Data analysis and processing method for advertising delivery
By collecting crowd image data in the bus stop and using deep learning algorithms for cluster analysis, the problem of mismatch between advertising content and crowd nature in the existing advertising delivery methods is solved, and precise division of crowds and reasonable allocation of advertising resources is achieved. , improves advertising delivery effect.
Patent Information
- Application Number
- CN202411188749.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-28
AI Technical Summary
The existing advertising delivery methods are not targeted and fail to effectively connect the advertising content with the nature of the population, resulting in poor advertising delivery results.
By setting up monitoring equipment in the bus stop to collect crowd image data, using deep learning algorithms to perform cluster analysis on character image data, people with similar characteristics are divided into one category, and advertising audience analysis and advertising delivery support are provided.
It has achieved accurate division of people and reasonable allocation of advertising resources, and improved the targeted and effective advertising delivery.
Smart Images

Figure CN118710342B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of advertising data analysis, and more specifically, to a data analysis and processing method for advertising delivery. Background Art
[0002] In today's increasingly competitive market environment, publicity and marketing have become an indispensable part of enterprises. Multimedia advertising design, as an innovative and efficient marketing method, is gradually being accepted and applied by more and more companies. Multimedia advertising refers to the use of various forms of media technology, such as images, sounds, animations, etc., to present advertising content vividly and vividly.
[0003] At present, multimedia advertisements can be found in the streets and alleys of the city, such as in bus stops. Because bus stops have a large complete surface, they are very suitable for multimedia advertising. In addition, the flow of people taking the bus is also large, which makes the advertisements have a better dissemination effect.
[0004] When placing advertisements, preliminary research is usually conducted. Generally speaking, the benefits of advertising in places with large traffic volume are bound to be better. However, existing advertising only considers the overall traffic volume and does not link the content of the advertisement with the nature of the crowd, thus lacking the targeting of advertising. Therefore, a data analysis and processing method for advertising is proposed. Summary of the invention
[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a data analysis and processing method for advertising delivery, which collects image data of passing people by using monitoring equipment installed in bus stops, and performs cluster analysis on the image data of people in combination with deep learning algorithms to classify people with similar characteristics into one category. In this way, clustering and dividing the passing people at the monitoring location can provide technical support for advertising audience analysis and advertising delivery, thereby promoting the rational allocation of advertising resources.
[0006] According to one aspect of the present application, a data analysis and processing method for advertisement delivery is provided, comprising:
[0007] Acquire a set of person image data collected by a monitoring device, wherein the monitoring device is set in a bus stop;
[0008] Extracting character image features from the character image data set to obtain a set of enhanced character image feature maps;
[0009] Based on the set of enhanced character image feature maps, constructing a feature distribution pseudo-topic;
[0010] Respectively calculating the semantic relevance between each enhanced person image feature map in the set of enhanced person image feature maps and the feature distribution pseudo-topic to obtain a set of semantic relevance;
[0011] Based on the set of semantic associations, cluster analysis is performed on the set of person image data to obtain a plurality of clustering results, each of which represents a class of people.
[0012] Compared with the prior art, the data analysis and processing method for advertising placement provided by the present application collects image data of people passing by by using monitoring equipment installed in bus stops, and clusters the image data of people in combination with deep learning algorithms to classify people with similar characteristics into one category. In this way, clustering and dividing the people passing by the monitoring location can provide technical support for advertising audience analysis and advertising placement, thereby promoting the rational allocation of advertising resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.
[0014] Figure 1 is a flow chart of a data analysis and processing method for advertising delivery according to an embodiment of the present application;
[0015] Figure 2 A system architecture diagram of a data analysis and processing method for advertisement delivery according to an embodiment of the present application;
[0016] Figure 3 is a flowchart of sub-step S2 of the data analysis and processing method for advertising delivery according to an embodiment of the present application;
[0017] Figure 4 Flow chart of sub-step S22 of the data analysis and processing method for advertising delivery according to an embodiment of the present application. DETAILED DESCRIPTION
[0018] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0019] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.
[0020] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.
[0021] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.
[0022] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0023] Existing advertising only considers the overall flow of people, and does not link the content of the advertisement with the nature of the crowd, so the advertising lacks pertinence. Therefore, a data analysis and processing method for advertising is proposed.
[0024] In the technical solution of the present application, a data analysis and processing method for advertisement delivery is proposed. Figure 1 The present invention is a flowchart of a data analysis and processing method for advertising delivery according to an embodiment of the present application. Figure 2 1 is a system architecture diagram of a data analysis and processing method for advertising delivery according to an embodiment of the present application. Figure 1 and Figure 2As shown, according to the embodiment of the present application, the data analysis and processing method for advertising placement includes the steps of: S1, obtaining a set of character image data collected by a monitoring device, wherein the monitoring device is arranged in a bus stop; S2, performing character image feature extraction on the set of character image data to obtain a set of enhanced character image feature maps; S3, constructing a feature distribution pseudo-topic based on the set of enhanced character image feature maps; S4, respectively calculating the semantic association between each enhanced character image feature map in the set of enhanced character image feature maps and the feature distribution pseudo-topic to obtain a set of semantic associations; S5, based on the set of semantic associations, performing cluster analysis on the set of character image data to obtain multiple clustering results, wherein each clustering result represents a class of people.
[0025] In particular, the S1 obtains a collection of character image data collected by a monitoring device, wherein the monitoring device is set in a bus stop. Here, the bus stop is a place where natural traffic is concentrated, and a large number of passengers wait and transfer here every day. By setting up monitoring equipment in the bus stop, character image data can be effectively collected to provide support for subsequent data analysis and advertising, and to achieve more accurate and effective market promotion. It should be noted that the above-mentioned collection of character image data is only used to analyze the characteristics of character images to obtain clustering results, and the clustering results are used to represent a class of people, such as male people, female people, children, etc.; after obtaining the clustering results, the collected character image data will not be saved, and personal privacy will not be violated.
[0026] In particular, the step S2 extracts character image features from the character image data set to obtain a set of enhanced character image feature maps. In particular, in a specific example of the present application, Figure 3 As shown, the S2 includes: S21, respectively passing each character image data in the set of character image data through a character image feature extractor based on a void convolutional neural network model to obtain a set of character image feature maps; S22, respectively passing each character image feature map in the set of character image feature maps through a feature multi-scale perception enhancement module to obtain a set of enhanced character image feature maps.
[0027] Specifically, in S21, each character image data in the character image data set is passed through a character image feature extractor based on a dilated convolutional neural network model to obtain a set of character image feature maps. Wherein, a dilated convolutional neural network is a convolutional neural network structure that increases the size of the receptive field by introducing dilation between convolution kernels, thereby capturing feature information in a wider range. The character image feature extractor is a model built based on a dilated convolutional neural network, which is used to extract a feature representation of a character from an input image. Each character image data in the character image data set is input into the character image feature extractor respectively, and after being processed by the model, a feature map corresponding to each character image is obtained. In this way, a set of character image feature maps is obtained, in which each feature map represents a feature representation of a character image.
[0028] Specifically, the S22 passes each character image feature map in the set of character image feature maps through a feature multi-scale perception enhancement module to obtain a set of enhanced character image feature maps. Considering that in complex real-life scenes, characters may exhibit diverse behavior patterns and behavior characteristics. Therefore, in the technical solution of the present application, it is expected to perform multi-scale perception enhancement on the character's image features so that the character's complex behavior characteristics can be fully expressed. Specifically, each character image feature map in the set of character image feature maps passes through a feature multi-scale perception enhancement module to obtain a set of enhanced character image feature maps. In particular, in a specific example of the present application, such as Figure 4 As shown, the S22 includes: S221, inputting the character image feature map into the low-frequency detection branch of the feature multi-scale perception enhancement module to obtain the character image local low-frequency activation feature vector; S222, inputting the character image feature map into the simplified compression branch of the feature multi-scale perception enhancement module to obtain the character image compression feature map; S223, inputting the character image feature map into the global space perception branch of the feature multi-scale perception enhancement module to obtain the character image receptive field expansion activation feature matrix; S224, performing feature superposition and integration on the character image local low-frequency activation feature vector, the character image compression feature map and the character image receptive field expansion activation feature matrix to obtain the enhanced character image feature map.
[0029] More specifically, in S221, the character image feature map is input into the low-frequency detection branch of the feature multi-scale perception enhancement module to obtain a local low-frequency activation feature vector of the character image. That is, in the technical solution of the present application, the feature multi-scale perception enhancement module uses the low-frequency detection branch to process and learn the detail information in the character image features.
[0030] In a specific example of the present application, the character image feature map is input into the low-frequency detection branch of the feature multi-scale perception enhancement module to obtain the character image local low-frequency activation feature vector, including: first, the character image feature map is subjected to point convolution processing to obtain the character image channel compression feature map. It should be understood that by applying point convolution processing, the number of channels in the character image feature map can be compressed to obtain the character image channel compression feature map. Next, the character image channel compression feature map is subjected to global mean pooling to obtain the character image local low-frequency feature vector. Among them, global mean pooling is a pooling operation for summarizing the features of the entire feature map and calculating the average value of all pixel positions in the feature map to obtain a global feature vector. By applying global mean pooling, the spatial information in the character image channel compression feature map can be integrated to obtain the character image local low-frequency feature vector. Finally, the character image local low-frequency feature vector is nonlinearly activated to obtain the character image local low-frequency activation feature vector. That is, by applying a nonlinear activation function, the character image local low-frequency feature vector is nonlinearly transformed to obtain the character image local low-frequency activation feature vector.
[0031] More specifically, in S222, the character image feature map is input into the simplified compression branch of the feature multi-scale perception enhancement module to obtain a compressed character image feature map. That is, in the technical solution of the present application, the simplified compression branch is used to reduce the number of channels of feature distribution to reduce the overhead of computing resources; at the same time, the global space perception branch is used to capture the overall spatial distribution pattern of character image features. More specifically, the global space perception branch uses a hole convolution coding method to expand the perception range of the original feature. Then, after the feature distribution of the simplified compression branch is superimposed with the other two branches, the perception range is expanded again by the hole convolution set at the tail, and finally the integrated features with enhanced details and overall features are obtained to form a multi-scale feature representation of the character.
[0032] More specifically, in S223, the character image feature map is input into the global spatial perception branch of the feature multi-scale perception enhancement module to obtain a character image receptive field expansion activation feature matrix. In a specific example of the present application, first, the character image feature map is subjected to hole convolution encoding to obtain a character image receptive field expansion feature map. Specifically, hole convolution is a convolution operation that can increase the receptive field of the convolution kernel without increasing parameters. By introducing holes (or expansion rates) in the convolution kernel, the convolution kernel can be sampled in a jumpy manner on the input feature map, thereby expanding the receptive field. In the hole convolution, by adjusting the hole rate, different degrees of receptive field expansion can be achieved. A larger hole rate causes the convolution kernel to jump a longer distance on the input feature map, thereby capturing a wider range of contextual information. Through hole convolution encoding, a character image receptive field expansion feature map is obtained. These feature maps contain a wider range of contextual information after the hole convolution operation. Next, the character image receptive field expansion feature map is subjected to point convolution processing to obtain a character image receptive field expansion global feature matrix. Among them, point convolution is a convolution operation, which is usually used for inter-channel feature integration and transformation in convolutional neural networks. It uses a 1x1 convolution kernel to adjust the number of channels of the feature map without changing the spatial dimension of the feature map. The expanded feature map of the receptive field of the human image is input into the point convolution layer. Through point convolution processing, the features of each pixel can be integrated and transformed between channels without changing the spatial dimension of the feature map, thereby extracting global features. After point convolution processing, the global feature matrix of the expanded receptive field of the human image is obtained. This matrix contains the global feature representation after the integration processing of point convolution, which can better capture the semantic information of the overall image.
[0033] Finally, the character image receptive field expansion global feature matrix is nonlinearly activated to obtain the character image receptive field expansion activation feature matrix. It should be understood that nonlinear activation functions such as ReLU (Rectified Linear Unit), Sigmoid, Tanh, etc. are applied to the global feature matrix. These activation functions can introduce nonlinear features so that the model can learn more complex data patterns. By applying nonlinear activation functions, the character image receptive field expansion activation feature matrix is obtained. This matrix contains the feature representation after nonlinear activation processing, which contains more complex and abstract feature information.
[0034] More specifically, the S224 performs feature superposition and integration on the local low-frequency activation feature vector of the character image, the compression feature map of the character image, and the expansion activation feature matrix of the character image receptive field to obtain the enhanced character image feature map. In a specific example of the present application, first, the character image receptive field expansion activation feature matrix is multiplied by position with each feature matrix along the channel dimension of the character image compression feature map to obtain the character image compression-expansion activation feature map. The character image receptive field expansion activation feature matrix is multiplied by position with each feature matrix along the channel dimension of the character image compression feature map. That is, the features at the corresponding positions are multiplied element by element, so as to realize the fusion and interaction of the features, and the character image compression-expansion activation feature map is obtained by the position point multiplication operation. This feature map contains the interaction information between the compression feature map and the expansion activation feature matrix, and integrates features of different scales and different abstract levels. Then, each eigenvalue in the local low-frequency activation feature vector of the character image is used as a weight, and each feature matrix along the channel dimension of the character image compression feature map is multiplied by position to obtain the character image compression-low-frequency fusion feature map. It should be understood that by taking each eigenvalue in the local low-frequency activation feature vector of the character image as a weight and performing dot multiplication on each feature matrix of the character image compression feature map along the channel dimension, the weighted fusion of low-frequency features can be achieved, thereby introducing more local information and enhancing feature representation. Then, the character image compression-expansion activation feature map and the character image compression-low-frequency fusion feature map are added by position to obtain a character image multi-scale fusion activation feature map. By adding the character image compression-expansion activation feature map and the character image compression-low-frequency fusion feature map by position, the fusion of multi-scale features can be achieved, combining feature information of different scales and different abstract levels, thereby improving the representation ability of the model when processing character image data. Finally, the second channel compression multi-scale fusion activation feature map is subjected to hole convolution coding to obtain the enhanced character image feature map. Among them, the second channel compression multi-scale fusion activation feature map is subjected to hole convolution coding to further enhance the character image feature map, and the model's ability to capture global image information is improved by introducing the nonlinear operation of hole convolution and a larger receptive field, thereby better representing the character image data.
[0035] In summary, in the above embodiment, each character image feature map in the set of character image feature maps is respectively subjected to a feature multi-scale perception enhancement module to obtain a set of enhanced character image feature maps, including: using the feature multi-scale perception enhancement module to perform feature multi-scale perception enhancement on each character image feature map in the set of character image feature maps using the following formula to obtain a set of enhanced character image feature maps; wherein the formula is:
[0036]
[0037] in, It is the character image feature map. represents 1×1 point convolution, and They represent 3×3 dilated convolutions with dilation rates of 2 and 3, respectively. represents the sigmoid activation function, represents global mean pooling, and denote dot product and positional addition respectively, It is an enhanced character image feature map.
[0038] Specifically, in other specific examples of the present application, each character image feature map in the set of character image feature maps can also be passed through a feature multi-scale perception enhancement module in other ways to obtain the set of enhanced character image feature maps, for example: input the set of character image feature maps; for each character image feature map, input it into a feature multi-scale perception enhancement module to extract multi-scale features from a single character image feature map and enhance these features. Among them, the feature multi-scale perception enhancement module includes: a multi-scale convolution layer: for capturing features at different scales; an attention mechanism: for learning the importance weights between features; a fusion layer: for fusing feature representations of different scales. Through the multi-scale convolution layer and attention mechanism in the module, feature representations of different scales are extracted and enhanced; after applying the feature multi-scale perception enhancement module to each character image feature map, an enhanced character image feature map is obtained to obtain the set of enhanced character image feature maps.
[0039] Specifically, in other specific examples of the present application, the set of character image data may be subjected to character image feature extraction in other ways to obtain a set of enhanced character image feature maps, for example: inputting the set of character image data; selecting a feature extraction model suitable for processing character images, such as a convolutional neural network (CNN) pre-trained model (such as ResNet, VGG, etc.) or a custom feature extraction network. These models are usually pre-trained on large-scale data sets and can extract rich image features; extracting features from each character image. This includes inputting each image into the selected feature extraction model and extracting high-level semantic features from the image. These features can be the activation values of the middle layer of the convolutional neural network or the output of the fully connected layer; for each character image, the obtained feature can be a vector or a matrix; if the data set contains multiple character images, the features extracted from each character image can be fused. This can be achieved by simply connecting feature vectors or performing more complex fusion operations (such as weighted averaging); after the features of each character image are fused, the set of enhanced character image feature maps is obtained.
[0040] In particular, the S3 constructs a feature distribution pseudo-topic based on the set of enhanced character image feature maps. That is, the feature distribution pseudo-topic is obtained by calculating the mean by position to summarize the general features of the character images in the entire data set, forming a representative feature distribution pseudo-topic. Specifically, the feature distribution pseudo-topic is an abstract representation of the character image features in the data set, which represents the commonality of the character features in the data set and helps to identify and understand the general characteristics of the crowd.
[0041] In particular, the S4, respectively calculates the semantic association between each enhanced character image feature map in the set of enhanced character image feature maps and the feature distribution pseudo-topic to obtain a set of semantic associations. That is, in the technical solution of the present application, the semantic association between each enhanced character image feature map in the set of enhanced character image feature maps and the feature distribution pseudo-topic is respectively calculated to obtain a set of semantic associations. By calculating the semantic association, the similarity between the individual behavior characteristics represented by each enhanced character image feature map and the general characteristics of the group represented by the feature distribution pseudo-topic can be quantified, thereby achieving personalized matching and providing an important data basis for cluster analysis. In particular, in a specific example, after calculating the positional subtraction of the enhanced character image feature map and the feature distribution pseudo-topic, a norm of the feature map obtained is calculated as the semantic association, including: after calculating the positional subtraction of the enhanced character image feature map and the feature distribution pseudo-topic according to the following formula, a norm of the feature map obtained is calculated as the semantic association; wherein, the formula is:
[0042]
[0043] in, It is the first in the set of enhanced character image feature maps. An enhanced character image feature map, is the feature distribution pseudo-topic, represents the one-norm of the feature map, It is The semantic relevance, Indicates subtraction by position.
[0044] In the technical solution of the present application, each enhanced character image feature map in the set of enhanced character image feature maps respectively represents the enhanced image semantic features of the character image. The feature distribution pseudo-topic is a position mean feature map of the set of enhanced character image feature maps, which is used to represent the statistical full-sample domain average image semantic feature density distribution of the set of character image data. Considering that there is a difference in feature distribution probability density caused by the difference in sample quantity between each enhanced character image feature map in the set of enhanced character image feature maps and the feature distribution pseudo-topic, this will cause a local mapping overflow between each enhanced character image feature map in the set of enhanced character image feature maps and the feature distribution pseudo-topic, thereby affecting the calculation accuracy of the semantic association between each enhanced character image feature map in the set of enhanced character image feature maps and the feature distribution pseudo-topic.
[0045] In a preferred embodiment, respectively calculating the semantic association between each enhanced person image feature map in the set of enhanced person image feature maps and the feature distribution pseudo-topic to obtain a set of semantic associations comprises the following steps:
[0046] The semantic association between each enhanced character image feature map and the feature distribution pseudo-topic is calculated using the following formula, wherein the formula is:
[0047]
[0048] in, and are the enhanced character image feature map and the feature distribution pseudo-topic, respectively. is the mean feature map of the enhanced character image feature map and the feature distribution pseudo-theme, and The feature maps representing the mean feature map Norms and feature maps norm, is the scale of the enhanced character image feature map and the feature distribution pseudo-theme, i.e., width multiplied by height multiplied by the number of channels, and and is the weighted sum weight as a hyperparameter, It means point multiplication by position. It means adding by position. It means subtracting by position. It means calculating the inverse of all eigenvalues in the feature map. It means to sum all the eigenvalues of the calculated feature map. represents the feature map after calculation, represents the eigenvalue of the calculated feature map, Indicates semantic relevance.
[0049] Therefore, based on the constrained representation of the structured foreground and background distinction of the supermanifold of the enhanced person image feature map and the feature distribution pseudo-topic based on the mean feature map norm between the enhanced person image feature map and the feature distribution pseudo-topic, the feature-level key differential correspondence between the enhanced person image feature map and the feature map of the feature distribution pseudo-topic is modeled, and the global differential association relationship between the corresponding features is adjusted, so as to make positive fine-grained differential correspondence suggestions through the unbalanced proportion control between the corresponding feature values of the enhanced person image feature map and the feature distribution pseudo-topic, so as to avoid the imbalance of the feature map-level image semantic distribution correspondence proportion of the enhanced person image feature map and the feature distribution pseudo-topic by focusing on the focus, so as to improve the mapping effect of each enhanced person image feature vector relative to the common semantic association space to which the clustering features of the person image semantic features expressed by the feature distribution pseudo-topic feature vector belong, and improve the calculation accuracy of the semantic association between the feature distribution pseudo-topic and each enhanced person image feature map.
[0050] In particular, in S5, based on the set of semantic associations, cluster analysis is performed on the set of character image data to obtain multiple clustering results, and each clustering result represents a group of people. That is, in the technical solution of the present application, based on the set of semantic associations, cluster analysis is performed on the set of character image data to obtain multiple clustering results, and each clustering result represents a group of people. It should be understood that different groups of people may be interested in different advertising content. Through cluster analysis, the advertising content can be matched with the characteristics of a specific group of people to achieve personalized advertising delivery.
[0051] In summary, the data analysis and processing method for advertising delivery according to the embodiment of the present application is explained, which collects image data of passing people by using monitoring equipment installed in the bus station, and performs cluster analysis on the image data of people in combination with a deep learning algorithm to classify people with similar characteristics into one category. In this way, clustering and dividing the passing people at the monitoring location can provide technical support for advertising audience analysis and advertising delivery, thereby promoting the rational allocation of advertising resources.
[0052] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.
Claims
1. A data analysis and processing method for advertising delivery, characterized in that: include: Acquire a set of person image data collected by a monitoring device, wherein the monitoring device is set in a bus stop; Extracting character image features from the character image data set to obtain a set of enhanced character image feature maps; Based on the set of enhanced character image feature maps, constructing a feature distribution pseudo-topic; Respectively calculating the semantic relevance between each enhanced person image feature map in the set of enhanced person image feature maps and the feature distribution pseudo-topic to obtain a set of semantic relevance; Based on the set of semantic associations, cluster analysis is performed on the set of person image data to obtain a plurality of clustering results, each of which represents a class of people; Calculating the semantic association between each enhanced person image feature map in the set of enhanced person image feature maps and the feature distribution pseudo-topic to obtain a set of semantic associations comprises the following steps: The semantic association between each enhanced character image feature map and the feature distribution pseudo-topic is calculated using the following formula, wherein the formula is: in, and are the enhanced character image feature map and the feature distribution pseudo-topic, respectively. is the mean feature map of the enhanced character image feature map and the feature distribution pseudo-theme, and The feature maps representing the mean feature map Norms and feature maps norm, is the scale of the enhanced character image feature map and the feature distribution pseudo-theme, i.e., width multiplied by height multiplied by the number of channels, and and is the weighted sum weight as a hyperparameter, It means point multiplication by position. It means adding by position. It means subtracting by position. It means calculating the inverse of all eigenvalues in the feature map. It means to sum all the eigenvalues of the calculated feature map. represents the feature map after calculation, represents the eigenvalue of the calculated feature map, Indicates semantic relevance.
2. The data analysis and processing method for advertising delivery according to claim 1, characterized in that: Extracting character image features from the character image data set to obtain a set of enhanced character image feature maps, including: Each character image data in the set of character image data is respectively passed through a character image feature extractor based on a dilated convolutional neural network model to obtain a set of character image feature maps; Each character image feature map in the set of character image feature maps is respectively passed through a feature multi-scale perception enhancement module to obtain the set of enhanced character image feature maps.
3. The data analysis and processing method for advertising delivery according to claim 2, characterized in that: The method of passing each character image feature map in the set of character image feature maps through a feature multi-scale perception enhancement module to obtain the set of enhanced character image feature maps comprises: Inputting the character image feature map into the low-frequency detection branch of the feature multi-scale perception enhancement module to obtain a local low-frequency activation feature vector of the character image; Inputting the character image feature map into the simplified compression branch of the feature multi-scale perception enhancement module to obtain a character image compression feature map; Inputting the character image feature map into the global space perception branch of the feature multi-scale perception enhancement module to obtain a character image receptive field expansion activation feature matrix; The local low-frequency activation feature vector of the character image, the character image compression feature map and the character image receptive field expansion activation feature matrix are subjected to feature superposition and integration to obtain the enhanced character image feature map.
4. The data analysis and processing method for advertisement delivery according to claim 3, characterized in that: Inputting the character image feature map into the low-frequency detection branch of the feature multi-scale perception enhancement module to obtain a local low-frequency activation feature vector of the character image, including: Performing point convolution processing on the character image feature map to obtain a character image channel compression feature map; Performing global mean pooling on the compressed feature map of the character image channel to obtain a local low-frequency feature vector of the character image; Nonlinear activation is performed on the local low-frequency feature vector of the character image to obtain the local low-frequency activated feature vector of the character image.
5. The data analysis and processing method for advertisement delivery according to claim 4, characterized in that: Inputting the character image feature map into the simplified compression branch of the feature multi-scale perception enhancement module to obtain a character image compression feature map, including: Performing point convolution processing on the character image feature map to obtain the character image compression feature map.
6. The data analysis and processing method for advertisement delivery according to claim 5, characterized in that: Inputting the character image feature map into the global space perception branch of the feature multi-scale perception enhancement module to obtain a character image receptive field expansion activation feature matrix, including: Performing dilated convolution coding on the character image feature map to obtain a character image receptive field expansion feature map; Performing point convolution processing on the character image receptive field expansion feature map to obtain a character image receptive field expansion global feature matrix; The character image receptive field expansion global feature matrix is nonlinearly activated to obtain the character image receptive field expansion activation feature matrix.
7. The data analysis and processing method for advertisement delivery according to claim 6, characterized in that: The local low-frequency activation feature vector of the character image, the character image compression feature map and the character image receptive field expansion activation feature matrix are superimposed and integrated to obtain the enhanced character image feature map, including: Multiplying the character image receptive field expansion activation feature matrix by each feature matrix along the channel dimension of the character image compression feature map by position points to obtain a character image compression-expansion activation feature map; Taking each eigenvalue in the local low-frequency activation feature vector of the character image as a weight, performing point multiplication on each feature matrix along the channel dimension of the character image compression feature map to obtain a character image compression-low-frequency fusion feature map; Adding the character image compression-expansion activation feature map and the character image compression-low-frequency fusion feature map according to position to obtain a character image multi-scale fusion activation feature map; The multi-scale fusion activation feature map of the character image is subjected to hole convolution coding to obtain the enhanced character image feature map.
8. The data analysis and processing method for advertisement delivery according to claim 7, characterized in that: Based on the set of enhanced character image feature maps, constructing a feature distribution pseudo-topic, including: The position-based mean feature map of the set of enhanced character image feature maps is calculated to obtain the feature distribution pseudo-theme.
9. The data analysis and processing method for advertisement delivery according to claim 8, characterized in that: The semantic association between each enhanced person image feature map in the set of enhanced person image feature maps and the feature distribution pseudo-topic is calculated respectively to obtain a set of semantic associations, including: After calculating the positional subtraction between the enhanced person image feature map and the feature distribution pseudo-topic, a norm of the obtained feature map is calculated as the semantic association degree.
Citation Information
Patent Citations
Character image clustering method and apparatus
CN108509994A
Data analysis processing method and system for advertising
CN117993957A
Device and method for pedestrian recognition using convolutional neural network
KR101818129B1
