model is used to predict the performance of the advertisement and guide the bidding strategy.

[0102] Specifically, the behavioral data includes whether the user clicks on the ad, skips the ad, or continues to watch the video after watching the ad; the feedback data includes click-through rate and viewing time, and the feedback data is transmitted to the advertising delivery system through the API interface.

[0103] In the precise and targeted advertising delivery system based on film and television scenes and objects, bidding strategy is the key link to ensure efficient allocation of advertising resources. The following is a detailed description of bidding strategy, including the basic principles, specific steps, optimization methods and examples of bidding.

[0104] Bidding basic principles:

[0105] Real-Time Bidding (RTB): RTB is a real-time online advertising purchase method. Advertisers decide whether to purchase ad display opportunities within a few milliseconds through bidding. When an ad space appears, the ad exchange (ADX) will send a bidding request to multiple demand-side platforms (DSPs). DSPs bid according to the needs and budgets of their own advertisers. The highest bidder gets the opportunity to display the ad.

[0106] Bidding logic: Bidding logic is usually based on factors such as the advertiser's bid, budget, target audience, historical performance, etc. The bidding algorithm will take these factors into consideration and calculate the optimal bid amount.

[0107] The specific steps of bidding are as follows:

[0108] 1. Advertisement request: The video playback system captures the current screen, identifies the objects in the scene through the object recognition model, and sends the recognition results to the advertisement delivery system. The advertisement delivery system generates an advertisement request based on the recognition results and user interest tags and sends it to ADX.

[0109] 2. Bidding request: After receiving the ad request, ADX sends a bidding request to multiple DSPs, including information about the ad space (such as location, size, user characteristics, etc.).

[0110] 3. DSP bidding: Each DSP calculates a bidding amount based on the advertiser's budget, target audience, historical performance, etc. DSP returns the bidding amount and ad creative information to ADX.

[0111] 4. Bid evaluation: ADX receives bid responses from all DSPs and evaluates the pros and cons of each bid based on factors such as bid amount, ad relevance, and historical performance. The evaluation process may also take into account the advertiser's credibility and historical cooperation.

[0112] 5. Bidding results: ADX selects the highest bidder based on the evaluation results and allocates the ad space to the DSP.

[0113] The bidding results are returned to the advertising delivery system through the API interface.

[0114] 6. Advertisement display: The advertisement delivery system sends the selected advertisement material to the video playback system based on the bidding results, and the video playback system displays the advertisement.

[0115] The following are the steps for bidding optimization:

[0116] 1. Dynamic bidding: Dynamically adjust the bid amount based on the quality of the ad space, the value of the user, and the advertiser's budget. For example, for high-quality ad space and high-value users, the bid can be appropriately increased to increase the probability of winning the auction.

[0117] 2. Historical performance analysis: Analyze historical bidding and ad display data to optimize bidding strategies. For example, if an ad has a high click-through rate during a specific time period, you can increase the bid during that time period.

[0118] 3. Multi-objective optimization: not only consider the advertiser's budget and bid, but also consider multiple objectives such as the click-through rate, conversion rate, user retention rate, etc. Through the multi-objective optimization algorithm, balance the advertiser's short-term benefits and long-term effects.

[0119] 4. Machine learning model: Use machine learning models to predict ad performance and guide bidding strategies. For example, use regression models to predict ad click-through rates and use classification models to predict user conversion rates.

[0120] An example of a supplementary bidding strategy is as follows: Assume that there is an advertiser A who wants to bid for a high-quality ad space to display an ad for a high-end watch. The following are the specific bidding strategy steps:

[0121] 1. Ad request: The video playback system captures the current screen and identifies a high-end watch in the screen. The ad delivery system generates an ad request based on the recognition result and the user's interest tag (such as liking high-end luxury goods) and sends it to ADX.

[0122] 2. Bidding request: After receiving the ad request, ADX sends a bidding request to multiple DSPs, including information about the ad space (such as location, size, user characteristics, etc.).

[0123] 3. DSP bidding: DSP A calculates a bidding amount, such as RMB 10, based on advertiser A’s budget, target audience (users who like high-end luxury goods), and historical performance (high ad click-through rate). DSP B calculates a bidding amount, such as RMB 8, based on advertiser B’s budget, target audience (users who like fashion products), and historical performance (average ad click-through rate).

[0124] 4. Bid evaluation: ADX receives bid responses from all DSPs and evaluates the pros and cons of each bid. The evaluation results show that DSP A has the highest bid amount and has good ad relevance and historical performance.

[0125] 5. Bidding results: ADX selects DSP A as the highest bidder and allocates the ad space to DSP A. The bidding results are returned to the advertising delivery system through the API interface.

[0126] 6. Advertisement display: Based on the bidding results, the advertisement delivery system sends the high-end watch advertising material provided by DSP A to the video playback system, and the video playback system displays the advertisement.

[0127] This solution introduces a real-time bidding (RTB) mechanism and a dynamic bidding strategy. The system dynamically adjusts the bid amount based on the quality of the ad space, the value of the user, and the advertiser's budget to ensure that advertisers can obtain high-quality ad display opportunities at the right time. This solution also collects and analyzes user feedback data in real time through a data management platform (DMP) to generate new user interest tags and behavior models. These optimized data are immediately fed back to the advertising delivery system to achieve continuous optimization of the advertising delivery strategy and improve advertising conversion rates and user experience.

[0128] Example 2: Figure 1 - Figure 3 As shown, a method for accurate and targeted advertising based on film and television scenes and objects includes the following steps:

[0129] Step 1: The user watches a video through an IPTV or OTT device. The user initiates a video request. The video playback system receives the user's request and starts playing the video.

[0130] Step 2: During video playback, the system captures the currently playing image periodically or on demand and sends it to the object recognition model.

[0131] Step 3: The object recognition model receives the image data and identifies the objects and their categories in the picture through technologies such as convolutional neural networks (CNN). For example, it identifies a high-end watch in the picture. The recognition results (such as the brand and model of the watch) are encapsulated into JSON format data and passed to the advertising delivery system through the API interface.

[0132] Step 4: The advertising delivery system receives the recognition results from the object recognition model, including the objects in the scene and their categories. The advertising delivery system also obtains the user's interest tags (such as liking high-end luxury goods).

[0133] Step 5: The ad delivery system searches for relevant ad materials from the ad library based on the identification results and user interest tags. For example, it finds ads related to high-end watches.

[0134] Step 6: DSP (Demand-Side Platform): The advertising delivery system sends the advertising request to DSP, and DSP participates in the bidding for the advertising space based on the advertiser's budget and bidding strategy.

[0135] ADX (Advertising Exchange): ADX receives bidding requests from multiple DSPs and decides which advertiser wins the ad space through a real-time bidding (RTB) mechanism.

[0136] Step 7: ADX feeds the bidding results back to the advertising delivery system, which selects the advertising material of the advertiser with the highest bid. The advertising delivery system sends the selected advertising material to the video playback system.

[0137] Step 8: The video playback system receives the advertising material, inserts the advertisement into the video during playback, and displays it to the user.

[0138] Step 9: After watching the ad, the user may click on the ad, skip the ad, or continue watching the video. These behavior data are collected by the video playback system. Feedback data (such as click-through rate, viewing time, etc.) is passed to the advertising delivery system through the API interface.

[0139] Step 10: DMP (Data Management Platform): The advertising delivery system sends feedback data to DMP, which analyzes the data and generates new user interest tags and behavior models.

[0140] Optimization strategy: DMP feeds the optimized data back to the advertising delivery system, which adjusts the advertising delivery strategy based on the new data to improve the conversion rate of advertising and user experience.

[0141] like Figure 2 and Figure 3 As shown, advertisers need to register on the DSP platform and undergo qualification review, then create an advertising plan, set a budget and bidding strategy, and submit advertising materials. DSP converts this information into advertising orders, and through interaction with DMP, uses the collected user data to accurately locate and optimize advertising delivery. At the same time, SSP, as a representative of media resources, manages the inventory of advertising space and provides real-time advertising display opportunities to DSP through advertising trading platforms (such as ADX).

[0142] The specific workflow of DSP and SSP interaction is as follows:

[0143] First, when there is a new advertising demand, advertisers will connect with the DSP platform through their agents or directly, and then produce and prepare advertising materials. Next, DSP will formulate a delivery strategy based on the advertiser's budget, target audience and other settings. After that, DSP will deliver the advertisements to multiple advertising exchange platforms (Ad Exchange), such as Google AdWords, Baidu Promotion, etc. In this process, DSP will bid in real time to purchase suitable advertising display opportunities to ensure that the advertisements can reach potential target customers. Once the advertisement is displayed, SSP (Sell-Side Platform) will intervene, which manages their inventory on behalf of publishers and trades with DSP. SSP is responsible for processing the data provided by the data management platform (DMP) to optimize the advertising delivery effect. Finally, the advertising data will be collected and sent to the data analysis system for analysis and evaluation of the advertising effect. Throughout the process, programmatic advertising technology realizes automated purchase, optimization and reporting, improving advertising efficiency and accuracy.

[0144] In the real-time bidding (RTB) process, DSP evaluates and bids for each display opportunity based on the parameters set by the advertiser and the data provided by the DMP. The winning bid will trigger the display of the ad, and this process is completed in a very short time. After the ad is displayed, the SSP is responsible for collecting the display and click data of the ad, which will be sent to the data analysis system for in-depth analysis to evaluate the advertising effect and generate feedback reports.

[0145] In addition, the financial management system tracks the consumption of advertising expenses to ensure that the financial settlement of advertisers and media resources is accurate. Throughout the process, the continuous management and update of user portrait tags is an indispensable part of improving the relevance of advertising. The logical relationship between these subsystems ensures the efficiency and transparency of the entire programmatic advertising purchase process, from the creation of advertising plans to the evaluation of delivery effects.

[0146] Embodiment 3: The following example illustrates advertisement delivery.

[0147] Case 1: High-end watch advertising

[0148] Traditional advertising method: Advertisements are placed based on basic information such as the user's region, age, interest tags, etc. Assuming that a user often watches high-end fashion channels, the system will infer that the user may be interested in high-end watches based on the user's viewing history and interest tags, and push relevant ads.

[0149] This application method: When a user watches a video, the system uses object recognition technology to identify the high-end watches that appear in the picture. Combined with the user's interest tags (such as liking high-end luxury goods), the system will push advertisements related to high-end watches more accurately.

[0150] Effect comparison: The traditional method may result in low relevance and attractiveness of advertisements due to inaccurate or single user interest tags. However, the present application method can more accurately match users' immediate interests through scene and object recognition, thereby improving the click-through rate and conversion rate of advertisements.

[0151] Case 2: Car Advertising

[0152] Traditional advertising method: Based on the user’s browsing history and search keywords, it is inferred that the user may be interested in a certain car, and relevant advertisements are pushed.

[0153] The method of this application is: when a user is watching a movie or TV series, the system uses object recognition technology to identify a certain car that appears in the picture. Combined with the user's interest tags (such as liking luxury cars), the system will push advertisements related to the car more accurately.

[0154] Effect comparison: The traditional method may lead to insufficient timeliness and relevance of advertisements due to the lag of users' browsing history and search keywords. However, the present application method can match users' immediate interests more timely through real-time scene and object recognition, thereby improving the exposure and conversion rate of advertisements.

[0155] Second: From the perspective of advertising efficiency, the formula calculation method is as follows:

[0156] Traditional advertising method: Assume that the average click-through rate of traditional advertising method is 1% and the number of clicks per 1000 impressions (CPM) is 10 times.

[0157] This application method: Assume that the average click-through rate of this application method is 5%, and the number of clicks per 1000 impressions (CPM) is 50 times.

[0158] Efficiency improvement calculation:

[0159] Efficiency improvement = (click rate of this application method / click rate of traditional method) - 1

[0160] Efficiency improvement = (5% / 1%)-1 = 400%

[0161] That is, the delivery efficiency of this application method is 5 times that of the traditional method, and the efficiency is improved by 400%.

[0162] Third: From the perspective of advertising conversion rate, the actual advertising cases are as follows:

[0163] Case 3: Tourism advertising

[0164] Traditional advertising method: Based on the user's browsing history and search keywords, it is inferred that the user may be interested in a certain travel destination, and relevant advertisements are pushed.

[0165] The method of this application is: when a user watches a travel documentary, the system uses object recognition technology to identify the tourist attractions that appear in the picture. Combined with the user's interest tags (such as liking outdoor activities), the system will push advertisements related to the tourist attractions more accurately.

[0166] Effect comparison: The traditional method may lead to a low conversion rate of advertisements due to the limitations of users' browsing history and search keywords. However, the present application method can more accurately match users' immediate interests and improve the conversion rate of advertisements through scene and object recognition.

[0167] Formula calculation method

[0168] Traditional advertising method: Assume that the average conversion rate of traditional advertising method is 0.5%, and the number of conversions per 1000 impressions (CPM) is 5 times.

[0169] This application method: Assume that the average conversion rate of this application method is 2%, and the number of conversions per 1000 impressions (CPM) is 20 times.

[0170] Conversion rate improvement calculation:

[0171] Conversion rate improvement = (conversion rate of this application method / conversion rate of traditional method) - 1

[0172] Conversion rate increase = (2% / 0.5%)-1 = 300%

[0173] That is, the conversion rate of this application method is 4 times that of the traditional method, and the conversion rate has increased by 300%.

[0174] Compared with existing advertising delivery technologies, this application has the following key innovations and inventions:

[0175] The innovative points of this application are explained from the following aspects:

[0176] 1. Application of scene and object recognition technology:

[0177] Existing technology: Traditional advertising technology mainly relies on basic information such as user's region, age, interest tags, as well as user's browsing history and search keywords. Although these data can provide a certain user portrait, they are limited in real-time and accuracy.

[0178] This application: This application introduces scene and object recognition technology, which uses advanced image recognition algorithms such as convolutional neural networks (CNN) to identify scenes and objects in real time when users are watching videos. This enables the advertising delivery system to push ads more accurately based on the user's immediate interests and current viewing content.

[0179] 2. Combination of user interest tags and scene object recognition:

[0180] Existing technology: Existing advertising delivery systems can usually only push advertisements based on users' historical behavior data, and cannot fully consider users' immediate interests in the current viewing scenario.

[0181] This application: This application implements a multi-dimensional advertising strategy by combining the user's interest tags and scene object recognition results. The system not only considers the user's long-term interests, but also combines the items in the current viewing content to further improve the relevance and attractiveness of the advertisement.

[0182] 3. Real-time bidding and dynamic bidding strategies:

[0183] Existing technology: Traditional advertising bidding systems are usually based on fixed bidding strategies and lack dynamic adjustments to ad space quality and user value.

[0184] This application: This application introduces a real-time bidding (RTB) mechanism and a dynamic bidding strategy. The system dynamically adjusts the bid amount based on the quality of the ad space, the value of the user, and the advertiser's budget to ensure that advertisers can obtain high-quality ad display opportunities at the right time.

[0185] 4. Closed-loop optimization of user feedback data:

[0186] Existing technology: After collecting user feedback data, existing advertising delivery systems usually take a long time to perform data analysis and adjust strategies, making it difficult to achieve real-time optimization.

[0187] This application: This application collects and analyzes user feedback data in real time through a data management platform (DMP) to generate new user interest tags and behavior models. These optimized data are immediately fed back to the advertising delivery system to achieve continuous optimization of advertising delivery strategies and improve advertising conversion rates and user experience.

[0188] Secondly, the invention of this application can also be explained from the following aspects:

[0189] 1. Advertising strategy based on multimodal data fusion:

[0190] Existing technology: Existing advertising delivery systems mainly rely on a single data source (such as user behavior data) and lack the integration of multimodal data.

[0191] This application: This application builds a more comprehensive and accurate user portrait by integrating multimodal data such as user interest tags, viewing history, scene object recognition results, etc. This enables the advertising delivery system to better understand and meet user needs and improve the click-through rate and conversion rate of advertisements.

[0192] 2. Advertising creative matching based on scene items:

[0193] Existing technology: Existing advertising delivery systems mainly rely on advertisers' preset rules and historical performance data when selecting advertising materials, and lack real-time matching of current viewing scenarios.

[0194] This application: This application uses scene object recognition technology to match advertising materials related to the current viewing content in real time. The system will give priority to advertisements that match the objects recognized in the scene, improving the relevance of advertisements and user acceptance.

[0195] 3. Combination of real-time bidding and dynamic bidding:

[0196] Existing technology: Existing advertising bidding systems usually adopt fixed bidding strategies and lack dynamic adjustments to ad space quality and user value.

[0197] This application: This application uses the real-time bidding (RTB) mechanism and dynamic bidding strategy to ensure that advertisers can obtain high-quality advertising display opportunities at the right time. The system will dynamically adjust the bid amount based on the quality of the ad space, the value of the user and the advertiser's budget to increase the exposure and conversion rate of the ad.

[0198] The object recognition technology in film and television scenes is developed through the collection of a large amount of data features and model training. The recognition rate depends on the size of the sample, so traditional regional tags, user viewing records, and user interest tags can also be used as the basis for our advertising. This technology is not only suitable for advertising, but also has a wide range of applications in advertising production, target positioning retrieval, face recognition, etc.

Claims

1. A method for accurate and targeted advertising based on film and television scenes and objects, characterized in that: The steps include: S1. The user initiates a video request. The video playback system receives the user's request and starts playing the video. S2. During video playback, capture the currently playing image regularly or on demand and send it to the object recognition model; S3, the object recognition model receives the image data, identifies the objects and their categories in the picture through a machine learning algorithm, and obtains a recognition result; S4. The advertising delivery system receives the recognition result and searches for relevant advertising materials from the advertising library in combination with the user's interest tags; S5. The advertising delivery system conducts bidding and displays the selected advertising material through the video playback system; S6. Behavior data and feedback data are collected through the video playback system, and the data management platform generates new user interest tags and behavior models based on the behavior data to optimize the delivery strategy.

2. A method for accurate and targeted advertising based on film and television scenes and objects according to claim 1, characterized in that: In S2, the object recognition model includes the following steps: S21, collect film and television scene pictures as training sets, and mark the categories of objects in the scenes in the pictures; S22, preprocessing the film and television scene images, including resizing, color space conversion, and noise removal; S23. Select a machine learning algorithm as the basic model, use the labeled training set to train the basic model, and continuously adjust the network weights through the back propagation algorithm to obtain the object recognition model.

3. The method for accurate and targeted advertising based on film and television scenes and objects according to claim 1, characterized in that: In S4, the interest tags include classification tags pre-stored by the user in the system and historical classification tags recorded by the system during the historical identification process.

4. The method for accurate and targeted advertising based on film and television scenes and objects according to claim 3, characterized in that: The interest tags also include classification tags selected by the user when using the system.

5. The method for accurate and targeted advertising based on film and television scenes and objects according to claim 1, characterized in that: The bidding process of the advertisement delivery system includes the following steps: Sa. The advertising delivery system sends the advertising request to the demand-side platform, which participates in the bidding for advertising space based on the advertiser's budget and bidding strategy; Sb, the advertising exchange platform receives bidding requests from multiple demand-side platforms and determines the ownership of advertising space through a real-time bidding mechanism; Sc. The advertising exchange platform feeds the bidding results back to the advertising delivery system, which selects the advertising material with the highest bid; S4. The advertisement delivery system sends the selected advertisement material to the video playback system to display the corresponding advertisement content.

6. A method for accurate and targeted advertising based on film and television scenes and objects according to claim 5, characterized in that: The determination of the ownership of the advertising space through the real-time bidding mechanism includes the real-time bidding mechanism taking into account the bid, budget, target audience, and historical performance of the advertiser.

7. A method for accurate and targeted advertising based on film and television scenes and objects according to claim 6, characterized in that: Dynamically adjust bid amounts based on the quality of the ad space, user value, and advertiser budget; analyze historical bidding and ad display data to optimize bidding strategies; consider ad click-through rates, conversion rates, and user retention rates, and balance advertisers’ short-term revenue and long-term results through multi-objective optimization algorithms; use machine learning models to predict ad performance and guide bidding strategies.

8. The method for accurate and targeted advertising based on film and television scenes and objects according to claim 1, characterized in that: The behavior data includes whether the user clicks on the advertisement, skips the advertisement, or continues to watch the video after watching the advertisement; the feedback data includes click-through rate and viewing time, and the feedback data is transmitted to the advertising delivery system through the API interface.

Images