Advertising image classification method and system based on deep convolutional neural network model

Abstract

The present invention discloses an advertising image classification method and system based on a deep convolutional neural network model, including: first classification of the collected target advertising images and related data, and first preprocessing for the first classification result; second classification of the first preprocessing result, and judging the number of model establishment according to the second classification result; according to the judged number of model establishment, a corresponding model is established in combination with a deep convolutional neural network, and the output of the model is the first customer attraction; combining a preset improved clustering algorithm, and the model output obtained by different second classification strategies, to obtain a third classification result. Through the above steps, the present invention can effectively classify advertising images, thereby providing more accurate decision support for advertising placement.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and in particular to an advertising image classification method and system based on a deep convolutional neural network model. Background Art

[0002] In the current digital advertising market, image advertising is an important means of attracting customer attention, and its effectiveness directly affects the conversion rate and revenue of advertising. However, due to the diversity and complexity of advertising images, how to efficiently and accurately classify advertising images in order to better meet the needs of different customer groups has always been a technical challenge faced by the advertising industry.

[0003] Traditional image classification methods rely on manual feature extraction and simple machine learning algorithms, which often fail to achieve satisfactory classification results when processing large-scale, high-dimensional advertising image data. With the development of deep learning technology, especially the breakthrough progress of deep convolutional neural networks (CNN) in the field of image recognition, new solutions have been provided for advertising image classification. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides an advertising image classification method and system based on a deep convolutional neural network model, which can solve the problems mentioned in the background technology.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, the present invention provides an advertising image classification method based on a deep convolutional neural network model, comprising:

[0009] Performing a first classification on the collected target advertisement images and related data, and performing a first preprocessing on the first classification results;

[0010] Performing a second classification on the first preprocessing result, and determining the number of model establishments according to the second classification result;

[0011] According to the determined number of model establishment, a corresponding model is established in combination with a deep convolutional neural network, and the output of the model is the first customer attraction;

[0012] The third classification result is obtained by combining the preset improved clustering algorithm and the model output obtained by different second classification strategies.

[0013] As a preferred solution of the advertising image classification method based on the deep convolutional neural network model described in the present invention, the first classification of the collected target advertising images and related data and the first preprocessing of the first classification results include:

[0014] The target advertising images include page advertising images, pop-up advertising images and inserted advertising images on mobile terminals and PC terminals;

[0015] The relevant data includes the length of time that mobile and / or PC clients stay on page advertisement images, the number of clicks and length of time that clients enter pop-up advertisement images, and the number of clicks on inserted advertisement images;

[0016] The first classification includes classifying the collected target advertising images into pure image-type advertising images and image-and-text-combined advertising images;

[0017] The first preprocessing includes performing a first color preprocessing on a pure image type advertisement image, and performing a first text preprocessing after performing a first color preprocessing on an image and text combination advertisement image.

[0018] As a preferred solution of the advertising image classification method based on the deep convolutional neural network model described in the present invention, the first preprocessing further includes:

[0019] The first color preprocessing is used to perform color analysis and layout analysis on pure image type advertisement images or / and image and text combined advertisement images;

[0020] The first text preprocessing is used to perform a relevance level analysis on the text in the image and text combined advertising image, and the relevance level analysis includes the number of judgments from "the text in the image and text combined advertising image" to the meaning of the target advertising image itself, and the number of judgments includes the number of manual judgments and / or the number of judgments in a preset judgment database.

[0021] As a preferred solution of the advertising image classification method based on the deep convolutional neural network model described in the present invention, wherein: performing a second classification on the first preprocessing result and determining the number of model establishments according to the second classification result includes:

[0022] The second classification includes performing a second classification on the first preprocessing result by using a preset second classification strategy;

[0023] The preset second classification strategy includes a second classification strategy A, a second classification strategy B and a second classification strategy C;

[0024] When the second classification strategy A is selected, the number of models established is twelve according to the second classification results;

[0025] When the second classification strategy B is selected, the number of models established is six according to the second classification results;

[0026] When the second classification strategy C is selected, the number of models to be established is determined to be three according to the second classification result.

[0027] As a preferred solution of the advertising image classification method based on the deep convolutional neural network model described in the present invention, wherein: the establishment of the corresponding model based on the determined number of model establishments and the deep convolutional neural network includes:

[0028] When the second classification strategy A is selected, the inputs of the twelve models are respectively the relevant data of the mobile terminal using pure image type advertising images as page advertising images and the corresponding first preprocessing results, the relevant data of the mobile terminal using pure image type advertising images as pop-up advertising images and the corresponding first preprocessing results, the relevant data of the mobile terminal using pure image type advertising images as inserted advertising images and the corresponding first preprocessing results, the mobile terminal using image and text combined advertising images as page advertising images and the corresponding first preprocessing results, the mobile terminal using image and text combined advertising images as pop-up advertising images and the corresponding first preprocessing results, and the mobile terminal using image and text combined advertising images as inserted advertising images. And corresponding to the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the page advertising image, and the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the pop-up advertising image, and the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the inserted advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the page advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the relevant data of the pop-up advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the relevant data of the inserted advertising image, and the first preprocessing result;

[0029] When the second classification strategy B is selected, the inputs of the six models are respectively the relevant data of the pure image type advertising image as the page advertising image and the corresponding first preprocessing result, the relevant data of the pure image type advertising image as the pop-up advertising image and the corresponding first preprocessing result, the relevant data of the pure image type advertising image as the inserted advertising image and the corresponding first preprocessing result, the relevant data of the image and text combined advertising image as the page advertising image and the corresponding first preprocessing result, the relevant data of the image and text combined advertising image as the pop-up advertising image and the corresponding first preprocessing result, and the relevant data of the image and text combined advertising image as the inserted advertising image and the corresponding first preprocessing result;

[0030] When the second classification strategy C is selected, the inputs of the three models are the relevant data of the page advertising image and the corresponding first preprocessing result, the relevant data of the pop-up advertising image and the corresponding first preprocessing result, and the relevant data of the inserted advertising image and the corresponding first preprocessing result.

[0031] As a preferred solution of the advertising image classification method based on the deep convolutional neural network model described in the present invention, wherein: the combination of the preset improved clustering algorithm and the model output obtained by different second classification strategies, obtaining the third classification result includes: dividing the model output obtained by different second classification strategies into three clustering results through the improved clustering algorithm.

[0032] As a preferred solution of the advertising image classification method based on the deep convolutional neural network model described in the present invention, it also includes: selecting a corresponding classification in the third classification result according to the target demand effect for the advertising image.

[0033] In a second aspect, the present invention provides an advertising image classification system based on a deep convolutional neural network model, comprising:

[0034] A first classification module, used to perform a first classification on the collected target advertisement images and related data, and perform a first preprocessing on the first classification results;

[0035] A second classification module is used to perform a second classification on the first preprocessing result and determine the number of models to be established according to the second classification result;

[0036] A model building module, used to build a corresponding model based on the determined model building quantity in combination with a deep convolutional neural network, wherein the output of the model is the first customer attraction;

[0037] The third classification module is used to obtain a third classification result by combining a preset improved clustering algorithm and model outputs obtained by different second classification strategies.

[0038] In a third aspect, the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned method when executing the computer program.

[0039] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method described above when executed by a processor.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes an advertising image classification method and system based on a deep convolutional neural network model, performs a first classification on the collected target advertising images and related data, and performs a first preprocessing on the first classification result; performs a second classification on the first preprocessing result, and determines the number of model establishments according to the second classification result; according to the determined number of model establishments, a corresponding model is established in combination with a deep convolutional neural network, and the output of the model is the first customer attraction; a third classification result is obtained by combining a preset improved clustering algorithm and the model output obtained by different second classification strategies. Through the above steps, the present invention can effectively classify advertising images, thereby providing more accurate decision support for advertising placement. The advertising image classification method and system of the present invention can not only improve the accuracy of advertising placement, but also can perform self-optimization according to customer feedback and market changes, thereby providing users with more efficient and personalized advertising placement solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0042] Figure 1 A method flow chart of a method and system for advertising image classification based on a deep convolutional neural network model provided by one embodiment of the present invention;

[0043] Figure 2 Another method flow chart of an advertising image classification method and system based on a deep convolutional neural network model provided by an embodiment of the present invention;

[0044] Figure 3 An internal structural diagram of a computer device of an advertising image classification method and system based on a deep convolutional neural network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0046] Example 1

[0047] Reference Figure 1-Figure 3 , which is the first embodiment of the present invention, and provides an advertisement image classification method and system based on a deep convolutional neural network model, comprising:

[0048] Before describing the embodiments of the present application in detail, some related concepts are first explained for the sake of clarity.

[0049] Deep Convolutional Neural Network: This is a deep learning model specifically designed to process data with a grid structure (such as images). It automatically identifies features in the input data through components such as convolutional layers, pooling layers, and fully connected layers, and performs well in tasks such as image recognition and classification.

[0050] Page advertising image: This refers to the advertising image displayed on the web page, which can be directly seen by customers when browsing the web page. This type of advertisement is usually a static or dynamic image, sometimes including video elements, placed in different positions of the page, such as the top, sidebar, etc.

[0051] Pop-up ad images: These ads appear suddenly in a new window while a customer is browsing a webpage. They may appear immediately when a page loads or when a customer completes a certain action, such as clicking a button, with the goal of grabbing the customer's attention and prompting them to take action.

[0052] Interstitial images: Interstitial images, also known as interstitials or banners, are embedded within the content a customer is reading, usually appearing as the customer scrolls. These ads are designed to grab the customer's attention while they are consuming content.

[0053] Clustering Algorithm: This is an unsupervised learning technique used to group data based on similarities between data objects. Clustering algorithms can help identify natural clusters or patterns between data points in a dataset that do not have explicit labels. In this context, clustering algorithms are used to further summarize and organize the results of advertising image classification.

[0054] The "customer" mentioned in this application is a person who generates relevant data behavior for the target advertising image.

[0055] The "user" mentioned in this application is a person who may implement the method or system of this application.

[0056] In the related technologies, there are some problems, such as low classification accuracy, complex model building and optimization processes, and difficulty in adapting to the ever-changing advertising market and customer preferences.

[0057] This application provides a method that can effectively solve the above problems. Next, we will combine multiple embodiments to explain in detail how to implement the advertising image classification method based on the deep convolutional neural network model.

[0058] Figure 1 A method flow chart of an advertisement image classification method and system based on a deep convolutional neural network model is shown, including:

[0059] S101, performing a first classification on the collected target advertisement image and related data, and performing a first preprocessing on the first classification result;

[0060] S102, performing a second classification on the first preprocessing result, and determining the number of models to be established according to the second classification result;

[0061] S103, establishing a corresponding model based on the determined number of model establishments in combination with a deep convolutional neural network, wherein the output of the model is the first customer attraction;

[0062] S104, combining the preset improved clustering algorithm and the model outputs obtained by different second classification strategies to obtain a third classification result.

[0063] Wherein, S101, performing a first classification on the collected target advertisement image and related data, and performing a first preprocessing on the first classification result;

[0064] In the embodiment of the present application, the target advertising image includes a page advertising image, a pop-up advertising image, and an inserted advertising image on a mobile terminal and a PC terminal;

[0065] It should be noted that this application analyzes the acceptance of different client (mobile and PC) customers for different advertising formats and their reactions to advertising content, thereby classifying advertising images. By analyzing customer behavior data, such as click-through rate, browsing time, etc., the impact of advertising images on the attractiveness of different client customers can be more accurately evaluated.

[0066] In the embodiment of the present application, the relevant data includes the length of time that mobile and / or PC clients stay on page advertisement images, the number of clicks and the length of time that they enter pop-up advertisement images, and the number of clicks on inserted advertisement images;

[0067] It should be noted that page ads are usually fixed in a certain position on the web page, such as the top, bottom or sidebar. Customers may accidentally notice these ads when browsing the web page, and whether they continue to browse these ads reflects the degree of interest of the customers in the ads. Therefore, page ad images mainly focus on the length of time customers stay, because a longer length of time may mean that customers are interested in the advertising content.

[0068] It should also be noted that pop-up ads are window ads that suddenly appear when customers are browsing the web. This type of ad is easy to attract customers' attention, but it is also easy to disturb customers' normal browsing experience. Therefore, for pop-up ads, in addition to caring whether customers click to enter the ad details (that is, the number of clicks), we also pay attention to the length of time customers stay on the pop-up ads. These two indicators together reflect the customer's attention to pop-up ads and the attractiveness of the ad content to customers.

[0069] It should also be noted that interstitial ads usually appear between content when customers are reading articles or browsing content, similar to traditional magazine ads. This form of advertising usually does not interrupt the customer's browsing process, but waits for the customer to actively click. Therefore, for interstitial advertising images, the main consideration is the number of customer clicks, because it directly reflects the customer's interest and intention in the advertisement.

[0070] In an optional embodiment, for each advertisement type (page advertisement, pop-up advertisement, insertion advertisement), a set of parameters may be defined for each customer. For example, for customer u, the dwell time of the page advertisement may be expressed as T pageu , the number of clicks on pop-up ads can be expressed as C popupu , the duration of the pop-up ad can be expressed as T popupu , and the number of clicks on the interstitial ad can be expressed as C insertu If we consider the set U of all customers, the interaction data of each customer can be represented as an element in a set. For example, for the customer set U, the page ad stay time of all customers can be represented as the set {T pageu |u∈U}, the number of clicks on pop-up ads can be expressed as a set {C popupu |u∈U}, the duration of pop-up ads can be expressed as a set {T popupu |u∈U}, and the number of clicks on the interstitial advertisement can be represented as the set {C insertu |u∈U}.

[0071] It should be noted that when the customer's relevant data is used in the subsequent model training process, the data in the collection can be directly called.

[0072] In the embodiment of the present application, the first classification includes classifying the collected target advertising images into pure image-type advertising images and image-and-text-combined advertising images;

[0073] It should be noted that the diversity of advertising images was taken into consideration at the beginning of the design of this application, including those containing only image content and those containing both image and text content. Advertising images that combine images and text are usually able to provide richer information because they not only attract customers through visual elements, but also convey more specific information and appeals through text. However, it is unknown whether this type of advertising is more effective than pure image advertising in certain circumstances, so it is necessary to distinguish and analyze these two types of advertising images.

[0074] In the embodiment of the present application, the first preprocessing includes performing a first color preprocessing on a pure image type advertisement image, and performing a first text preprocessing after performing a first color preprocessing on an image and text combination advertisement image.

[0075] Specifically, the first color preprocessing is used to perform color analysis and layout analysis on pure image type advertisement images or / and image and text combined advertisement images;

[0076] It should be noted that color is one of the key factors affecting visual effects, and different color combinations can convey different emotions and information. By analyzing the colors of advertising images, we can understand the colors used in the image and their matching, which is very useful for evaluating the visual appeal of the advertisement. Color analysis helps to identify the main tones in the image, which is very necessary for subsequent image processing (such as contrast enhancement, color correction, etc.). In addition, color analysis can also help identify important areas in the image, which may be the parts that users want to highlight.

[0077] It should also be noted that layout refers to the positional relationship of the various elements in the image, how the visual elements, including colors, are distributed in space. A good layout can make the advertising message easier for the recipient to understand, thereby improving the effectiveness of the advertisement. Analyzing the color layout helps to understand which colors are placed where and how these colors interact with each other. For example, bright colors are often used to attract attention, while contrasting color combinations can be used to emphasize specific information. In advertising images, the layout of colors is very important for creating a sense of visual hierarchy, which can help distinguish the main information and supporting information of the advertisement, thereby guiding the customer's eyes to move as the designer intended.

[0078] In the embodiment of the present application, the first color preprocessing is specifically performed as follows:

[0079] The steps for color analysis can be:

[0080] Step 1: Extract the color value of each pixel from the image. For the RGB color space, the color value of each pixel can be represented as a triplet (R i ,G i ,B i ), where R i , G i , and B i Represents the intensity values ​​of the red, green, and blue color channels respectively.

[0081] Step 2: Count the frequency of each color in the image to form a color histogram. The color histogram can be used to describe the distribution of colors in an image. Assuming there are n pixels in the image, the color histogram can be expressed as:

[0082]

[0083] Here, δ is a function that compares whether a given color value (R, G, B) matches the color value of pixel i.

[0084] Step 3: Convert the RGB color space to HSV (hue, saturation, value) or other color space more suitable for analysis to better understand the color properties.

[0085] Step 4: Extract color features based on the HSV color space, such as the average value of hue, standard deviation of saturation, etc. This application uses the average value of hue as the color feature. The average value of hue can be obtained by calculating the average value of all pixel values ​​of the hue channel in the HSV color space. The average value of hue helps to understand the overall hue tendency of the advertising image, thereby evaluating its possible psychological impact on the target audience.

[0086] The steps for layout analysis can be:

[0087] Step 1: Divide the image into m×n grids, and analyze each grid as an independent region. The size of each grid can be determined according to the resolution of the image and the level of detail that needs to be extracted.

[0088] Step 2: For each grid, repeat the above color analysis steps to obtain the color histogram of each grid. Construct the layout matrix L, where each element in the matrix represents the color feature vector of a grid. The matrix can be expressed as:

[0089]

[0090] Among them, H ij Represents the color histogram of the grid at row i and column j.

[0091] Step 3: Normalize the color histogram of each grid to have the same scale:

[0092]

[0093] Where l and k represent the column index and row index in the grid matrix L, respectively. Specifically, k and l are used to traverse all grids in the entire grid matrix to obtain the sum of all grid color histograms. This ensures that all color histograms have the same scale, which is convenient for comparison and further processing.

[0094] In an embodiment of the present application, the first text preprocessing is used to perform a relevance level analysis on the text in the image and text combined advertising image. The relevance level analysis includes the number of judgments from "the text in the image and text combined advertising image" to the meaning of the target advertising image itself. The number of judgments includes the number of manual judgments and / or the number of judgments in a preset judgment database.

[0095] It should be noted that the judgment database is a preset database set of judgment times from "text in the image and text combined advertising image" to the meaning of the target advertising image itself;

[0096] It should also be noted that the number of manual judgments is similar to the judgment effect of the database, except that manual judgment is more random and the judgments of different people are uncertain. Therefore, in this application, it is more recommended to establish a database.

[0097] Exemplarily, the establishment of the judgment database can be established according to the following specific example. For tourism advertisements, the text in the initial advertisement image is: "Tianya Haijiao", and the judgment path is as follows: "Tianya Haijiao" -> "Beach" (first judgment), "Beach" -> "Vacation" (second judgment), "Vacation" -> "Tourism" (third judgment), advertising intention: promotion of tourist attractions, judgment times: 3 times;

[0098] For Valentine's Day gift advertisements, the text in the initial advertisement image is: "The sweetness of love", and the judgment path is as follows: "The sweetness of love" -> "Love" (first judgment), "Love" -> "Valentine's Day" (second judgment), "Valentine's Day" -> "Gift" (third judgment), advertising intention: to promote Valentine's Day gifts, judgment times: 3 times;

[0099] For car ads, the text in the initial ad image is: "Fast and Furious", and the judgment path is as follows: "Fast and Furious" -> "Speed" (first judgment), "Speed" -> "Driving" (second judgment), "Driving" -> "Car" (third judgment), advertising intention: to promote car brands, judgment times: 3 times;

[0100] For beverage advertisements, the text in the initial advertisement image is: "Refreshing summer", and the judgment path is as follows: "Refreshing summer" -> "summer" (first judgment), "summer" -> "cold drink" (second judgment), "cold drink" -> "beverage" (third judgment), advertising intention: to promote summer beverages, judgment times: 3 times;

[0101] For the bank financial product advertisement, the text in the initial advertisement image is: "Steady Return", and the judgment path is as follows: "Steady Return" -> "Investment" (first judgment), "Investment" -> "Financial Management" (second judgment), "Financial Management" -> "Bank" (third judgment), advertising intention: to promote bank financial products, judgment times: 3 times;

[0102] For holiday promotion ads, the text in the initial ad image is: "Merry Christmas", and the judgment path is as follows: "Merry Christmas" -> "Christmas" (first judgment), "Christmas" -> "Promotion" (second judgment), advertising intent: Christmas promotion, number of judgments: 2 times.

[0103] It should be noted that this application takes into account the impact of this number of judgments on advertising, and can more accurately evaluate the relevance between advertising images and target audiences. By analyzing the correlation between the text in the advertising image and the image content, the intent of the advertisement and the potential market response can be better understood. For example, for tourism advertisements, by judging the path "End of the Earth" -> "Seaside" -> "Vacation" -> "Travel", it can be determined that the intention of the advertisement is to promote tourist attractions. This analysis not only helps users optimize advertising content, but also provides market researchers with valuable information about consumer preferences.

[0104] In practical applications, the establishment and maintenance of the judgment database is crucial. The database should contain a wide range of associated paths for text and image content to ensure that it can cover all types of advertisements. In addition, the database should be updated regularly to reflect changes in market trends and consumer behavior. By continuously optimizing the judgment path and database, the accuracy and efficiency of advertising analysis will be significantly improved.

[0105] In an optional embodiment, the first preprocessing may also include extracting features such as font, font size, and color of the text in the image. These features are crucial to understanding the visual performance and appeal of the text in the advertising image. By analyzing the font type and size of the text, the degree of emphasis placed by the advertiser on a specific message can be inferred; and the color of the text can correspond to the overall tone of the advertisement, enhancing the overall effect of the advertisement.

[0106] In an optional embodiment, the first preprocessing may also include analyzing the position and arrangement of text in the image for the text layout in the advertisement image. For example, whether the text is located in a prominent position in the image, and whether it is presented in a specific arrangement (such as horizontal, vertical or diagonal), all of which will affect the attractiveness and readability of the text in the advertisement. Through layout analysis, the visual weight and guiding role of the text in the advertisement image can be evaluated, providing strong support for subsequent evaluation of the advertising effect.

[0107] In an optional embodiment, the first preprocessing may also include a more in-depth analysis of the relevance between the text in the advertisement image and the image content. For example, by analyzing the spatial relationship between the text and specific elements in the image (such as products, characters, scenes, etc.), the role played by the text in the advertisement and the information conveyed may be inferred. This analysis helps to better understand the intent of the advertisement and the audience's response, and provide users with more targeted optimization suggestions.

[0108] However, this application does not elaborate on other optional contents, but is only intended to inform other technicians that they can consider designs in other directions.

[0109] In summary, step S101 can pre-process the text in the image-text combined advertising image in a systematic and standardized manner, thereby improving the efficiency and accuracy of advertising analysis. Through correlation level analysis, the degree of correlation between the text in the advertising image and the meaning of the target advertising image itself can be quickly and accurately understood, thereby further revealing the intention of the advertisement and potential market response. This not only provides users with a scientific basis for optimizing advertising content, but also provides market researchers with valuable consumer preference information.

[0110] In addition, the method of establishing a judgment database proposed in this application can effectively reduce the subjectivity and uncertainty of manual judgment and improve the objectivity and accuracy of advertising analysis. Each judgment path in the database is based on the statistics and analysis of a large amount of sample data and has high credibility and representativeness. Through this method, the relationship between the text and image content in the advertising image can be grasped more accurately, and the intention of the advertisement and the audience response can be deeply understood.

[0111] It is worth noting that with the continuous innovation and diversification of advertising forms, the judgment database also needs to be continuously updated and improved. Only by maintaining the timeliness and comprehensiveness of the database can the accuracy and effectiveness of advertising analysis be ensured. Therefore, in practical applications, it is necessary to regularly review and update the database, and timely incorporate new advertising types and judgment paths to adapt to the needs of market changes.

[0112] S102, performing a second classification on the first preprocessing result, and determining the number of models to be established according to the second classification result;

[0113] In the embodiment of the present application, performing a second classification on the first preprocessing result, and determining the number of model establishments according to the second classification result includes:

[0114] The second classification includes performing a second classification on the first preprocessing result by using a preset second classification strategy;

[0115] The preset second classification strategies include second classification strategy A, second classification strategy B and second classification strategy C;

[0116] When the second classification strategy A is selected, the number of models established is twelve according to the second classification results;

[0117] When the second classification strategy B is selected, the number of models established is six according to the second classification results;

[0118] When the second classification strategy C is selected, the number of models to be established is determined to be three according to the second classification result.

[0119] It should be noted that when the user implementing the present application selects the preset second classification strategy, the present application provides three different strategy selection methods. For example, when the user's funds are insufficient, it is necessary to determine which client or which type of advertisement (page advertisement, pop-up advertisement, insert advertisement) can help him to get back his investment more. Then he can choose the second classification strategy A; when the user's funds are slightly sufficient and he wants to test the effect of the advertisement on different clients (such as mobile terminals and PC terminals), but he is not sure whether the advertisement content should be pure image type or image and text combination type, it may be appropriate to choose strategy B. In this case, the user may want to test two types of advertisements on different clients to observe which combination can bring better customer appeal and conversion rate. Therefore, choosing to establish six models can cover different combinations of clients and advertisement types while maintaining relatively low cost investment. When the user has sufficient funds and plans to place advertisements on all clients at the same time, but the main focus is on the effect of the advertisement type (page advertisement, pop-up advertisement, insert advertisement), strategy C should be selected. In this case, the user may have determined the form of the advertisement (pure image type or image and text combination type), and now only needs to verify the performance in different scenarios. At this point, establishing three models to focus on different types of ad display methods can concentrate resources on optimizing specific ad types, thereby achieving the best advertising results on all platforms.

[0120] When the second classification strategy A is selected, twelve models are built:

[0121] Mobile: 3 types of ads (page ads, pop-up ads, and insert ads) × 2 types of ad images (pure image type, image and text combination type) = 6 models.

[0122] PC: Again, 3 ad types x 2 ad image types = 6 additional models.

[0123] Total: 6+6=12 models.

[0124] When the second classification strategy B is selected, six models are built:

[0125] For two types of advertising images (pure image type and image and text combination type), models are established for each type in three scenarios: page advertising, pop-up advertising, and inserted advertising.

[0126] Total: 2 ad image types x 3 ad types = 6 models.

[0127] When the second classification strategy C is selected, three models are built:

[0128] For all ad images, whether they are page ads, pop-up ads, or interstitial ads, a model is built to handle them.

[0129] Total: 3 ad types = 3 models.

[0130] It should be noted that the benefit of S102 is that it provides a flexible selection space, allowing users to choose the appropriate second classification strategy according to their own financial situation, testing needs and target market, and build a corresponding number of models accordingly. This personalized selection method not only helps users to allocate resources reasonably, but also ensures the effectiveness and pertinence of advertising testing.

[0131] Specifically, by choosing the second classification strategy A, users can build more models to comprehensively cover various types of advertisements, thereby exploring and optimizing advertising effects on a larger scale. This method is suitable for users who have insufficient funds and have high requirements for advertising effects.

[0132] The second classification strategy B allows you to test across clients and ad types while keeping costs low. This approach is suitable for users with relatively limited funds who want to make initial attempts on different platforms and ad formats.

[0133] As for the second classification strategy C, it is more focused on testing the type of advertisement. By concentrating resources on a specific type of advertisement, users can gain a deeper understanding of the performance of this type of advertisement in different scenarios, thereby optimizing the advertisement content and delivery strategy. This method is suitable for users who have already determined the form of advertisement but need to further verify its effect.

[0134] In summary, step S102 provides users with a flexible and efficient advertising testing and optimization solution by providing a variety of second classification strategies and corresponding model building quantity options. This not only helps users improve advertising effects, but also maximizes market returns with limited resources.

[0135] S103, establishing a corresponding model based on the determined number of model establishments in combination with a deep convolutional neural network, wherein the output of the model is the first customer attraction;

[0136] It should be noted that the first customer attraction is the probability that a customer is attracted by this advertisement.

[0137] In the embodiment of the present application, according to the determined number of model establishment, establishing a corresponding model in combination with a deep convolutional neural network includes:

[0138] When the second classification strategy A is selected, the inputs of the twelve models are respectively the relevant data of the mobile terminal using pure image type advertising images as page advertising images and the corresponding first preprocessing results, the relevant data of the mobile terminal using pure image type advertising images as pop-up advertising images and the corresponding first preprocessing results, the relevant data of the mobile terminal using pure image type advertising images as inserted advertising images and the corresponding first preprocessing results, the mobile terminal using image and text combined advertising images as page advertising images and the corresponding first preprocessing results, the mobile terminal using image and text combined advertising images as pop-up advertising images and the corresponding first preprocessing results, and the mobile terminal using image and text combined advertising images as inserted advertising images. And corresponding to the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the page advertising image, and the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the pop-up advertising image, and the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the inserted advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the page advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the relevant data of the pop-up advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the relevant data of the inserted advertising image, and the first preprocessing result;

[0139] When the second classification strategy B is selected, the inputs of the six models are respectively the relevant data of the pure image type advertising image as the page advertising image and the corresponding first preprocessing result, the relevant data of the pure image type advertising image as the pop-up advertising image and the corresponding first preprocessing result, the relevant data of the pure image type advertising image as the inserted advertising image and the corresponding first preprocessing result, the relevant data of the image and text combined advertising image as the page advertising image and the corresponding first preprocessing result, the relevant data of the image and text combined advertising image as the pop-up advertising image and the corresponding first preprocessing result, and the relevant data of the image and text combined advertising image as the inserted advertising image and the corresponding first preprocessing result;

[0140] When the second classification strategy C is selected, the inputs of the three models are the relevant data of the page advertising image and the corresponding first preprocessing result, the relevant data of the pop-up advertising image and the corresponding first preprocessing result, and the relevant data of the inserted advertising image and the corresponding first preprocessing result.

[0141] It should be noted that when a pure image type advertisement image is selected, the first preprocessing result only includes the color analysis and layout analysis results, and when an image and text combined advertisement image is selected, the first preprocessing result includes the color analysis result, the layout analysis result, and the correlation level analysis result;

[0142] When the second classification strategy A is selected, for each model, its input is the advertising image and its related data corresponding to the selected advertising type (page advertising, pop-up advertising, and inserted advertising) and advertising image type (pure image type, image and text combination type), combined with the result of the first preprocessing. The output of the model is the first customer attractiveness, that is, the attractiveness of the advertisement to customers predicted by the model.

[0143] In the embodiment of the present application, when the second classification strategy A is selected, the loss functions used by the twelve models are (of course, other types of loss functions can be selected):

[0144] L improved =α·L+β·R

[0145] Among them, R is a regularization term used to penalize excessive model complexity, and α and β are weight coefficients that balance the contribution of the two parts.

[0146]

[0147] Among them, y i is the actual customer attractiveness of the ith sample, is the customer attractiveness predicted by the model for the sample, and N is the number of samples.

[0148] Exemplarily, when the second classification strategy A is selected, the grid architecture of the deep convolutional neural network model established can be:

[0149] Input layer: accepts images of shape (224, 224, 3).

[0150] Convolutional layer 1: 64 3x3 filters, ReLU activation, followed by a batch normalization layer.

[0151] Max pooling layer: 2x2 window.

[0152] Convolutional layer 2: 128 3x3 filters, ReLU activation, followed by a batch normalization layer.

[0153] Max pooling layer: 2x2 window.

[0154] Convolutional layer 3: 256 3x3 filters, ReLU activation, followed by a batch normalization layer.

[0155] Max pooling layer: 2x2 window.

[0156] Flatten layer: Converts the feature map into a one-dimensional vector.

[0157] Fully connected layer: 512 nodes, ReLU activation.

[0158] Dropout layer: randomly discard a certain proportion of nodes to prevent overfitting.

[0159] Fully connected layer: 256 nodes, ReLU activation.

[0160] Output layer: single node, linear activation.

[0161] When the second classification strategy B is selected, for each model, the input of each model is the relevant data of a specific type of advertising image (pure image type or image and text combination type) in a specific advertising context (page advertisement, pop-up advertisement, inserted advertisement) and the corresponding first preprocessing result. The output of the model is the first customer attractiveness, that is, the attractiveness of the advertisement to customers predicted by the model.

[0162] In the embodiment of the present application, when the second classification strategy B is selected, the loss functions used by the six models are (of course, other types of loss functions can be selected):

[0163] L improved =αL+βR+γS

[0164] Among them, L represents the basic mean square error loss function (MSE), which is used to measure the gap between the first customer attractiveness predicted by the model and the actual customer attractiveness.

[0165]

[0166] Among them, R is a regularization term, which is used to penalize the complexity of the model and prevent overfitting. You can use L2 regularization:

[0167]

[0168] Among them, λ represents the weight decay parameter, which controls the strength of the regularization term. A larger λ will increase the regularization strength, while a smaller λ allows for greater model complexity. S represents the supplementary loss term related to ad interaction, such as the penalty term for click-through rate or conversion rate, which can be calculated as follows:

[0169]

[0170] Among them, δ is the coefficient for adjusting the impact of the supplementary loss term, CTR i is the click-through rate of the i-th sample, CTR avg is the average click rate of all samples. The coefficients α, β, and γ are weight coefficients used to balance the contribution of the three loss parts and can be adjusted according to actual conditions.

[0171] Exemplarily, when the second classification strategy B is selected, the grid architecture of the deep convolutional neural network model established can be:

[0172] Input layer: Receives image data of size (width, height, channels). Assuming the width is 224 pixels, the height is 224 pixels, and the number of color channels is 3 (RGB image), the shape of the input layer is (224, 224, 3).

[0173] Convolutional layer 1: uses 64 3x3 kernels, stride 1, padding 'same' to keep the input and output dimensions the same, and activation function is ReLU.

[0174] Batch Normalization Layer 1: Normalizes the output of Convolutional Layer 1.

[0175] Pooling layer 1: Use a 2x2 max pooling layer to reduce the spatial dimension.

[0176] Convolutional layer 2: Use 128 3x3 kernels, stride 1, padding 'same', and activation function ReLU.

[0177] Batch Normalization Layer 2: Normalizes the output of Convolutional Layer 2.

[0178] Pooling layer 2: A 2x2 max pooling layer is used to reduce the spatial dimension.

[0179] Convolutional layer 3: uses 256 3x3 kernels, stride 1, padding 'same', and activation function ReLU.

[0180] Batch Normalization Layer 3: Normalizes the output of Convolutional Layer 3.

[0181] Pooling layer 3: A 2x2 max pooling layer is used to reduce the spatial dimension.

[0182] Flatten layer: Flattens the output of the convolutional layer into a one-dimensional vector.

[0183] Fully connected layer 1: has 512 nodes and the activation function is ReLU.

[0184] Dropout layer: Dropout is used to prevent overfitting.

[0185] Fully connected layer 2: has 256 nodes and the activation function is ReLU.

[0186] Output layer: A single node that uses a linear activation function to predict customer attractiveness.

[0187] When the second classification strategy C is selected, the input of each model is the relevant data of a specific advertisement type (page advertisement, pop-up advertisement, insert advertisement) and the corresponding first preprocessing result. For all advertisement images, whether they are pure image type or image and text combination type, the same model will be used for processing. The output of the model is the first customer attractiveness, that is, the degree of attractiveness of the advertisement to customers predicted by the model.

[0188] In the embodiment of the present application, when the second classification strategy C is selected, the loss functions used by the three models are (of course, other types of loss functions can be selected):

[0189] L improved =αL+βR+γW

[0190] Among them, L represents the basic mean square error loss function (MSE), which is used to measure the gap between the first customer attractiveness predicted by the model and the actual customer attractiveness.

[0191]

[0192] Among them, R is a regularization term, which is used to penalize the complexity of the model and prevent overfitting. You can use L2 regularization:

[0193]

[0194] Where W represents the supplementary loss term related to the ad type, such as the penalty term for click-through rate or conversion rate, which can be calculated as follows:

[0195]

[0196] Among them, ωk is the weight of the k-th advertising type, f k is a characteristic function related to the k-th type of advertisement (such as the browsing time of page advertisements, the click-through rate of pop-up advertisements). The coefficients α, β, and γ are weight coefficients used to balance the contributions of the three loss parts and can be adjusted according to the actual situation.

[0197] Exemplarily, when the second classification strategy B is selected, the grid architecture of the deep convolutional neural network model established can be:

[0198] Input layer: Receives image data of size (width, height, channels). Assuming the width is 224 pixels, the height is 224 pixels, and the number of color channels is 3 (RGB image), the shape of the input layer is (224, 224, 3).

[0199] Convolutional layer 1: Use 32 3x3 kernels, stride 1, padding 'same' to keep the input and output dimensions the same, and activation function is ReLU.

[0200] Batch Normalization Layer 1: Normalizes the output of Convolutional Layer 1.

[0201] Pooling layer 1: Use a 2x2 max pooling layer to reduce the spatial dimension.

[0202] Convolutional layer 2: Use 64 3x3 kernels, stride 1, padding 'same', and activation function ReLU.

[0203] Batch Normalization Layer 2: Normalizes the output of Convolutional Layer 2.

[0204] Pooling layer 2: A 2x2 max pooling layer is used to reduce the spatial dimension.

[0205] Convolutional layer 3: uses 128 3x3 kernels, stride 1, padding 'same', and activation function ReLU.

[0206] Batch Normalization Layer 3: Normalizes the output of Convolutional Layer 3.

[0207] Pooling layer 3: A 2x2 max pooling layer is used to reduce the spatial dimension.

[0208] Flatten layer: Flattens the output of the convolutional layer into a one-dimensional vector.

[0209] Fully connected layer 1: has 256 nodes and the activation function is ReLU.

[0210] Dropout layer: Dropout is used to prevent overfitting.

[0211] Fully connected layer 2: has 128 nodes and the activation function is ReLU.

[0212] Output layer: A single node that uses a linear activation function to predict customer attractiveness.

[0213] It should be noted that by customizing the corresponding loss function and model architecture for different ad types, the unique characteristics of each type of ad can be captured more accurately, thereby improving the model's prediction accuracy of customer attractiveness. Specifically, the introduction of the supplementary loss term allows the model to not only focus on improving prediction accuracy during training, but also consider additional information related to the ad type, such as click-through rate or conversion rate, which is crucial for evaluating ad effectiveness. In addition, by adjusting the weight coefficients α, β, and γ, the contribution of different loss components to model training can be flexibly balanced to adapt to different application scenarios and needs.

[0214] S104, combining the preset improved clustering algorithm and the model outputs obtained by different second classification strategies to obtain a third classification result.

[0215] In an embodiment of the present application, in combination with a preset improved clustering algorithm and model outputs obtained by different second classification strategies, obtaining a third classification result includes: dividing the model outputs obtained by different second classification strategies into three clustering results through the improved clustering algorithm.

[0216] In an optional embodiment, the specific steps of dividing the model output obtained by the second classification strategy A into three clustering results by improving the clustering algorithm may be:

[0217] Step 1: When the second classification strategy A or B is selected, the output results of all twelve models or six models are collected, which represent customer attractiveness under different clients (mobile, PC) and different advertising types (page ads, pop-up ads, and insert ads).

[0218] Step 2: Randomly select three initial cluster centers as the initial center points.

[0219] Step 3: Calculate the distance from each model output to each cluster center. Commonly used distance measures include Euclidean distance, Manhattan distance, etc. For Euclidean distance, the following formula can be used:

[0220]

[0221] Among them, x is the output of a model, c is a cluster center, and n is the dimension of the feature.

[0222] Step 4: Assign the output of each model to the closest cluster center. This can be done by comparing the distance of each model output to all cluster centers.

[0223] Step 5: Introduce an adaptive learning rate so that each update is adjusted according to the distribution of the current data points and the cluster center is recalculated. The formula is as follows:

[0224]

[0225] Among them, C j is the set of all model outputs assigned to the jth cluster center, η t It is the learning rate adjusted with the number of iterations t. The initial value is large and gradually decreases with the increase of the number of iterations.

[0226] Step 6: Repeat steps 3 to 5 and continue iterating until the cluster center no longer changes or the preset number of iterations is reached.

[0227] Step 7: When the change of cluster center is less than a certain threshold or the maximum number of iterations is reached, the algorithm stops.

[0228] Step 8: Output the final three cluster centers and the cluster to which each model output belongs.

[0229] In an optional embodiment, when the demand changes from three clustering results to other numbers, it is only necessary to adjust the number of center points in the improved aggregation step accordingly;

[0230] In an embodiment of the present application, when the second classification strategy C is selected, no clustering operation is required, and the customer attractiveness of the outputs of the three models is directly used as the clustering result. Regardless of how the clustering type changes, no clustering operation is performed when the second classification strategy C is selected, and the model output results are directly used as the final criterion for attractiveness judgment.

[0231] It should be noted that by combining the preset improved clustering algorithm, the complex model outputs obtained under different second classification strategies can be effectively processed and converted into more representative classification results. This capability enables the system to adapt to changes in demand in different scenarios and improves overall adaptability and practicality. By adjusting the number of cluster centers, the clustering results can be easily expanded from three to other types to meet a wider range of needs. This flexibility enables the system to demonstrate stronger adaptability and competitiveness when facing a variety of application scenarios. The prediction accuracy of the model and additional information related to the type of advertisement are also taken into account. By introducing supplementary loss terms and adjusting weight coefficients, the model can more comprehensively capture the characteristics of various types of advertisements during the training process, thereby improving the accuracy of predictions of customer attractiveness. This is of great significance for improving advertising effectiveness and optimizing advertising delivery strategies.

[0232] Figure 2Another method flow chart of an advertisement image classification method and system based on a deep convolutional neural network model is shown, including:

[0233] S101, performing a first classification on the collected target advertisement image and related data, and performing a first preprocessing on the first classification result;

[0234] S102, performing a second classification on the first preprocessing result, and determining the number of models to be established according to the second classification result;

[0235] S103, establishing a corresponding model based on the determined number of model establishments in combination with a deep convolutional neural network, wherein the output of the model is the first customer attraction;

[0236] S104, combining the preset improved clustering algorithm and the model outputs obtained by different second classification strategies to obtain a third classification result.

[0237] S105: Select a corresponding category from the third classification results according to the target demand effect for the advertisement image.

[0238] In the embodiment of the present application, if the user has insufficient funds, the result of clustering the twelve models of the second classification strategy A is selected. For example, when the user requires a high attractiveness, any one of the several models with high attractiveness after model clustering is selected, and the method of placing advertisements is determined according to the model type;

[0239] If the user has sufficient funds and wants to test the effectiveness of the advertisement on different clients (such as mobile and PC), but is not sure whether the advertisement content should be pure image or a combination of image and text, then the user can select the model that meets the user's attractiveness from the results of the clustering of the six models of the second classification strategy B, and determine the advertisement delivery method based on the type of the selected model.

[0240] When the user has sufficient funds and plans to place ads on all clients at the same time, but the main focus is on the effect of the ad type (page ads, pop-up ads, and insert ads), then the model of the three models in strategy C that meets the user's needs and attractiveness should be selected, and the ad delivery method should be determined based on the type of model selected.

[0241] For example, suppose user Xiao Li is currently short of funds, and he needs to determine which client (mobile or PC) or which type of advertising (page advertising, pop-up advertising, insert advertising) can better help him make his money back. In this case, Xiao Li chose the second classification strategy A, and established and trained twelve models, each corresponding to a different combination of client and advertising image type. Use the improved clustering algorithm to cluster the outputs of the twelve models to find the most attractive models. Xiao Li can choose any one or several of the more attractive models to determine the final advertising delivery method. If the clustering results show that the mobile image and text combined page advertising model has the highest customer appeal, then Xiao Li can choose this model and place the advertisement on the mobile page advertising position.

[0242] In summary, the present invention proposes an advertising image classification method based on a deep convolutional neural network model, which performs a first classification on the collected target advertising images and related data, and performs a first preprocessing on the first classification result; performs a second classification on the first preprocessing result, and determines the number of model establishments according to the second classification result; according to the determined number of model establishments, a corresponding model is established in combination with a deep convolutional neural network, and the output of the model is the first customer attraction; a third classification result is obtained by combining a preset improved clustering algorithm and the model output obtained by different second classification strategies. Through the above steps, the present invention can effectively classify advertising images, thereby providing more accurate decision support for advertising placement. The advertising image classification method and system of the present invention can not only improve the accuracy of advertising placement, but also can perform self-optimization according to customer feedback and market changes, thereby providing users with more efficient and personalized advertising placement solutions.

[0243] This embodiment also provides an advertising image classification system based on a deep convolutional neural network model, including:

[0244] A first classification module, used to perform a first classification on the collected target advertisement images and related data, and perform a first preprocessing on the first classification results;

[0245] A second classification module is used to perform a second classification on the first preprocessing result and determine the number of models to be established according to the second classification result;

[0246] A model building module is used to build a corresponding model based on the determined model building quantity and in combination with a deep convolutional neural network, and the output of the model is the first customer attraction;

[0247] The third classification module is used to obtain a third classification result by combining a preset improved clustering algorithm and model outputs obtained by different second classification strategies.

[0248] The above-mentioned unit modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above-mentioned modules.

[0249] This embodiment also provides a computer device, which may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 3 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a method for classifying advertising images based on a deep convolutional neural network model is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0250] This embodiment further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0251] Performing a first classification on the collected target advertisement images and related data, and performing a first preprocessing on the first classification results;

[0252] Performing a second classification on the first preprocessing result, and determining the number of model establishments according to the second classification result;

[0253] According to the determined number of model establishment, a corresponding model is established in combination with a deep convolutional neural network, and the output of the model is the first customer attraction;

[0254] The third classification result is obtained by combining the preset improved clustering algorithm and the model output obtained by different second classification strategies.

[0255] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0256] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0257] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0258] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0259] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0260] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0261] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. The advertising image classification method based on the deep convolutional neural network model is characterized by: include: Performing a first classification on the collected target advertisement images and related data, and performing a first preprocessing on the first classification results; Performing a second classification on the first preprocessing result, and determining the number of model establishments according to the second classification result includes: the second classification includes performing a second classification on the first preprocessing result by a preset second classification strategy; The preset second classification strategy includes a second classification strategy A, a second classification strategy B and a second classification strategy C; When the second classification strategy A is selected, the number of models established is twelve according to the second classification results; When the second classification strategy B is selected, the number of models established is six according to the second classification results; When the second classification strategy C is selected, the number of models to be established is three according to the second classification results; According to the determined number of model establishment, a corresponding model is established in combination with a deep convolutional neural network, and the output of the model is the first customer attraction; Combined with the preset improved clustering algorithm and the model outputs obtained by different second classification strategies, the third classification result is obtained. Specifically, the model output obtained by the second classification strategy A is divided into three clustering results by the improved clustering algorithm, including: Step 1: When the second classification strategy A or B is selected, the output results of all twelve models or six models are collected; Step 2: Randomly select three initial cluster centers as the initial center points; Step 3: Calculate the distance from each model output to each cluster center; Step 4: Assign the output of each model to the nearest cluster center by comparing the distance between each model output and all cluster centers; Step 5: Introduce an adaptive learning rate so that each update is adjusted according to the distribution of the current data points and the cluster center is recalculated. The formula is as follows: Among them, C j is the set of all model outputs assigned to the jth cluster center, η t is the learning rate adjusted with the number of iterations t, which gradually decreases with the increase of the number of iterations, c j represents the current j-th cluster center, c j ′ represents the updated j-th cluster center, x i Represents the set C j A data point in Step 6: Repeat steps 3 to 5 and continue iterating until the cluster center no longer changes or the preset number of iterations is reached; Step 7: When the change of cluster center is less than a certain threshold or reaches the maximum number of iterations, the algorithm stops; Step 8: Output the final three cluster centers and the cluster to which each model output belongs.

2. The advertising image classification method based on a deep convolutional neural network model according to claim 1, characterized in that: The first classification of the collected target advertisement images and related data and the first preprocessing of the first classification results include: The target advertising images include page advertising images, pop-up advertising images and inserted advertising images on mobile terminals and PC terminals; The relevant data includes the length of time that mobile and / or PC clients stay on page advertisement images, the number of clicks and length of time that clients enter pop-up advertisement images, and the number of clicks on inserted advertisement images; The first classification includes classifying the collected target advertising images into pure image-type advertising images and image-and-text-combined advertising images; The first preprocessing includes performing a first color preprocessing on a pure image type advertisement image, and performing a first text preprocessing after performing a first color preprocessing on an image and text combination advertisement image.

3. The advertising image classification method based on the deep convolutional neural network model as claimed in claim 2, characterized in that: The first preprocessing further comprises: The first color preprocessing is used to perform color analysis and layout analysis on pure image type advertisement images or / and image and text combined advertisement images; The first text preprocessing is used to perform a relevance level analysis on the text in the image and text combined advertising image, and the relevance level analysis includes the number of judgments from "the text in the image and text combined advertising image" to the meaning of the target advertising image itself, and the number of judgments includes the number of manual judgments and / or the number of judgments in a preset judgment database.

4. The advertising image classification method based on a deep convolutional neural network model as claimed in claim 3, characterized in that: The method of establishing a corresponding model based on the determined number of model establishments and combining a deep convolutional neural network comprises: When the second classification strategy A is selected, the inputs of the twelve models are respectively the relevant data of the mobile terminal using pure image type advertising images as page advertising images and the corresponding first preprocessing results, the relevant data of the mobile terminal using pure image type advertising images as pop-up advertising images and the corresponding first preprocessing results, the relevant data of the mobile terminal using pure image type advertising images as inserted advertising images and the corresponding first preprocessing results, the mobile terminal using image and text combined advertising images as page advertising images and the corresponding first preprocessing results, the mobile terminal using image and text combined advertising images as pop-up advertising images and the corresponding first preprocessing results, and the mobile terminal using image and text combined advertising images as inserted advertising images. And corresponding to the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the page advertising image, and the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the pop-up advertising image, and the first preprocessing result, the PC end uses a pure image type advertising image as the relevant data of the inserted advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the page advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the relevant data of the pop-up advertising image, and the first preprocessing result, the PC end uses an image and text combined advertising image as the relevant data of the inserted advertising image, and the first preprocessing result; When the second classification strategy B is selected, the inputs of the six models are respectively the relevant data of the pure image type advertising image as the page advertising image and the corresponding first preprocessing result, the relevant data of the pure image type advertising image as the pop-up advertising image and the corresponding first preprocessing result, the relevant data of the pure image type advertising image as the inserted advertising image and the corresponding first preprocessing result, the relevant data of the image and text combined advertising image as the page advertising image and the corresponding first preprocessing result, the relevant data of the image and text combined advertising image as the pop-up advertising image and the corresponding first preprocessing result, and the relevant data of the image and text combined advertising image as the inserted advertising image and the corresponding first preprocessing result; When the second classification strategy C is selected, the inputs of the three models are the relevant data of the page advertising image and the corresponding first preprocessing result, the relevant data of the pop-up advertising image and the corresponding first preprocessing result, and the relevant data of the inserted advertising image and the corresponding first preprocessing result.

5. The advertising image classification method based on a deep convolutional neural network model as claimed in claim 4, characterized in that: The method of obtaining the third classification result by combining the preset improved clustering algorithm and the model outputs obtained by different second classification strategies includes: dividing the model outputs obtained by different second classification strategies into three clustering results by using the improved clustering algorithm.

6. The advertising image classification method based on a deep convolutional neural network model as claimed in claim 5, characterized in that: Also includes: According to the target demand effect for the advertisement image, a corresponding category in the third classification result is selected.

7. An advertising image classification system based on a deep convolutional neural network model, applied to the method of claim 1, characterized in that: include: A first classification module, used to perform a first classification on the collected target advertisement images and related data, and perform a first preprocessing on the first classification results; A second classification module is used to perform a second classification on the first preprocessing result and determine the number of models to be established according to the second classification result; A model building module, used to build a corresponding model based on the determined model building quantity in combination with a deep convolutional neural network, wherein the output of the model is the first customer attraction; The third classification module is used to obtain a third classification result by combining a preset improved clustering algorithm and model outputs obtained by different second classification strategies.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images