Intelligent analysis algorithm for enterprise identification data and optimization method thereof

Abstract

The invention discloses an enterprise identification data intelligent analysis algorithm and an optimization method thereof, and relates to the technical field of analysis algorithms, and the method comprises the steps: S1, selecting an OfficientNet as a basic model, and integrating a plurality of defense mechanisms for an adversarial attack; s2, carrying out lightweight processing on the model by adopting a model pruning or knowledge distillation technology; s3, collecting different styles of enterprise identification data, and carrying out cleaning processing and enhancement processing; s4, training the model by using the preprocessed data; s5, designing an online learning mechanism, and adjusting and optimizing an algorithm according to feedback and recognition results in practical application; and S6, inputting a to-be-identified enterprise identifier into the trained model, and extracting useful enterprise information. According to the enterprise identification data intelligent analysis algorithm and the optimization method thereof, continuous optimization can be carried out by occupying relatively few computing resources and time cost, and relatively high value and competitive power are achieved in practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of parsing algorithms, and specifically to an intelligent parsing algorithm for enterprise logo data and an optimization method thereof. Background Art

[0002] Enterprise logo data refers to the core elements representing the visual image of an enterprise, usually including the enterprise logo, pattern, text, etc. These elements convey the enterprise's business philosophy, cultural characteristics, scale, business content and features through specific designs. The enterprise logo is the core of the visual image, constitutes the basic characteristics of the enterprise image, reflects the internal quality of the enterprise, is the leading force mobilizing all visual elements, and is also the representative that the general public recognizes the enterprise brand.

[0003] The intelligent parsing algorithm for enterprise logo data is a technical means specifically used to identify, parse and manage enterprise logo data. This algorithm can automatically detect and identify the enterprise logo in an image, and further parse relevant enterprise information, such as enterprise name, brand, industry, etc., so as to effectively manage and store enterprise logo data for subsequent data analysis and application. At present, there are many application scenarios for this algorithm. For example, through the intelligent parsing algorithm, counterfeit and shoddy products can be quickly identified and combated to protect the enterprise's brand image and intellectual property rights. Another example is in the process of advertising placement, the algorithm can monitor the enterprise logo in the advertisement in real time to ensure the accuracy and compliance of the advertisement, and at the same time optimize the advertisement effect. For another example, in a smart city, through this algorithm, various enterprise logos in the city can be identified and managed to improve the efficiency and intelligent level of urban management.

[0004] However, in practical applications, due to the high diversity and complexity of enterprise logo data, the logo design styles of different enterprises are diverse, and elements such as shape, color, and font are rich and varied. This makes the algorithm need to process a large number of different styles of logos, thus increasing the difficulty of algorithm training and optimization. And the training and optimization of the algorithm require a large amount of computing resources and time costs. As the model complexity increases, the training time will also be correspondingly extended. Especially in the case of limited resources, this brings great inconvenience to the continuous optimization of the algorithm. If the optimization of the parsing algorithm is restricted, its recognition accuracy will gradually decline over time, which will in turn cause the algorithm to be unable to accurately identify enterprise logos in practical applications, affecting user experience and satisfaction.

[0005] Therefore, it is urgent to improve this shortcoming. The present invention studies and improves the existing technology and deficiencies, and provides an intelligent parsing algorithm for enterprise logo data and an optimization method thereof. Summary of the Invention

[0006] The object of the present invention is to provide an intelligent parsing algorithm for enterprise logo data and its optimization method to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] In the first aspect, an intelligent parsing algorithm for enterprise logo data includes the steps:

[0009] S1. Select EfficientNet with strong feature extraction ability as the basic model. This model performs well in the field of image recognition and has good generalization ability. For adversarial attacks, integrate multiple defense mechanisms to improve the robustness of the algorithm;

[0010] S2. Use model pruning or knowledge distillation technology to lightweight the model to reduce computing resources and time costs. The lightweighted model can run more efficiently while maintaining a high recognition accuracy;

[0011] S3. Collect a large amount of enterprise logo data in different styles from multiple sources to ensure the diversity and comprehensiveness of the data, and perform cleaning processing to remove noise, redundancy, and invalid data to improve the data quality. Then perform enhancement processing on the cleaned data, such as flipping, rotating, scaling, cropping, etc., to increase the diversity and richness of the data, which helps to improve the adaptability of the algorithm to complex scenarios;

[0012] S4. Use the preprocessed data to train the model, and improve the recognition accuracy by adjusting the model parameters and optimization algorithms;

[0013] S5. Design an online learning mechanism so that the algorithm can continuously receive new data and learn and update in actual applications. This helps the algorithm to adapt to the continuously changing enterprise logo styles and features. According to the feedback and recognition results in actual applications, adjust and optimize the algorithm. For example, retrain the samples with recognition errors or adjust the strategies of feature extraction and model training;

[0014] S6. Input the enterprise logo to be recognized into the trained model for feature extraction and classification recognition, and parse the recognition results to extract useful enterprise information, such as enterprise name, brand, industry, etc.

[0015] Further, in the step S1, the defense mechanisms include adversarial training, gradient regularization, defensive distillation, and ensemble learning. These defense mechanisms can be used alone or in combination to achieve the best defense effect, specifically as follows:

[0016] Adversarial training: By adding adversarial samples to the training data, the model learns how to defend against these attacks during training to improve the model's robustness to adversarial perturbations;

[0017] Gradient regularization: By penalizing the sensitivity of the model output to input changes to enhance the model's robustness. Specifically, a regularization term related to the input gradient is added to the loss function to limit the model's over-sensitivity to small changes in the input;

[0018] Defensive distillation: By using knowledge distillation technology to improve the model's robustness. Specifically, a teacher model is trained, and the output of the teacher model is used as soft labels to train the student model. This method can enable the student model to learn a smoother and more robust decision boundary;

[0019] Ensemble learning: By combining the prediction results of multiple models to improve the overall performance. Specifically, in terms of defending against adversarial attacks, the robustness is improved by integrating multiple EfficientNet models with different structures and parameters.

[0020] Furthermore, in the step S2, the model pruning lightweight processing flow is as follows:

[0021] Sparse training: By adding methods such as the L1 regularization term, the model weights are gradually sparsified during training to determine which weights are redundant;

[0022] Determine the pruning strategy: On the premise of not affecting the performance of the basic model, select a pruning method, such as structured pruning or unstructured pruning. Structured pruning usually removes the entire convolution kernel or neuron, and unstructured pruning removes individual weights. According to the structure and performance requirements of the EfficientNet model, determine the pruning granularity and ratio;

[0023] Implement pruning: Use the pre-trained EfficientNet model as the starting point, and according to the pruning strategy, remove redundant parameters layer by layer or kernel by kernel;

[0024] Model fine-tuning: Fine-tune the pruned model to recover the accuracy lost due to pruning. During the fine-tuning process, a smaller learning rate can be used, and the performance change of the model on the validation set can be monitored.

[0025] Furthermore, in the step S2, the knowledge distillation lightweight processing flow is as follows:

[0026] Prepare the teacher model and the student model: Select a trained EfficientNet model as the teacher model, and design a smaller neural network structure as the student model, whose architecture is simpler and more compact than the teacher model;

[0027] Generate soft labels: Use the teacher model to predict the training dataset and generate soft labels containing probability distributions. These soft labels will be used to train the student model so that the student model can learn the output distribution of the teacher model;

[0028] Train the student model: Input the training dataset and the corresponding soft labels into the student model for training. At the same time, use the hard labels (i.e., true labels) of the original data as supervision information to compare with the output of the student model, calculate the loss function, and optimize the loss function to make the output of the student model as close as possible to the output of the teacher model.

[0029] Evaluate and optimize: Evaluate the performance of the student model on the validation dataset to ensure that its accuracy and generalization ability meet the requirements. Based on the evaluation results, further optimize and adjust the student model to improve its performance.

[0030] Furthermore, in step S3, the sources of enterprise identification data include:

[0031] Enterprise internal database: The enterprise's own database, such as CRM systems, ERP systems, etc. These systems contain information related to enterprise identification, such as brand history, logo design drafts, usage specifications, etc. Collection method: Through the enterprise's internal data management system or IT department, use data analysis or data mining techniques to extract data related to the identification;

[0032] Government and industry institutions: Public data, reports, or guidelines released by government agencies, industry associations, chambers of commerce, etc. Collection method: Visit government websites, industry association websites, or chamber of commerce websites to find data, reports, or guidelines related to the identification. These resources may include industry trends, market sizes, market shares, etc. information, which helps to understand the performance of enterprise identification in a specific market environment;

[0033] Professional research institutions and consulting companies: Industry reports, market research reports, white papers, etc. released by market research institutions, consulting companies, etc. Collection method: Subscribe to the reports released by these institutions or purchase their consulting services. These reports usually contain information such as brand logos, market shares, and competitive situations of enterprises in the industry, which helps to understand the logo design trends and competitor situations in the industry;

[0034] Social media and online platforms: Enterprise identification data on social media platforms (such as Weibo, WeChat, Douyin, etc.), online design platforms (such as Canva, Gaoding Design, etc.), and e-commerce platforms (such as Taobao, JD.com, etc.). Collection method: Use the search function of social media platforms to find relevant enterprise accounts and collect the logo information they release. Browse and search for different styles and designs of enterprise logos on online design platforms. Search for products or services related to enterprise logos on e-commerce platforms to understand the logo design and application situations in the market;

[0035] Market research and consumer research: Data collected through market research activities, consumer research questionnaires, etc. Collection methods: Design and conduct market research activities such as online surveys, telephone interviews, face-to-face interviews, etc. to collect consumers' views and feedback on the corporate logo, and formulate consumer research questionnaires, which are sent to the target audience through online questionnaire platforms or emails to collect opinions and suggestions;

[0036] Copyright library and Creative Commons platform: Logo design materials and cases provided by copyright libraries, Creative Commons platforms, etc. Collection methods: Access these platforms, search for and browse corporate logo design materials and cases in different styles and forms;

[0037] Professional exhibitions and events: Professional exhibitions, seminars, forums and other activities related to corporate logos. Collection methods: Participate in these activities, communicate with industry professionals, and collect information such as the latest trends, design concepts and application cases of corporate logos.

[0038] Furthermore, in step S4, the model parameters are optimized by the backpropagation algorithm. During the training process, strategies such as learning rate decay and momentum optimization are adopted to improve the training efficiency;

[0039] The backpropagation algorithm calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the opposite direction of the gradient, thereby minimizing the loss. During the backpropagation process, the gradient of each layer needs to be calculated and passed to the previous layer until the parameters of all layers are updated;

[0040] The learning rate decay dynamically adjusts the learning rate according to the training progress, and as the training progresses, the learning rate gradually decreases, which helps the model to more finely adjust the parameters in the later stage of training;

[0041] The momentum optimization introduces a momentum term, which is the weighted sum of the direction of the previous gradient update and the current gradient. Through momentum optimization, the model updates the parameters faster along the correct direction and reduces the oscillation in the wrong direction.

[0042] Furthermore, in step S5, the design of the online learning mechanism includes:

[0043] Online learning framework: Select or develop a deep learning framework that supports online learning, such as TensorFlow, PyTorch, etc., and the framework supports incremental learning and continuous learning so that the algorithm can continuously receive new data and update the model;

[0044] Model Update Strategy: Design a model update strategy, including learning rate adjustment, model fine-tuning, weight update, etc. Adopt a progressive learning method, that is, when receiving new data each time, only update a part of the model's parameters to reduce the computational load and avoid overfitting;

[0045] Feedback and Evaluation: Design a feedback mechanism to collect the algorithm recognition results and user feedback in actual applications, and regularly evaluate the algorithm. Adjust the model parameters and learning strategy according to the evaluation results.

[0046] Furthermore, in step S6, the model will output one or more possible recognition results, which may include information such as the category to which the enterprise logo belongs, similarity scores, etc. According to the similarity scores or other metrics of the recognition results, screen and sort the output results to determine the most likely recognition result.

[0047] Furthermore, in step S6, the process of parsing and extracting information from the recognition results: Match the recognition results with the information in the enterprise database. This requires establishing a database containing information such as enterprise names, brands, industries, etc., and comparing the recognition results with the entries in the database. Once the recognition results match the entries in the database successfully, useful information related to the enterprise logo can be extracted from the database, such as enterprise name, brand name, industry, contact information, etc., and the extracted information is verified to ensure the accuracy and integrity of the information. If necessary, missing information can also be supplemented from other reliable sources.

[0048] In the second aspect, an optimization method for an intelligent parsing algorithm of enterprise logo data, applied to the intelligent parsing algorithm of enterprise logo data as described above, includes the following steps:

[0049] S1. Online Learning and Update:

[0050] Adopt an online learning mechanism to enable the algorithm to continuously receive new data and learn and update in actual applications to adapt to the continuously changing enterprise logo styles and features;

[0051] S2. Computational Resource Optimization:

[0052] Use technical means such as distributed computing and GPU acceleration to optimize the computational resource occupancy and time cost of the algorithm, so as to improve the running efficiency of the algorithm and reduce the demand for hardware resources;

[0053] S3. Deployment and Monitoring:

[0054] Deploy the optimized algorithm to the actual application scenario and conduct continuous monitoring. By collecting user feedback and recognition results, timely discover and solve potential problems to ensure the stability and accuracy of the algorithm.

[0055] The present invention provides an intelligent parsing algorithm for enterprise logo data and its optimization method, with the following

[0056] Beneficial effects:

[0057] The present invention selects EfficientNet as the basic model and uses model pruning or knowledge distillation technology to lightweight it, significantly reducing the computational resources and time cost of the model. At the same time, a large number of enterprise logo data of different styles are collected and processed, including steps such as data cleaning and enhancement processing, improving the usability and quality of the data. The model is trained with the preprocessed data, further enhancing the generalization ability and recognition accuracy of the model. In addition, an online learning mechanism is designed to enable the algorithm to continuously receive new data and learn and update in actual applications, so that it can adapt to changing environments and data distributions. In summary, the algorithm of the present invention can be continuously optimized with less computational resources and time cost, and has high intelligence parsing and recognition accuracy for enterprise logo data, and has high value and competitiveness in actual applications. Description of the Drawings

[0058] Figure 1 It is a step schematic diagram of an intelligent parsing algorithm for enterprise logo data of the present invention;

[0059] Figure 2 It is a step flow chart of an optimization method for an intelligent parsing algorithm for enterprise logo data of the present invention. Detailed Embodiments

[0060] The following further describes in detail the embodiments of the present invention in conjunction with the drawings and embodiments. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0061] As Figure 1 shown, an intelligent parsing algorithm for enterprise logo data includes the steps:

[0062] S1. Select EfficientNet with strong feature extraction ability as the basic model. This model performs well in the field of image recognition and has good generalization ability. For adversarial attacks, multiple defense mechanisms are integrated to improve the robustness of the algorithm. In this step, the defense mechanisms include adversarial training, gradient regularization, defense distillation, and ensemble learning. These defense mechanisms can be used alone or in combination to achieve the best defense effect, specifically as follows:

[0063] Adversarial training: By adding adversarial samples to the training data, the model learns how to resist these attacks during training to improve the model's robustness to adversarial perturbations. In this embodiment, during the training of EfficientNet, adversarial samples are generated by FGSM (Fast Gradient Sign Method) or other adversarial attack methods and added to the training set. The model learns the features of both normal samples and adversarial samples during training, thereby improving its ability to resist adversarial attacks.

[0064] Gradient regularization: Enhances the model's robustness by penalizing the sensitivity of the model output to input changes. Specifically, a regularization term related to the input gradient is added to the loss function to limit the model's over-sensitivity to small changes in the input. In this embodiment, in the loss function of EfficientNet, a regularization term related to the input gradient can be added. This regularization term penalizes the sensitivity of the model output to small changes in the input, thereby forcing the model to learn a smoother decision boundary and improving its robustness to adversarial attacks.

[0065] Defensive distillation: Improves the model's robustness through knowledge distillation technology. Specifically, a teacher model is trained, and the output of the teacher model is used as soft labels to train the student model. This method enables the student model to learn a smoother and more robust decision boundary. In this embodiment, in the defensive distillation of EfficientNet, a larger EfficientNet model can be trained first as the teacher model. Then, the output of the teacher model is used as soft labels to train a smaller EfficientNet model as the student model. In this way, the student model can inherit the robust features of the teacher model and improve its ability to resist adversarial attacks.

[0066] Ensemble learning: Improves the overall performance by combining the prediction results of multiple models. Specifically, in terms of defending against adversarial attacks, the robustness is improved by integrating multiple EfficientNet models with different structures and parameters. In this embodiment, multiple EfficientNet models with different structures and parameters are trained, and during testing, methods such as voting mechanism or weighted average are used to combine their prediction results, thereby leveraging the complementarity between different models to improve the overall robustness.

[0067] S2. The model is lightweight processed using model pruning or knowledge distillation technology to reduce computational resources and time costs. The lightweight model can operate more efficiently while maintaining a high recognition accuracy.

[0068] 1) The process of lightweight processing by model pruning is as follows:

[0069] Sparse training: By methods such as adding L1 regularization terms, the model weights are gradually sparsified during training to determine which weights are redundant;

[0070] Determine the pruning strategy: On the premise of ensuring that the performance of the basic model is not affected, select a pruning method, such as structured pruning or unstructured pruning. Structured pruning usually removes the entire convolutional kernel or neuron, and unstructured pruning removes individual weights. According to the structure and performance requirements of the EfficientNet model, determine the pruning granularity and ratio;

[0071] Implement pruning: Use the pre-trained EfficientNet model as a starting point. According to the pruning strategy, remove redundant parameters layer by layer or kernel by kernel. In this embodiment, redundant parameters with small weight values can be removed by setting a threshold, or pruning can be performed based on importance scores. During the pruning process, the indices of the removed weights can be recorded for subsequent reconstruction of the model structure;

[0072] Model fine-tuning: Fine-tune the pruned model to recover the accuracy lost due to pruning. During the fine-tuning process, a smaller learning rate can be used, and the performance changes of the model on the validation set can be monitored;

[0073] 2) The knowledge distillation lightweight processing process is as follows:

[0074] Prepare the teacher model and the student model: Select a trained EfficientNet model as the teacher model, and design a smaller neural network structure as the student model, whose architecture is simpler and more compact than the teacher model;

[0075] Generate soft labels: Use the teacher model to predict the training data set to generate soft labels containing probability distributions. These soft labels will be used to train the student model so that the student model can learn the output distribution of the teacher model;

[0076] Train the student model: Input the training data set and the corresponding soft labels into the student model for training. At the same time, use the hard labels (i.e., true labels) of the original data as supervision information to compare with the output of the student model, calculate the loss function, and optimize the loss function to make the output of the student model as close as possible to the output of the teacher model.

[0077] Evaluate and optimize: Evaluate the performance of the student model on the validation data set to ensure that its accuracy and generalization ability meet the requirements, and further optimize and adjust the student model according to the evaluation results to improve its performance.

[0078] S3. Collect a large amount of enterprise logo data in various styles from multiple sources to ensure data diversity and comprehensiveness. Then, clean the data by removing noise, redundancy, and invalid data to improve data quality. After that, perform augmentation on the cleaned data, such as flipping, rotating, scaling, and cropping, to increase data diversity and richness, which helps improve the algorithm's adaptability to complex scenarios.

[0079] In this embodiment, the sources of enterprise logo data include:

[0080] Enterprise internal database: The enterprise's own database, such as CRM systems, ERP systems, etc. These systems contain relevant information about enterprise logos, such as brand history, logo design drafts, usage specifications, etc. Collection method: Through the enterprise's internal data management system or IT department, use data analysis or data mining techniques to extract logo-related data;

[0081] Government and industry institutions: Public data, reports, or guidelines released by government agencies, industry associations, chambers of commerce, etc. Collection method: Visit government websites, industry association websites, or chamber of commerce websites to search for logo-related data, reports, or guidelines. These resources may include information such as industry trends, market size, and market share, which helps understand the performance of enterprise logos in a specific market environment;

[0082] Professional research institutions and consulting companies: Industry reports, market research reports, white papers, etc. released by market research institutions, consulting companies, etc. Collection method: Subscribe to the reports released by these institutions or purchase their consulting services. These reports usually contain information such as the brand logos of enterprises in the industry, market share, and competitive landscape, which helps understand the logo design trends and competitor situations in the industry;

[0083] Social media and online platforms: Enterprise logo data on social media platforms (such as Weibo, WeChat, Douyin, etc.), online design platforms (such as Canva, Gaoding Design, etc.), and e-commerce platforms (such as Taobao, JD.com, etc.). Collection method: Use the search function of social media platforms to search for relevant enterprise accounts and collect the logo information they release. Browse and search for different styles and designs of enterprise logos on online design platforms. Search for products or services related to enterprise logos on e-commerce platforms to understand the logo design and application in the market;

[0084] Market research and consumer research: Data collected through market research activities, consumer research questionnaires, etc. Collection method: Design and conduct market research activities, such as online surveys, telephone interviews, face-to-face interviews, etc., to collect consumers' views and feedback on enterprise logos. Also, develop consumer research questionnaires and send them to the target audience through online questionnaire platforms or emails to collect opinions and suggestions;

[0085] Copyright Library and Creative Commons Platform: Logo design materials and cases provided by copyright libraries, creative commons platforms, etc. Collection method: Access these platforms, search for and browse corporate logo design materials and cases in different styles and fashions;

[0086] Professional Exhibitions and Events: Professional exhibitions, seminars, forums and other events related to corporate logos. Collection method: Participate in these events, communicate with industry professionals, and collect information such as the latest trends, design concepts and application cases of corporate logos.

[0087] S4. Use the preprocessed data to train the model, and improve the recognition accuracy by adjusting the model parameters and optimizing the algorithm; in this step, optimize the model parameters through the backpropagation algorithm. During the training process, adopt strategies such as learning rate decay and momentum optimization to improve the training efficiency.

[0088] The backpropagation algorithm calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the opposite direction of the gradient, thereby minimizing the loss. During the backpropagation process, it is necessary to calculate the gradient of each layer and pass these gradients to the previous layer until the parameters of all layers are updated.

[0089] Learning rate decay dynamically adjusts the learning rate according to the training progress, and as the training progresses, the learning rate gradually decreases, which helps the model to more finely adjust the parameters in the later stage of training.

[0090] Momentum optimization introduces a momentum term, which is the weighted sum of the direction of the previous gradient update and the current gradient. Through momentum optimization, the model updates the parameters faster along the correct direction and reduces the oscillation in the wrong direction at the same time.

[0091] S5. Design an online learning mechanism so that the algorithm can continuously receive new data and learn and update in actual applications. This helps the algorithm to adapt to the continuously changing corporate logo styles and features. According to the feedback and recognition results in actual applications, adjust and optimize the algorithm. For example, retrain the samples with recognition errors or adjust the strategies of feature extraction and model training.

[0092] In this embodiment, the design of the online learning mechanism includes:

[0093] Online learning framework: Select or develop a deep learning framework that supports online learning, such as TensorFlow, PyTorch, etc., and the framework supports incremental learning and continuous learning so that the algorithm can continuously receive new data and update the model;

[0094] Model update strategy: Design a model update strategy, including learning rate adjustment, model fine-tuning, weight update, etc. Adopt a progressive learning method, that is, each time new data is received, only a part of the model's parameters are updated to reduce the computational cost and avoid overfitting;

[0095] Feedback and evaluation: Design a feedback mechanism to collect the algorithm recognition results and user feedback in actual applications, and regularly evaluate the algorithm. Adjust the model parameters and learning strategy according to the evaluation results.

[0096] S6. Input the enterprise logo to be recognized into the trained model for feature extraction and classification recognition. The model will output one or more possible recognition results, which may include information such as the category to which the enterprise logo belongs and the similarity score. According to the similarity score or other metrics of the recognition results, screen and sort the output results to determine the most likely recognition result, and then parse this recognition result to extract useful enterprise information. The extraction process: Match the recognition result with the information in the enterprise database. This requires building a database containing information such as enterprise names, brands, industries, etc., and comparing the recognition result with the entries in the database. Once the recognition result matches an entry in the database, useful information related to the enterprise logo, such as the enterprise name, brand name, industry, contact information, etc., can be extracted from the database, and the extracted information is verified to ensure the accuracy and integrity of the information. If necessary, missing information can also be supplemented from other reliable sources.

[0097] Such as Figure 2 As shown, an optimization method for an intelligent parsing algorithm of enterprise logo data, applied to the intelligent parsing algorithm of enterprise logo data as described above, includes the following steps:

[0098] S1. Online learning and update:

[0099] Adopt an online learning mechanism to enable the algorithm to continuously receive new data and learn and update in actual applications to adapt to the continuously changing enterprise logo styles and features. In this step, regarding data reception, specifically, design a data collection mechanism to obtain new enterprise logo data from various sources in real-time or regularly.

[0100] S2. Computational resource optimization:

[0101] Utilize technical means such as distributed computing and GPU acceleration to optimize the computational resource occupancy and time cost of the algorithm to improve the running efficiency of the algorithm and reduce the demand for hardware resources. In this embodiment, this step includes the following implementation details:

[0102] Distributed computing: Split the computational tasks of the algorithm into multiple subtasks and distribute them to multiple computing nodes for parallel processing.

[0103] GPU Acceleration: Utilize the powerful computing capabilities of the GPU to accelerate intensive computing tasks such as matrix operations and convolution operations in the algorithm. This can significantly shorten the running time of the algorithm.

[0104] Algorithm Optimization: Conduct low-level optimization of the algorithm, such as memory management, caching strategies, and reduction of algorithm complexity, to reduce computational overhead and time costs.

[0105] Hardware Selection: Based on the computational requirements of the algorithm and budget constraints, select a suitable hardware platform (such as high-performance servers, cloud computing services, etc.) to deploy the algorithm.

[0106] S3, Deployment and Monitoring:

[0107] Deploy the optimized algorithm into the actual application scenario and conduct continuous monitoring. By collecting user feedback and recognition results, promptly discover and solve potential problems to ensure the stability and accuracy of the algorithm. In this embodiment, this step includes the following implementation details:

[0108] Deployment Strategy: According to the requirements of the application scenario and the performance characteristics of the algorithm, select a suitable deployment strategy (such as cloud services, on-premises deployment, etc.). Ensure that the algorithm can run stably and efficiently in the actual environment.

[0109] Monitoring Mechanism: Establish a comprehensive monitoring mechanism, including indicators such as the running status of the algorithm, recognition accuracy rate, and response time. Through real-time monitoring and alarm systems, promptly discover and solve potential problems.

[0110] User Feedback: Actively collect user feedback and recognition results to continuously improve and optimize the algorithm. User feedback can be obtained through user surveys, online reviews, customer service, etc.

[0111] Regular Evaluation: Regularly evaluate and test the performance of the algorithm to ensure that it can adapt to changing application scenarios and requirements. If problems or performance degradation are found, promptly adjust the algorithm or retrain the model.

[0112] The embodiments of the present invention are given for purposes of illustration and description, and are not exhaustive or limit the present invention to the disclosed form. Many modifications and variations are obvious to those of ordinary skill in the art. The embodiments are selected and described to better illustrate the principles of the present invention and its practical applications, and to enable those of ordinary skill in the art to understand the present invention and design various embodiments with various modifications suitable for specific purposes.

Claims

1. An intelligent analysis algorithm for enterprise identification data, characterized in that: Includes steps: S1. EfficientNet is selected as the basic model, and multiple defense mechanisms are integrated against adversarial attacks. S2. Use model pruning or knowledge distillation technology to lightweight the model to reduce computing resources and time costs; S3. Collect enterprise logo data of different styles, and perform cleaning and enhancement processing; S4, use the preprocessed data to train the model, and improve the recognition accuracy by adjusting the model parameters and optimizing the algorithm; S5. Design an online learning mechanism to enable the algorithm to continuously receive new data and learn and update in actual applications, and adjust and optimize the algorithm based on feedback and recognition results in actual applications; S6. Input the corporate logo to be identified into the trained model, perform feature extraction and classification identification, analyze the identification results, and extract useful corporate information.

2. According to claim 1, the intelligent analysis algorithm for enterprise identification data is characterized in that: In step S1, the defense mechanism includes adversarial training, gradient regularization, defense distillation and ensemble learning, as follows: Adversarial training: By adding adversarial samples to the training data, the model learns how to resist these attacks during the training process, thereby improving the model's robustness to adversarial perturbations; Gradient regularization: Enhance model robustness by penalizing the sensitivity of model output to input changes. Specifically, a regularization term related to the input gradient is added to the loss function to limit the model's oversensitivity to small changes in the input. Defensive distillation: Improve model robustness through knowledge distillation techniques, specifically by training a teacher model and using the teacher model's output as soft labels to train the student model; Ensemble learning: Improve overall performance by combining the prediction results of multiple models. Specifically, in terms of defending against adversarial attacks, robustness is improved by integrating multiple EfficientNet models with different structures and parameters.

3. According to claim 1, the intelligent analysis algorithm for enterprise identification data is characterized in that: In step S2, the model pruning and lightweight processing flow is as follows: Sparse training: By adding the L1 regularization term, the model weights are gradually sparse during the training process to determine which weights are redundant; Determine the pruning strategy: Select the pruning method while ensuring that the performance of the basic model is not affected, and determine the pruning granularity and proportion based on the structure and performance requirements of the EfficientNet model; Implement pruning: Use the pre-trained EfficientNet model as a starting point and remove redundant parameters layer by layer or kernel by kernel according to the pruning strategy; Model fine-tuning: Fine-tune the pruned model to recover the accuracy lost due to pruning.

4. According to claim 1, the intelligent analysis algorithm for enterprise identification data is characterized in that: In step S2, the knowledge distillation lightweight processing flow is as follows: Prepare the teacher model and student model: select a trained EfficientNet model as the teacher model, and design a neural network structure as the student model; Generate soft labels: Use the teacher model to predict the training data set and generate soft labels containing probability distribution; Training the student model: Input the training data set and the corresponding soft labels into the student model for training. At the same time, use the hard labels of the original data as supervision information, compare them with the output of the student model, calculate the loss function, and optimize the loss function to make the output of the student model as close as possible to the output of the teacher model. Evaluation and optimization: Evaluate the performance of the student model on the validation dataset to ensure that its accuracy and generalization ability meet the requirements, and further optimize and adjust the student model based on the evaluation results.

5. The intelligent analysis algorithm for enterprise identification data according to claim 1 is characterized in that: In step S3, the sources of the enterprise identification data include: Internal enterprise database: the enterprise's own database, containing relevant information about the enterprise's identity; Government and industry organizations: public data, reports or guidelines issued by government agencies, industry associations and chambers of commerce; Professional research institutions and consulting companies: industry reports, market research reports, and white papers issued by market research institutions and consulting companies; Social media and online platforms: corporate logo data on social media platforms, online design platforms, and e-commerce platforms; Market research and consumer research: data collected through market research activities and consumer survey questionnaires; Copyright library and creative sharing platform: logo design materials and cases provided by copyright library and creative sharing platform; Professional exhibitions and activities: professional exhibitions, seminars, forums and activities related to corporate identity.

6. The intelligent analysis algorithm for enterprise identification data according to claim 1 is characterized in that: In step S4, the model parameters are optimized by back propagation algorithm, and during the training process, learning rate decay and momentum optimization strategies are adopted to improve the training efficiency; The back propagation algorithm minimizes the loss by calculating the gradient of the loss function with respect to the model parameters and updating the parameters in the opposite direction of the gradient; The learning rate decay dynamically adjusts the learning rate according to the training progress, and as the training progresses, the learning rate gradually decreases; The momentum optimization introduces a momentum term, which is a weighted sum of the direction of the previous gradient update and the current gradient. Through momentum optimization, the model can update parameters in the correct direction faster while reducing oscillations in the wrong direction.

7. The intelligent analysis algorithm for enterprise identification data according to claim 1 is characterized in that: In step S5, the design of the online learning mechanism includes: Online learning framework: Select or develop a deep learning framework that supports online learning and supports incremental learning and continuous learning; Model update strategy: Design a model update strategy that uses a progressive learning approach, that is, each time new data is received, only a portion of the model parameters are updated; Feedback and evaluation: Design a feedback mechanism to collect algorithm recognition results and user feedback in actual applications, evaluate the algorithm regularly, and adjust model parameters and learning strategies based on the evaluation results.

8. The intelligent analysis algorithm for enterprise identification data according to claim 1 is characterized in that: In step S6, the model will output one or more possible recognition results, and the output results will be screened and sorted according to the similarity scores or other indicators of the recognition results to determine the most likely recognition result.

9. The intelligent analysis algorithm for enterprise identification data according to claim 8 is characterized in that: In step S6, the recognition result analysis and information extraction process: the recognition result is matched with the information in the enterprise database. Once the recognition result successfully matches the entry in the database, useful information related to the enterprise logo can be extracted from the database and the extracted information can be verified.

10. A method for optimizing an intelligent parsing algorithm for enterprise identification data, applied to the intelligent parsing algorithm for enterprise identification data as claimed in any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Online learning and updating: Adopting online learning mechanism, the algorithm can continuously receive new data and learn and update in actual application to adapt to the ever-changing corporate logo style and characteristics; S2. Computing resource optimization: Use distributed computing and GPU acceleration to optimize the computing resource usage and time cost of the algorithm to improve the algorithm's operating efficiency and reduce the demand for hardware resources; S3, deployment and monitoring: Deploy the optimized algorithm to actual application scenarios and conduct continuous monitoring to promptly discover and resolve potential problems by collecting user feedback and identifying results.

Images