Method for automatically identifying design manuscript element generation activity through deep learning

Abstract

The invention discloses a method for automatically identifying design draft elements to generate activities through deep learning, and relates to the technical field of artificial intelligence, and the method comprises the following steps: obtaining a marketing activity scene based on a pre-obtained marketing activity design drawing and user information through a customer group identification and marketing scene identification mode; identifying marketing elements in the marketing activity design drawing by using a convolutional neural network model to obtain activity component element features, and extracting position information of an activity component according to the activity component element features to obtain description type data; and based on a pre-acquired component library, acquiring an activity component template in combination with the description type data and the marketing activity scene, and performing page rendering according to the acquired activity component template to generate a marketing activity page. According to the method, the marketing activity page with complete interaction is automatically generated through the convolutional neural network model, so that the development efficiency of the marketing activity page and the customer satisfaction are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence. Specifically, it relates to a method for automatically identifying design draft elements through deep learning to generate activities. Background Art

[0002] In terms of digitalization, the popularization of the Internet has made online marketing the mainstream; such as the use of social media platforms; big data technology can help collect and analyze a large amount of user data to achieve precise marketing and personalized recommendations; the development of mobile technology has promoted the tilt of marketing activities towards the mobile side, such as in-app marketing and location-based service (LBS) marketing; artificial intelligence and machine learning are used in marketing to predict user behavior and optimize advertising placement strategies; marketing automation tools can automate processes, such as SMS marketing automation, email marketing automation, and social media management automation, etc.; cloud computing provides strong support for marketing data storage and processing; at the same time, blockchain technology also plays a role in ensuring data security and trust.

[0003] The continuous integration and innovation of the above technologies have continuously promoted the development and transformation of marketing activities. However, existing technical solutions, which are based on deep learning to identify elements in pictures, usually generate simple user interfaces (UIs) and cannot generate activity pages with complete interactions and business logics according to different customer groups and business scenarios. Secondary development still needs to be carried out on the generated simple UI pages to improve relevant business interactions; this has greatly reduced the efficiency of developing marketing activities, thus greatly increasing the human and time costs; and then has greatly affected the adaptation rate to market changes and marketing needs.

[0004] For the problems in the related technologies, no effective solutions have been proposed yet. Summary of the Invention

[0005] In view of the problems in the related technologies, the present invention proposes a method for automatically identifying design draft elements through deep learning to generate activities, so as to overcome the above technical problems existing in the existing related technologies.

[0006] To this end, the specific technical solution adopted by the present invention is as follows:

[0007] A method for automatically identifying design draft elements through deep learning to generate activities, the method comprising the following steps:

[0008] S1. Based on the pre-acquired marketing activity design drawings and user information, and through the methods of customer group identification and marketing scenario identification, obtain the marketing activity scenario;

[0009] S2. Use a convolutional neural network model to identify the marketing elements in the marketing campaign design diagram, obtain the feature of the activity component elements, and extract the position information of the activity components based on the feature of the activity component elements to obtain descriptive data;

[0010] S3. Based on the pre-obtained component library, combine the descriptive data and the marketing campaign scenario to obtain the activity component template, and perform page rendering according to the obtained activity component template to generate the marketing campaign page.

[0011] Furthermore, based on the pre-obtained marketing campaign design diagram and user information, and through the customer group identification and marketing scenario identification methods, obtaining the marketing campaign scenario includes the following steps:

[0012] S11. Based on the pre-obtained user information, perform customer group segmentation through the customer group identification method to obtain customer group label data;

[0013] S12. Based on the pre-obtained marketing campaign design diagram, perform marketing scenario binding through the marketing scenario identification method to obtain the marketing scenario;

[0014] S13. Associate the customer group label data and the marketing scenario to obtain the marketing campaign scenario.

[0015] Furthermore, based on the pre-obtained user information, performing customer group segmentation through the customer group identification method to obtain customer group label data includes the following steps:

[0016] S111. Characterize the user portrait based on the pre-obtained user information, analyze the user behavior according to the user image, and extract the user features;

[0017] S112. Segment the users into customer groups according to the extracted user features to obtain the customer group segmentation result;

[0018] S113. Perform labeled management on the customer group segmentation result to obtain the customer group label data.

[0019] Furthermore, using a convolutional neural network model to identify the marketing elements in the marketing campaign design diagram, obtaining the feature of the activity component elements, and extracting the position information of the activity components based on the feature of the activity component elements to obtain descriptive data includes the following steps:

[0020] S21. Annotate the marketing campaign design diagram to obtain the marketing campaign data, and train the convolutional neural network model through the marketing campaign data to obtain the activity component element recognition model;

[0021] S22. Input the marketing campaign design diagram into the activity component element recognition model to identify the marketing elements in the marketing campaign design diagram and obtain the feature of the activity component elements;

[0022] S23. Use the object detection algorithm of the convolutional neural network model, and combine the characteristics of the activity component elements to identify the bounding box positions of the activity component elements, and obtain the position information of the extracted activity components;

[0023] S24. Extract the key element information of the marketing activity according to the characteristics and position information of the activity component elements, and convert the key element information into descriptive data through data conversion.

[0024] Further, annotate the marketing activity design drawing to obtain marketing activity data, and train the convolutional neural network model through the marketing activity data. The steps for obtaining the activity component element recognition model are as follows:

[0025] S211. Annotate the marketing activity design drawing to obtain marketing activity data, and initialize the parameters of the convolutional neural network model;

[0026] S212. Input the marketing activity data into the initialized convolutional neural network model, optimize the convolutional neural network model through the loss function, and use the gradient descent method to calculate the gradient of the loss function with respect to the parameters of the convolutional neural network model;

[0027] S213. Based on the gradient of the convolutional neural network model parameters, use the optimization algorithm to update the parameters of the convolutional neural network model to obtain the activity component element recognition model, and evaluate the performance of the activity component element recognition model through accuracy, recall rate, and precision.

[0028] Further, input the marketing activity design drawing into the activity component element recognition model to identify the marketing elements of the marketing activity design drawing. The steps for obtaining the activity component element characteristics are as follows:

[0029] S221. Input the marketing activity design drawing into the activity component element recognition model, and identify the marketing elements through convolutional operations to obtain the local characteristics of the activity component elements;

[0030] S222. Reduce the feature dimension of the local characteristics of the activity component elements through pooling operations to obtain the main characteristics of the activity component elements;

[0031] S223. According to the local characteristics and main characteristics of the activity component elements, use the fully connected operation for feature synthesis to obtain the activity component element characteristics.

[0032] Further, use the object detection algorithm of the convolutional neural network model, and combine the characteristics of the activity component elements to identify the bounding box positions of the activity component elements. The steps for obtaining the position information of the extracted activity components are as follows:

[0033] S231. According to the characteristics of the activity component elements, use the region proposal network to generate candidate regions to obtain a candidate region dataset;

[0034] S232. Convert the candidate region dataset into a feature map of a fixed size, and extract features of the fixed size through a pooling operation;

[0035] S233. Classify the fixed-size features using object classification, and adjust the bounding box position of the fixed-size features using bounding box regression to obtain the position information of the extracted active components.

[0036] Furthermore, the key element information includes the name, type, width, height, position, background, and text of the active component.

[0037] Furthermore, based on the pre-obtained component library, combine the descriptive data and the marketing activity scenario to obtain the active component template, and perform page rendering according to the obtained active component template to generate the marketing activity page, including the following steps:

[0038] S31. Parse the descriptive data to obtain the attributes of each active component, and extract the component type according to the attributes of each active component;

[0039] S32. Obtain the active component template from the pre-obtained component library according to the component type and the marketing activity scenario, and convert the active component template into active message data in combination with the descriptive data;

[0040] S33. Parse the active message data, and perform page rendering according to the parsing result to generate the marketing activity page.

[0041] Furthermore, performing page rendering according to the parsing result includes registering and searching for active components, laying out active components on the page, setting the attributes and styles of active components, and setting the interaction logic of active components.

[0042] The beneficial effects of the present invention are as follows:

[0043] 1. Through the convolutional neural network model, the present invention identifies the marketing elements in the marketing activity design drawing, and based on the pre-obtained component library, automatically generates a marketing activity page with complete interactions and suitable for the current customer group and marketing scenario, thereby greatly improving the development efficiency of the marketing activity page and customer satisfaction, further improving the adaptation rate to market changes and marketing needs, and reducing the labor and time costs.

[0044] 2. By obtaining the active component template through descriptive data and marketing activity scenarios, the present invention greatly shortens the time from design to activity planning, can quickly respond to market demands, seize marketing opportunities; also reduces the costs of manual planning and design communication, and avoids resource waste caused by repeated modification and adjustment.

[0045] 3. Through customer group identification and marketing scenario identification, the present invention can quickly adjust the marketing activity page according to different marketing activity design drawings, enhancing the adaptability to marketing scenarios and target audiences; it is also applicable to simultaneously processing multiple marketing activity design drawings and generating marketing activity pages, which helps to maintain the high efficiency and high quality of generating marketing activity pages in large-scale marketing activities. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0047] Figure 1 is a flowchart of a method for automatically identifying design draft elements through deep learning to generate activities according to an embodiment of the present invention;

[0048] Figure 2 is an architecture diagram of a method for automatically identifying design draft elements through deep learning to generate activities according to an embodiment of the present invention;

[0049] Figure 3 is a flowchart of generating activity message data in a method for automatically identifying design draft elements through deep learning to generate activities according to an embodiment of the present invention;

[0050] Figure 4 is a flowchart of page rendering in a method for automatically identifying design draft elements through deep learning to generate activities according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] To further illustrate the embodiments, the present invention provides drawings. These drawings are part of the disclosure of the present invention, mainly used to illustrate the embodiments, and can be combined with the relevant descriptions in the specification to explain the operation principle of the embodiments. With reference to these contents, those of ordinary skill in the art should be able to understand other possible implementation manners and the advantages of the present invention.

[0052] According to an embodiment of the present invention, a method for automatically identifying design draft elements through deep learning to generate activities is provided.

[0053] Now, the present invention will be further described in combination with the drawings and specific implementation manners. As Figure 1 shown, a method for automatically identifying design draft elements through deep learning to generate activities according to an embodiment of the present invention includes the following steps:

[0054] S1. Based on the pre-acquired marketing activity design drawings and user information, and through customer group identification and marketing scenario identification methods, obtain the marketing activity scenarios.

[0055] Specifically, based on the pre-acquired marketing activity design drawings and user information, and through customer group identification and marketing scenario identification methods, obtaining the marketing activity scenarios includes the following steps:

[0056] S11. Based on the pre-acquired user information, conduct customer group segmentation through customer group identification methods to obtain customer group label data.

[0057] Specifically, based on the pre-acquired user information, conducting customer group segmentation through customer group identification methods to obtain customer group label data includes the following steps:

[0058] S111. Characterize the user portrait based on the pre-acquired user information, analyze the user behavior according to the user image, and extract user characteristics;

[0059] S112. According to the extracted user characteristics, conduct customer group segmentation on the users to obtain the customer group segmentation result;

[0060] S113. Conduct labeling management on the customer group segmentation result to obtain customer group label data.

[0061] Specifically, customer group label data refers to the characteristic identifiers or attributes used to classify and describe user groups; these labels can be set based on various information and behaviors of users, such as age, gender, region, consumption habits, purchase frequency, purchase amount, preferred product types, etc.; by assigning different customer group labels to users, activities such as market segmentation, precision marketing, personalized recommendation, and user service can be carried out more targeted to improve operation efficiency and user satisfaction.

[0062] Specifically, the obtained customer group label data (customerGroupId) will be persistently saved to the database.

[0063] Specifically, analyzing user behavior includes analyzing the user's gender, age, occupation, investment risk tolerance, historical behavior data, and information on handling business, etc.

[0064] Specifically, according to the extracted user characteristics, conducting customer group segmentation on the users means dividing users with the same user characteristics into one customer group.

[0065] S12. Based on the pre-acquired marketing activity design drawings, conduct marketing scenario binding through marketing scenario identification methods to obtain the marketing scenario;

[0066] S13. Associate the customer group label data with the marketing scenario to obtain the marketing activity scenario.

[0067] Specifically, the customer group label data is associated with the marketing scenarios, that is, the customer group label data is matched with the marketing scenarios. Each marketing scenario has corresponding customer group label data. By comparing the consistency between the customer group label data of the current user and the customer group label data of the marketing scenario, it is determined whether the current user can participate in the marketing scenario. The users who can participate in the marketing scenario and the marketing scenario form the current marketing activity scenario. At the same time, the marketing scenario and the corresponding customer group label data are persisted in the memory for subsequent calling of generating the activity component template.

[0068] Specifically, marketing scenarios such as wealth management scenarios, loan scenarios, etc. The marketing scenario includes the name of the marketing activity, description, validity period, etc. Among them, the threshold for participating in the activity or the whitelist users participating in the activity or users with specific customer group label data can also be set through the marketing scenario configuration.

[0069] Specifically, the data format generated by the marketing scenario is as follows:

[0070]

[0071] S2. Use the convolutional neural network model to identify the marketing elements in the marketing activity design drawing, obtain the activity component element features, and extract the position information of the activity components according to the activity component element features to obtain descriptive data.

[0072] Specifically, marketing elements such as button elements, text fields, product promotion pictures, lottery component element information, etc. in the marketing activity design drawing.

[0073] Specifically, using the convolutional neural network model to identify the marketing elements in the marketing activity design drawing, obtaining the activity component element features, and extracting the position information of the activity components according to the activity component element features to obtain descriptive data includes the following steps:

[0074] S21. Annotate the marketing activity design drawing to obtain marketing activity data, and train the convolutional neural network model with the marketing activity data to obtain an activity component element recognition model.

[0075] Specifically, annotating the marketing activity design drawing to obtain marketing activity data, and training the convolutional neural network model with the marketing activity data to obtain an activity component element recognition model includes the following steps:

[0076] S211. Annotate the marketing activity design drawing to obtain marketing activity data, and initialize the parameters of the convolutional neural network model;

[0077] S212. Input the marketing activity data into the initialized convolutional neural network model, optimize the convolutional neural network model through the loss function, and use the gradient descent method to calculate the gradient of the loss function with respect to the parameters of the convolutional neural network model.

[0078] S213. Based on the gradient of the parameters of the convolutional neural network model, use the optimization algorithm to update the parameters of the convolutional neural network model to obtain the activity component element recognition model, and evaluate the performance of the activity component element recognition model through accuracy, recall rate, and precision.

[0079] Specifically, the convolutional neural network model mainly consists of a convolutional layer, a pooling layer, and a fully connected layer.

[0080] Specifically, annotate the marketing activity design drawing to obtain the marketing activity data, that is, mark the features related to the marketing activity, such as text fields, buttons, tables (TAB), product promotion pictures, etc., and divide the marketing activity data into a training set, a validation set, and a test set.

[0081] Specifically, initialize the parameters of the convolutional neural network model, that is, determine the number of convolutional layers, the size and number of convolutional kernels, the type and size of the pooling layer, the number of nodes in the fully connected layer, etc.

[0082] Specifically, input the marketing activity data into the initialized convolutional neural network model, that is, the marketing activity data passes through the convolutional layer, the pooling layer, and the fully connected layer to calculate the output value of each neuron; for the convolutional layer, calculate the value of each position of the output feature map through the convolution formula; for the fully connected layer, the output value can be calculated through the weighted sum activation function of the neurons. For example, use the Rectified Linear Unit (ReLU) activation function. ReLU is a commonly used activation function and is widely used in neural networks.

[0083] The calculation expression of the ReLU activation function is:

[0084] f(x) = max(0, x);

[0085] In the formula, f(x) represents the ReLU activation function; x represents the input feature vector.

[0086] When the input x is less than 0, the output is 0; when the input x is greater than or equal to 0, the output is equal to the input value x. The main advantages of the ReLU activation function include simple calculation, fast calculation speed, improving the training efficiency of the convolutional neural network model; effectively alleviating the gradient vanishing problem, which helps to train deeper convolutional neural network models.

[0087] Suppose the input feature vector is x with dimension n, the number of neurons in the fully connected layer is m, the connection weight matrix is W (dimension m×n), and the bias vector is b (dimension m); calculate the linear combination as z = Wx + b, where the multiplication here is matrix multiplication and the addition is vector addition.

[0088] Generally, activation functions such as ReLU, sigmoid curve (Sigmoid), hyperbolic tangent (Tanh), etc. are used to process z to obtain the output a. For example, for the ReLU function, a = max(0, z).

[0089] Specifically, optimizing the convolutional neural network model through the loss function includes selecting an appropriate loss function, such as the cross-entropy loss function, to measure the gap between the model's predicted value and the true label.

[0090] For binary classification problems, the calculation expression of the cross-entropy loss function is:

[0091]

[0092] In the formula, L represents the loss value, that is, the difference measure between the prediction result of the convolutional neural network model and the true result; N represents the number of samples; i represents the index of the sample, traversing all samples from 1 to N; y i represents the true label of the i-th sample, taking values of 0 or 1, that is, belonging to class 0 or class 1; represents the predicted probability of the convolutional neural network model for the i-th sample, that is, the probability of predicting as class 1.

[0093] When y i = 1, the first term in the loss function takes effect, and at this time the second term is 0; the cross-entropy hopes that the probability predicted by the convolutional neural network model is as close to 1 as possible, because increases as approaches 1.

[0094] When y i = 0, the second term in the loss function takes effect, and at this time the first term is 0; the cross-entropy hopes that the probability predicted by the model is as close to 0 as possible, because increases as approaches 0.

[0095] Summing and averaging these two losses for all samples gives the average loss L over the entire dataset, which is used to measure the performance of the model in binary classification tasks. The smaller the loss, the better the performance of the model.

[0096] Specifically, the gradient of the loss function with respect to the parameters of the convolutional neural network model is calculated using the gradient descent method, that is, backpropagation. The gradient of the loss function with respect to the network parameters (such as convolutional kernel weights, fully connected layer weights, etc.) is calculated, and the parameters of the model are continuously adjusted through an optimization algorithm (such as gradient descent) to minimize this loss function.

[0097] The functions of backpropagation include calculating the output error, that is, calculating the error of the output layer according to the loss function; calculating the gradient, that is, calculating the gradient with respect to the weight W and bias b; and weight update, that is, using an optimization algorithm (such as stochastic gradient descent) to update the weight W and bias b according to the calculated gradient.

[0098] In practical applications, fully connected layers are usually used in combination with convolutional layers, pooling layers, etc. to form a deep neural network to jointly complete complex learning tasks.

[0099] Specifically, based on the gradient of the convolutional neural network model parameters, the parameters of the convolutional neural network model are updated using an optimization algorithm, that is, updating the parameters of the convolutional neural network model. The network parameters are updated using an optimization algorithm (such as Stochastic Gradient Descent (SGD), Adagrad, Adadelta, etc.) according to the calculated gradient. For example, in SGD, the parameter update formula can be expressed as:

[0100]

[0101] where θ t represents the parameter at the t-th iteration (such as the weights and biases in a neural network); θ t+1 represents the updated parameter, that is, the parameter at the (t + 1)-th iteration; η represents the learning rate, which can control the step size of parameter update. An overly large learning rate may cause the algorithm to not converge, while an overly small one may make the convergence speed too slow. represents the gradient of the loss function J with respect to the parameter θ t .

[0102] Training is repeated and updated through multiple iterations until the performance of the model on the validation set reaches a satisfactory level, and an active component element recognition model is obtained.

[0103] Specifically, the performance of the active component element recognition model is evaluated through accuracy, recall rate, and precision rate, that is, evaluating indicators such as the accuracy, recall rate, and F1 value of the active component element recognition model on the test set to determine the performance of the model.

[0104] The steps to use Region of Interest (RoI) pooling to evaluate indicators such as the accuracy, recall rate, and F1 value of the model are as follows:

[0105] First, prediction result generation, that is, for a given test data set, the prediction results of each sample are obtained through the model, including the predicted region of interest (RoI) and the corresponding class label.

[0106] Second, ground truth annotation, that is, there are already ground truth annotations for the test data set, including the true RoI and class label.

[0107] Third, matching predictions with ground truth, that is, matching the predicted RoI with the true RoI to determine which are true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN).

[0108] Calculate accuracy, Accuracy = (number of correctly predicted samples) / (total number of samples), that is, accuracy = (TP + TN) / (TP + TN + FP + FN).

[0109] Calculate recall, Recall = (number of samples correctly predicted as positive) / (number of actual positive samples), that is, recall = TP / (TP + FN).

[0110] Calculate precision, Precision = (number of samples correctly predicted as positive) / (number of samples predicted as positive), that is, precision = TP / (TP + FP).

[0111] Calculate the F1 value, F1 = 2 * (Precision * Recall) / (Precision + Recall), that is, the F1 value = 2 * (precision * recall) / (precision + recall).

[0112] In actual calculations, according to the specific tasks and data characteristics, appropriate definitions and processing should be carried out for the matching of RoI and class judgment to accurately calculate these metrics.

[0113] S22. Input the marketing activity design diagram into the activity component element recognition model to recognize the marketing elements in the marketing activity design diagram and obtain the activity component element features.

[0114] Specifically, inputting the marketing activity design diagram into the activity component element recognition model to recognize the marketing elements in the marketing activity design diagram and obtain the activity component element features includes the following steps:

[0115] S221. Input the marketing activity design diagram into the activity component element recognition model and recognize the marketing elements through convolutional operations to obtain the local features of the activity component elements;

[0116] S222. Reduce the feature dimension of the local features of the active component elements through pooling operations to obtain the main features of the active component elements;

[0117] S223. Based on the local features and main features of the active component elements, use fully connected operations for feature synthesis to obtain the features of the active component elements.

[0118] Specifically, the convolution operation is to slide the convolution kernel on the input marketing activity design diagram to extract the local features of the active component elements.

[0119] For the input marketing activity design diagram X and the convolution kernel K, the convolution operation can be expressed as:

[0120] N = X * K;

[0121] In the formula, * represents the convolution operation; N represents the result of the convolution operation.

[0122] Specifically, N is the feature map obtained after the convolution operation of the input marketing activity design diagram X and the convolution kernel K.

[0123] Each element in the feature map is obtained by performing convolution calculations at the corresponding positions on the input data by the convolution kernel; these feature maps contain the feature information extracted and filtered from the input data, providing a basis for subsequent processing and classification tasks.

[0124] For example, if the input X is a marketing activity design diagram, then N is the image feature representation obtained after being processed by a specific convolution kernel.

[0125] Specifically, the pooling operation is to reduce the dimension of the feature map, reduce the amount of calculation, and at the same time retain the main features of the active component elements; the commonly used ones are max pooling and average pooling.

[0126] Assume that the size of the input feature map is H×W (height is H, width is W), the pooling window size is kH×kW (pooling window height is kH, width is kW), and the stride is sH and sW (stride in the height and width directions).

[0127] Max pooling can be expressed as:

[0128]

[0129] In the formula, Y represents the output after the pooling operation, which is the result obtained after performing max pooling or average pooling calculations on the input feature map, representing the feature values after pooling; Y ijDenote the element value at the position of the \(i\)-th row and \(j\)-th column in the output feature map; \(i\) represents the row index of the output feature map; \(j\) represents the column index of the output feature map; \(m\) and \(n\) represent the indices for traversing the element positions in the corresponding pooling window area \(R\) in the input feature map ij in the input feature map; max represents the operation of taking the maximum value; \(X\) m，n denotes the element value at the position of the \(m\)-th row and \(n\)-th column in the input feature map.

[0130] The pooling operation reduces the dimension of the feature map. The new feature map composed of \(Y\) reduces the data volume and computational complexity while retaining the main features, which is helpful for subsequent processing and analysis.

[0131] The meaning of the max-pooling formula is that the element value \(Y\) at the position \((i, j)\) in the output feature map ij is the maximum value among all the element values in the corresponding pooling window area \(R\) ij in the input feature map.

[0132] Average pooling can be expressed as:

[0133]

[0134] In the formula, \(Q\) ij denotes the element value at the position of the \(i\)-th row and \(j\)-th column in the output feature map.

[0135] The meaning of the average-pooling formula is to first sum all the element values \(X\) ij in the corresponding pooling window area \(R\) m，n in the input feature map, and then divide the sum result by the total number of elements in the pooling window area (i.e., \(kH\times kW\)). The obtained result is the element value \(Y\) at the position \((i, j)\) in the output feature map ij .

[0136] Specifically, the specific steps of the fully connected operation include data preparation, that is, receiving the input feature vector \(x\) from the previous layer (usually a convolutional operation or a pooling operation); weight initialization, that is, the fully connected layer has a weight matrix \(W\) and a bias vector \(b\), and the parameters are randomly initialized before the start of training; linear transformation, that is, calculating the linear combination \(z = Wx + b\); here, \(Wx\) is matrix multiplication, that is, multiplying the input feature vector by the weight matrix and then adding the bias vector; activation function application, that is, inputting the result \(z\) of the linear transformation into the activation function to obtain the output \(a\); common activation functions such as ReLU, Sigmoid, Tanh, etc.; output transmission, that is, transmitting the activated output \(a\) to the next layer or as the final output (if this is the last layer of the network); during the training process, the gradient is calculated according to the loss function, and the weights \(W\) and biases \(b\) of the fully connected layer are updated through the backpropagation algorithm to optimize the performance of the model.

[0137] S23. Use the object detection algorithm of the convolutional neural network model, and combine the characteristics of the active component elements to identify the bounding box position of the active component elements, so as to obtain the position information of the extracted active components.

[0138] Specifically, using the object detection algorithm of the convolutional neural network model and combining the characteristics of the active component elements to identify the bounding box position of the active component elements, the steps for obtaining the position information of the extracted active components include the following:

[0139] S231. According to the characteristics of the active component elements, use the region proposal network to generate candidate regions to obtain a candidate region data set;

[0140] S232. Convert the candidate region data set into a feature map set of a fixed size, and extract the fixed-size features through a pooling operation;

[0141] S233. Use object classification to classify the fixed-size features, and use bounding box regression to adjust the bounding box position of the fixed-size features to obtain the position information of the extracted active components.

[0142] Specifically, using the object detection algorithm of the convolutional neural network model and combining the characteristics of the active component elements to identify the bounding box position of the active component elements, the position information of the extracted active components is obtained, that is, by using the output result of the CNN to infer or approximate the position information of the elements.

[0143] The object detection algorithm, that is, the object detection algorithm based on CNN, such as the faster region-based convolutional neural network (Faster R-CNN), YOLO, etc., can not only identify the object categories in the picture, but also give the approximate bounding box position of the object, so as to indirectly obtain the position range of the object in the image.

[0144] Among them, Faster R-CNN includes a region proposal network (Region Proposal Network, RPN), whose function is to generate candidate regions that may contain objects; the specific calculation is to use a convolutional neural network to slide a window on the feature map, and predict the existence probability and position offset of multiple candidate boxes (anchors) at each window position to obtain a candidate region data set.

[0145] For each anchor, two values are predicted, that is, the probability of containing an object (foreground / background) and the position offset (dx, dy, dw, dh), which are used to correct the position and size of the anchor.

[0146] Region of Interest (RoI) pooling, that is, converting candidate regions of different sizes into a feature map of a fixed size; the specific calculation is to extract the corresponding region from the feature map according to the position and size of the candidate region, and obtain the feature of a fixed size through a pooling operation.

[0147] Implementation steps for extracting the corresponding region from the feature map according to the position and size of the candidate region:

[0148] First, determine the coordinates of the candidate region on the feature map. That is, assume that the upper left corner coordinates of the candidate region are (x1, y1), the lower right corner coordinates are (x2, y2), and the size of the feature map is H×W (the height is H and the width is W).

[0149] Second, extract the corresponding region, that is, intercept the part from (y1, x1) to (y2, x2) from the feature map to obtain the feature map block corresponding to the candidate region.

[0150] For the calculation of obtaining a feature of a fixed size (assumed to be n×n) through a pooling operation, taking max pooling as an example, the formula can be expressed as:

[0151]

[0152] In the formula, F ij represents the value at the i-th row and j-th column in the fixed-size feature; w represents the width of the extracted candidate region; h represents the height of the extracted candidate region; f(x, y) is the value at the coordinate (x, y) in the candidate region feature map.

[0153] The meaning of the pooling operation formula is that for each position (i, j) in the fixed-size feature, the maximum value in the sub-region corresponding to the candidate region (determined by i and j) is taken as the feature value at this position.

[0154] Taking average pooling as an example, the formula can be expressed as:

[0155]

[0156] The meaning is similar, except that here the average value of the feature values in the sub-region is calculated.

[0157] Third, classification and regression, that is, performing object classification and position refinement on the candidate region; using a fully connected layer to classify the RoI feature to obtain the object category probability; at the same time, predicting the position offsets (dx, dy, dw, dh) again to refine the position and size of the candidate region.

[0158] The calculation formula for the position offset is usually based on the following form:

[0159]

[0160] Among them, (x a , y a ) is the center coordinate of the anchor, w a and h a are the width and height of the anchor, and (dx, dy, dw, dh) are the predicted offsets.

[0161] According to actual requirements, data augmentation, regularization, hyperparameter tuning, etc. can be selected to improve the performance and generalization ability of the model.

[0162] S24. Extract the key element information of the marketing activity according to the characteristics and location information of the activity component elements, and convert the key element information into descriptive data through data conversion.

[0163] Specifically, converting the key element information into descriptive data, that is, the data format for generating the descriptive key element summary data is as follows:

[0164]

[0165] Specifically, the key element information includes the name, type, width, height, location, background, and text of the activity component.

[0166] Specifically, the key element information is the marketing key information contained in buttons, texts, product pictures, tables, etc.

[0167] S3. Based on the pre-acquired component library, combine the descriptive data and the marketing activity scenario to obtain the activity component template, and perform page rendering according to the obtained activity component template to generate the marketing activity page.

[0168] Specifically, performing page rendering according to the obtained activity component template is to perform rendering through a rendering engine and display the complete activity page on the terminal screen.

[0169] Specifically, based on the pre-acquired component library, combining the descriptive data and the marketing activity scenario to obtain the activity component template, and performing page rendering according to the obtained activity component template to generate the marketing activity page includes the following steps:

[0170] S31. Parse the descriptive data to obtain the attributes of each activity component, and extract the component type according to the attributes of each activity component;

[0171] S32. According to the component type and the marketing activity scenario, obtain the activity component template in the pre-acquired component library, and convert the activity component template into activity message data in combination with the descriptive data.

[0172] Specifically, a component template refers to a general model or paradigm created for a certain type of components with similar functions and structures in software development or page design; it usually includes the basic structure, style, property settings, and default behavior logic of the component; the component template provides a standardized starting point for developers or designers, enabling them to customize and modify based on the template when creating specific component instances, thereby improving development efficiency, ensuring consistency, and reducing the possibility of errors.

[0173] For example, when building a web page, the "button" component has a template that stipulates the default size, color, font style, mouse hover effect, etc. of the button. When developers need to use a button, they can create a specific button based on this template and adjust certain properties according to specific requirements, such as the text content and the operation after clicking.

[0174] Specifically, the activity message data has a unique activity identifier (activityId) and persists the data storage.

[0175] S33. Parse the activity message data and perform page rendering according to the parsing result to generate a marketing activity page.

[0176] Specifically, perform data parsing on the descriptive data, that is, read and parse the descriptive data, and extract the attributes of each activity component, such as name, type, width, height, position, background, text, etc.

[0177] Specifically, convert the activity component template into activity message data in combination with the descriptive data. First, define a unified basic data structure for the activity components, and all activity components and activity component templates follow the unified data structure.

[0178] Specifically, obtain the activity component template from the pre-acquired component library, that is, retrieve the corresponding activity component template from the corresponding component library; the activity component template has encapsulated the functional implementation related to the marketing activity scenario; only need to render the component to the corresponding position according to the position, width, and height data described in the component type; fill the data such as width (width), height (height), and position information in the design draft in the component template. The unified data structure of the activity components is as follows:

[0179]

[0180]

[0181] Among them, "attributes" represents the custom attributes in the component, which are applied to the component to achieve specific styles or functions; "action" indicates that the data in the array is the operation id supported in the current component, and the specific operation type and parameters can be queried in the "actionList" data through the id; "actionList" indicates that the data in the array is the detailed information of the operations supported in the current component; through this configuration information, after rendering and parsing, the corresponding operations can be completed.

[0182] The data structure stored in "actionList" is as follows:

[0183]

[0184]

[0185] The above configuration means a button component. There are two event handling ids, id1 and id2, in "action". The event type "actionType" of id1 is "init", that is, the trigger timing of the component. When the component is loaded and initialized, the id1 event handling is executed; when the rendering engine processes the id1 event, it further determines what specific actions need to be performed according to the type of "type"; for example, when "type" = 7, it prompts the user "Data is loading"; the rendering engine then takes the content of "content" as the prompt message and displays the specific prompt message.

[0186] The data format of the converted activity message data is as follows:

[0187]

[0188] Specifically, page rendering according to the parsing result includes the registration and search of activity components, the layout of activity components in the page, the setting of attributes and styles of activity components, and the setting of interaction logic of activity components.

[0189] Specifically, parsing the activity message data means that page rendering sends a request to the server. The request contains the "activityId" of the activity to be obtained. After the server receives the request, it sends down the data; receives the JavaScript Object Notation (JSON) data sent down by the server; and converts the data into a data object, the structure of which is as follows:

[0190] [{Component 1}, {Component 2},..., {Component N}].

[0191] Specifically, the parsing result, that is, traversing the parsed data, includes:

[0192] First, for each type of component (such as button, tab, text, input, etc.), obtain the corresponding component template from the corresponding component library.

[0193] Second, set the style and position of the component in the page according to the extracted attributes (width, height, position, etc.).

[0194] Third, if the component has text content (such as the Txt attribute), fill it into the component.

[0195] Fourth, if the component has a background color or image (such as the Background, Sr attributes), make the corresponding settings.

[0196] Specifically, the registration and lookup of active components, that is, the components of various component libraries registered during the initialization of the rendering engine, and find the component template corresponding to the scenario according to the values of the business scenario id and the componentType field.

[0197] For example, if "componentType" is "COMPONT_BUTTON", that is, a button component, obtain the relevant information and functions of the button component registered in the corresponding component library.

[0198] Specifically, layout the active components in the page, that is, extract attributes such as top, left, width, and height from the rectangle object to determine the position and size of the component in the page; set the rotation angle of the component according to the rotate attribute; process the zIndex attribute to determine the stacking order of the components; determine whether the component is fixed at a certain position in the page according to the fixed attribute.

[0199] Specifically, set the attributes and styles of the active components, that is, set the optional animation effects; if the animate object exists and has relevant attributes, obtain the animation name from the name field; use duration to set the duration of the animation, delay to set the delay time, and repeat to set the number of repetitions; border settings, that is, obtain attributes such as style, color, and width from the border object, and set the border style, color, and width of the component; set the rounded corner effect of the border according to the radius attributes of the four corners; apply the custom attributes in the attributes object to the component to achieve specific styles or functions.

[0200] Specifically, set the interaction logic of the active components. If it is defined in the action array or the actionList object, set the corresponding actions and interaction logic for the components, such as click events, mouse hover effects, etc.

[0201] The process of executing actions and interaction functions is to configure the action of the component and the message data in the actionList and load them into the currently active button component. For example, when parsing the "attributes" configured in the component template, which contains a custom attribute named "buttonLable" with a value of "Execute by clicking the button", when displaying the button, the lable content of the button component is set to "Execute by clicking the button". When parsing that the "action" array lists two event handling ids, namely "licai_001" and "daikuan_002", the configuration data corresponding to the ids in the "actionList" is retrieved according to the ids.

[0202] The "actionList" object details the specific information of these two events: one is that the trigger timing of the "licai_001" event is "click", and the business scenario type is 1. Assuming type: 1, it means that when the user clicks the button, a network request is sent. Then, relevant configuration parameters such as "apiId" and "param" are further obtained. When the user clicks, the network request interface of the button component is called, and the parsed "apiId" and "param" information is passed in. The apiId uniquely corresponds to the interface of the current business scenario of the service system. After the service system receives the data, it processes it and returns the processing result. The button component completes the processing of the data returned by the server and displays it on the page. The other is the "licai_002" event, where the actionType is "doubleClick", which means that double-clicking the button executes the relevant event. Similarly, the "ApiId", "params", etc. are parsed and the component method is called to complete the data processing.

[0203] An example of the active message data in the execution process is as follows:

[0204]

[0205]

[0206] Note that attributes are the data associated with the current component and can store any data associated with the component. Action stores the operation ids corresponding to the marketing scenarios supported by the current component. There are corresponding specific operations in the actionList. The actionList is the action of the current component corresponding to the marketing scenario, such as sending an interface request, etc. According to the above configuration information, when the component is rendered, the operations of the relevant scenarios are executed.

[0207] The complete example message is as follows:

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214]

[0215]

[0216] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for automatically identifying design draft element generation activities through deep learning, characterized in that: The method comprises the following steps: S1. Based on the pre-acquired marketing activity design and user information, the marketing activity scenario is obtained through customer group identification and marketing scenario identification; S2. Using the convolutional neural network model, identify the marketing elements in the marketing activity design diagram, obtain the element features of the activity components, and extract the location information of the activity components according to the element features of the activity components to obtain descriptive data; S3. Based on the pre-acquired component library, the activity component template is obtained in combination with the descriptive data and the marketing activity scenario, and the page is rendered according to the acquired activity component template to generate a marketing activity page.

2. A method for automatically identifying design draft element generation activities through deep learning according to claim 1, characterized in that: The method of obtaining the marketing activity scenario based on the pre-acquired marketing activity design diagram and user information and through customer group identification and marketing scenario identification includes the following steps: S11, based on the pre-acquired user information, divide the customer groups by customer group identification to obtain customer group label data; S12, based on the pre-acquired marketing activity design diagram, bind the marketing scenarios through a marketing scenario recognition method to obtain a marketing scenario; S13. Associate the customer group tag data with the marketing scenario to obtain the marketing activity scenario.

3. The method for automatically identifying design draft element generation activities through deep learning according to claim 2, characterized in that: The method of dividing the customer groups by customer group identification based on the pre-acquired user information to obtain customer group label data includes the following steps: S111, creating a user portrait based on the pre-acquired user information, analyzing the user behavior based on the user image, and extracting user features; S112, dividing the users into customer groups according to the extracted user features to obtain customer group division results; S113. Label management is performed on the customer group segmentation results to obtain customer group label data.

4. The method for automatically identifying design draft element generation activities through deep learning according to claim 1, characterized in that: The method of using a convolutional neural network model to identify marketing elements in a marketing activity design diagram, obtaining activity component element features, and extracting location information of the activity component according to the activity component element features to obtain descriptive data includes the following steps: S21. Annotate the marketing activity design drawing to obtain marketing activity data, train the convolutional neural network model through the marketing activity data, and obtain an activity component element recognition model; S22, inputting the marketing activity design drawing into the activity component element recognition model, identifying the marketing elements in the marketing activity design drawing, and obtaining the activity component element features; S23, using the object detection algorithm of the convolutional neural network model and combining the characteristics of the activity component elements to identify the bounding box position of the activity component elements, and obtain the position information of the extracted activity component; S24. Extract key element information of the marketing activity according to the element characteristics and position information of the activity components, and convert the key element information into descriptive data through data conversion.

5. The method for automatically identifying design draft element generation activities through deep learning according to claim 4, characterized in that: The step of annotating the marketing activity design drawing to obtain marketing activity data, training the convolutional neural network model through the marketing activity data, and obtaining the activity component element recognition model comprises the following steps: S211, annotating the marketing activity design drawing, obtaining marketing activity data, and initializing the convolutional neural network model parameters; S212, inputting the marketing activity data into the initialized convolutional neural network model, optimizing the convolutional neural network model through a loss function, and calculating the gradient of the loss function to the convolutional neural network model parameters using a gradient descent method; S213. Based on the gradient of the convolutional neural network model parameters, the parameters of the convolutional neural network model are updated using an optimization algorithm to obtain an active component element recognition model, and the performance of the active component element recognition model is evaluated by accuracy, recall rate, and precision.

6. The method for automatically identifying design draft element generation activities through deep learning according to claim 4, characterized in that: The step of inputting the marketing activity design drawing into the activity component element recognition model, identifying the marketing elements of the marketing activity design drawing, and obtaining the activity component element features comprises the following steps: S221, inputting the marketing activity design drawing into the activity component element recognition model, and identifying the marketing elements through convolution operation to obtain the local features of the activity component elements; S222, reducing the feature dimensions of local features of the activity component elements through a pooling operation to obtain main features of the activity component elements; S223. Based on the local features of the activity component elements and the main features of the activity component elements, a full connection operation is used to perform feature synthesis to obtain the features of the activity component elements.

7. The method for automatically identifying design draft element generation activities through deep learning according to claim 4, characterized in that: The object detection algorithm using the convolutional neural network model and combining the characteristics of the active component elements to identify the bounding box position of the active component elements and obtain the position information of the extracted active component includes the following steps: S231, generating candidate regions using a region proposal network according to the element features of the activity component, and obtaining a candidate region dataset; S232, converting the candidate region dataset into a fixed-size feature atlas, and extracting fixed-size features through a pooling operation; S233, classifying the fixed-size features by using object classification, and adjusting the bounding box position of the fixed-size features by using bounding box regression to obtain position information of the extracted active component.

8. The method for automatically identifying design draft element generation activities through deep learning according to claim 4, characterized in that: The key element information includes the name, type, width, height, position, background, and text of the activity component.

9. The method of automatically identifying design draft element generation activities through deep learning according to claim 1, characterized in that: The method of obtaining an activity component template based on the pre-acquired component library in combination with the descriptive data and the marketing activity scenario, and performing page rendering according to the acquired activity component template to generate a marketing activity page includes the following steps: S31, parsing the descriptive data to obtain the attributes of each activity component, and extracting the component type according to the attributes of each activity component; S32. According to the component type and the marketing activity scenario, obtain an activity component template from a pre-acquired component library, and convert the activity component template into activity message data in combination with the descriptive data; S33. Parse the activity message data, and render the page according to the parsing result to generate a marketing activity page.

10. The method for automatically identifying design draft element generation activities through deep learning according to claim 9, characterized in that: The page rendering according to the parsing result includes registering and searching for active components, laying out the active components in the page, setting the properties and styles of the active components, and setting the interaction logic of the active components.

Images