A self-optimizing image recognition method, device, medium, and product that integrates feature distillation and data management.

By using a pre-trained model with frozen parameters and a student model trained with knowledge distillation, combined with a data management system, the image recognition system achieved self-optimization, solved the model update problem, and improved the system's adaptability and recognition robustness.

CN123071797APending Publication Date: 2026-11-10XISHUANGBANNA TROPICAL BOTANICAL GARDEN CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611334438.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing image recognition systems are difficult to continuously update after model deployment, lack linkage between knowledge distillation and data management, are not adaptable enough to dynamically changing scenarios, and the robustness of feature retrieval and recognition decisions needs to be improved.

Method used

Feature extraction is performed using a pre-trained visual basic model with parameter freezing, and student models are trained by combining knowledge distillation. A vector database is established and automatic linkage between data management and models is realized. Similarity retrieval and recognition are performed by compressing feature vectors.

Benefits of technology

It achieves a closed-loop linkage of efficient feature compression, automatic data management, and continuous model optimization without the need for repeated training of large-scale pre-trained models, thereby improving recognition accuracy and category coverage while reducing computational resource consumption and time costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071797A_ABST
    Figure CN123071797A_ABST
Patent Text Reader

Abstract

This application discloses a self-optimizing image recognition method, device, medium, and product that integrates feature distillation and data management, relating to the field of image recognition. The method includes: extracting features from a target image using a pre-trained visual baseline model with frozen parameters; constructing a labeled feature dataset; training a student model using knowledge distillation based on the feature dataset; converting global feature vectors into compressed feature vectors using the trained student model; extracting compressed feature vectors from the image to be recognized using the pre-trained visual baseline model and the trained student model; performing a similarity search in a vector database; and determining the recognition result based on the search results. When an external data management platform generates new labeled image data, feature extraction and vector conversion are automatically performed. This application achieves a closed-loop linkage between efficient feature compression, automatic data management, and continuous model optimization without repeatedly training a large-scale pre-trained model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of image recognition, and in particular to a self-optimizing image recognition method, device, medium, and product that links feature distillation with data management. Background Technology

[0002] With the rapid development of deep learning technology, image recognition models based on architectures such as Convolutional Neural Networks (CNN) and Visual Transformers (ViT) have been widely used in fields such as species classification, medical image analysis, and industrial quality inspection. In particular, in recent years, the emergence of large-scale visual foundational models (such as the DINO series and CLIP) based on self-supervised pre-training has significantly improved the ability to extract general image features, providing a strong feature representation foundation for downstream recognition tasks.

[0003] However, current deep learning-based image recognition systems generally face the following technical bottlenecks in practical deployment and long-term operation and maintenance: First, it's difficult to continuously update the model after deployment. Existing image recognition systems typically employ a fixed process of "offline training—deployment—static service." Once a model is trained and deployed to the production environment, its parameters are frozen. When new recognition categories emerge in real-world applications (e.g., newly discovered species, new records), or when the accuracy of existing categories decreases due to changes in sample distribution, the system cannot adapt autonomously. Manual intervention is required to re-collect data, label samples, and fully train the entire neural network model before replacement and deployment. This process is not only time-consuming and labor-intensive but also demands extremely high computing resources. Especially for large-scale pre-trained models with hundreds of millions or even billions of parameters, each full fine-tuning or retraining represents a huge computational burden and time cost.

[0004] Second, there is a lack of effective linkage between knowledge distillation technology and data management. Knowledge distillation, as a model compression and transfer technique, has been widely used to transfer knowledge from large teacher models to lightweight student models to reduce the computational burden during the inference phase. However, existing distillation schemes are mostly one-off, static training processes, meaning that after distillation is completed on a fixed dataset, the student model also enters a fixed state. There is a lack of structured linkage between the distillation process and the backend data management system. When new labeled data is generated, vector updates or model iterations cannot be automatically triggered, resulting in the value of new data not being incorporated into the recognition system in a timely and efficient manner. Consequently, the system's species coverage or category coverage capability cannot automatically improve with data accumulation.

[0005] Third, the adaptability to dynamically changing scenarios is insufficient. In specific application areas such as biodiversity monitoring and citizen science data collection, there are many species and new samples and records are constantly emerging, which puts forward high requirements for the category expansion capability and continuous learning capability of image recognition systems. Traditional solutions include: (1) using online learning strategies to update model parameters in real time, but this method is prone to catastrophic forgetting, which leads to a sharp degradation of the model's recognition performance for existing categories and poor stability; (2) using large-scale vision-language models (such as CLIP) for full parameter fine-tuning. Although it has a certain open category recognition capability, the computational overhead of fine-tuning is significant, and the deployment flexibility on edge devices is limited; (3) having maintenance personnel manually trigger model retraining periodically. This method still relies on manual operation in essence, has a low degree of automation, and is difficult to meet the needs of high-frequency and fine-grained data updates.

[0006] Fourth, the robustness of feature retrieval and recognition decisions needs improvement. In the image retrieval and recognition process based on vector similarity, directly using the similarity ranking results as the final recognition output is easily affected by noisy samples, labeling bias, or local clustering effects in the feature space. This can lead to low-confidence erroneous matches being incorrectly adopted, affecting the reliability and robustness of the recognition results.

[0007] In summary, existing image recognition technologies still have significant shortcomings in terms of model iteration efficiency, the linkage between data management and model updates, automated system operation and maintenance capabilities, and robustness of retrieval and recognition. Therefore, there is an urgent need for an image recognition method that can achieve efficient feature compression, automated data management, and a closed-loop linkage for continuous model optimization without repeatedly training large-scale pre-trained models. Summary of the Invention

[0008] The purpose of this application is to provide a self-optimizing image recognition method, device, medium, and product that links feature distillation and data management, which can achieve a closed-loop linkage of efficient feature compression, automatic data management, and continuous model optimization without repeatedly training large-scale pre-trained models.

[0009] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a self-optimizing image recognition method that links feature distillation with data management, including: The target image is used to extract features using a pre-trained visual basic model with frozen parameters to obtain a global feature vector of a first preset dimension; and the global feature vector is stored in a vector database. Based on the image recognition objective, the global feature vectors in the vector database are grouped according to the classification labels to obtain a labeled feature dataset. Based on the labeled feature dataset, a student model is trained using knowledge distillation; the student model is used to map the global feature vector of the first preset dimension to the compressed feature vector of the second preset dimension; the second preset dimension is smaller than the first preset dimension; Using the trained student model, all global feature vectors of the first preset dimension in the corresponding group are converted into compressed feature vectors of the second preset dimension and stored in the corresponding vector database; The compressed feature vector of the second preset dimension of the image to be identified is extracted using a pre-trained visual basic model and a trained student model. Similarity is then searched in the corresponding vector database, and the identification result is determined based on the search results. When an external data management platform generates new labeled image data, it automatically uses a pre-trained visual basic model with frozen parameters and a trained student model to extract features and transform vectors from the new labeled image data, and stores the obtained compressed feature vectors and corresponding labels in the vector database.

[0010] Secondly, this application provides a self-optimizing image recognition device that links feature distillation with data management, comprising: The feature extraction unit integrates a pre-trained visual basic model with frozen parameters, which is used to extract features from the input target image and output a global feature vector of the first preset dimension. According to the image recognition purpose, the global feature vectors in the vector database are grouped based on classification labels to obtain a labeled feature dataset. The distillation training unit is configured with a student network structure to receive a labeled feature dataset and obtain a trained student model through knowledge distillation training; the student model is used to map the global feature vector of the first preset dimension to a compressed feature vector of the second preset dimension. A vector management unit is used to store the global feature vector and the compressed feature vector, and to provide a retrieval service based on vector similarity. The recognition service unit is used to respond to recognition requests, call the feature extraction unit and the trained student model to obtain the compressed feature vector of the image to be recognized, and perform similarity retrieval in the vector management unit to determine the recognition result; The data management unit communicates with an external data management platform. When the external data management platform generates new labeled image data, it automatically triggers the feature extraction unit and the trained student model to perform feature extraction and vector transformation on the new labeled image data, and writes the obtained compressed feature vector and corresponding classification label into the vector management unit to achieve automatic updating of recognition capabilities.

[0011] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the self-optimizing image recognition method that links feature distillation and data management.

[0012] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the self-optimizing image recognition method that links feature distillation and data management.

[0013] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the self-optimizing image recognition method that links feature distillation and data management.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a self-optimizing image recognition method, device, medium, and product that integrates feature distillation and data management. By freezing the parameters of a pre-trained visual base model, it is used solely as a fixed feature extractor throughout its entire lifecycle, eliminating the need to retrain a large-scale pre-trained model with each data update or category expansion. When expanding recognition categories or updating the model, only distillation training and vector transformation operations on a lightweight student model are required, significantly reducing the number of training parameters and computational overhead compared to full fine-tuning of a large-scale pre-trained model. Through knowledge distillation training, high-dimensional global feature vectors of a first preset dimension are compressed into compressed feature vectors of a second preset dimension, significantly reducing vector dimensionality while preserving the original feature semantic information. When performing similarity retrieval based on compressed feature vectors, storage usage and computational complexity are effectively reduced, improving retrieval response speed and making this application suitable for real-time retrieval scenarios with large-scale vector libraries. By establishing an automatic linkage mechanism between an external data management platform and a vector database, when the data management platform generates new labeled image data, it automatically calls the fixed pre-trained visual base model and student model to complete feature extraction and vector transformation, and writes the new vectors into the vector database. The entire process requires no manual data collection, re-labeling, or triggering of a complete training cycle. It automatically improves recognition accuracy and category coverage without human intervention, achieving a self-optimization goal of "deploy once, evolve continuously." Newly labeled image data, after review, only requires one feature extraction and one vector transformation before being incorporated into the retrieval system, without waiting for a full training cycle. Newly labeled data can be instantly converted into searchable compressed feature vectors, significantly shortening the time window from data generation to the effectiveness of recognition capabilities, and accelerating the system's learning and coverage of new categories. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of a self-optimizing image recognition method that links feature distillation and data management in one embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] In one exemplary embodiment, such as Figure 1 As shown, a self-optimizing image recognition method that links feature distillation with data management is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, and includes the following steps S101 to S106. Wherein: S101, using a pre-trained visual basic model with frozen parameters to extract features from the target image, obtaining a global feature vector of a first preset dimension; and storing the global feature vector in a vector database; wherein, the first preset dimension can be 1280 dimensions; As a specific implementation, the vector database includes fields such as image ID, image address, and vector. As a specific example, the parameter-frozen pre-trained visual base model includes, but is not limited to, the open-source DINOv3 model and other visual base models such as CLIP; As a specific embodiment, S101 specifically includes: S11, Obtain the target image from the external data management platform; S12, perform preprocessing operations on the target image; the preprocessing operations include image scaling, center cropping, pixel value normalization (0~1) and standardization processing; Specifically, the official AutoImageProcessor can be used to preprocess the target image; S13, freeze the pre-trained visual base model with the pre-processed image input parameters, and perform forward propagation of the encoder; S14: Obtain the feature output corresponding to the zeroth CLS marker in the sequence output by the pre-trained visual base model. The dimension of a single vector is 1280, which directly represents the global semantic representation of the entire image. S15, the feature output corresponding to the zeroth CLS marker in the sequence is used as the global feature vector of the first preset dimension. Other preprocessed images are processed in a loop until all vectors are stored in the database.

[0020] As a specific example, the external data management platform is the Jiaxin Notes Citizen Science Platform (https: / / inewrecords.cubg.cn / ). The Xishuangbanna Tropical Botanical Garden of the Chinese Academy of Sciences is the owner of the Jiaxin Notes Citizen Science Platform and has the right to use the images for public benefit. Data reviewed by experts on the Jiaxin Notes Citizen Science Platform is marked as research-level, has relatively authoritative species tags, and can be applied to image recognition models. The experts on the Jiaxin Notes Citizen Science Platform are nature enthusiasts or researchers with strong taxonomic backgrounds. The system administrator sets the criteria based on the quality of data uploaded or assessed by users, with an accuracy rate of over 90%.

[0021] S102, according to the image recognition purpose, the global feature vectors in the vector database are grouped based on the classification labels to obtain a labeled feature dataset; S102 specifically includes: S21, Obtain the multi-level classification label information corresponding to each target image; The multi-level classification label information includes kingdom, phylum, class, order, family, genus, and species; the following are examples of data grouping methods for different application purposes: (1) General identification: The vectors are grouped by species-level labels, and species with more than 30 vectors in each group are used for distillation; (2) Plant identification: The vectors are grouped by species and the boundary field value is limited to the plant kingdom. Species with more than 30 vectors in each group are used for distillation. The grouping principles of groups such as the animal kingdom, fungi kingdom, birds, and insects are similar. (1) Disease and pest identification: First, compile a list of diseases and pests, then group them according to the list, and select species with more than 30 vectors in each group for distillation. The application methods for other special purposes are similar; S22, determine the grouping strategy according to the preset recognition purpose, and group the global feature vector according to the classification level corresponding to the grouping strategy; S23, count the number of global feature vectors in each group, and select the groups whose number of global feature vectors meets the preset threshold as valid groups; S24, the global feature vectors and corresponding multi-level classification label information in the effective grouping are used to form the labeled feature dataset.

[0022] S103, Based on the labeled feature dataset, a student model is trained using knowledge distillation; the student model is used to map the global feature vector of the first preset dimension to the compressed feature vector of the second preset dimension; the second preset dimension is smaller than the first preset dimension; the first preset dimension is 1280 dimensions, and the second preset dimension here can be 640 dimensions (or other dimensions); that is, a lightweight student neural network model is trained to compress 1280-dimensional teacher features into 640-dimensional student features.

[0023] The specific process of S103 is as follows: (1) Data input Obtain labeled high-dimensional feature datasets generated by the teacher model ,in, For the first Teacher feature vectors for each sample, ; The corresponding species label is used; the teacher feature vector is processed by L2 normalization, i.e. .

[0024] (2) Construction of student model Building a lightweight student neural network ,in, Its structure includes: one or two fully connected layers, a batch normalization and ReLU output layer, and an L2 normalization module following the output layer; the L2 normalization module ensures student characteristics satisfy .

[0025] (3) Design of multi-objective distillation loss function The multi-objective distillation loss function adopts a joint loss function, which is as follows: ; in, For the joint loss function, This is the feature alignment loss, used to constrain the consistency of student features and teacher features in direction. , The total number of samples, For the first Teacher feature vectors for each sample, Let T represent the student features of the i-th sample, and let T denote the transpose. This is a similarity distillation loss, used to maintain the relative similarity structure between samples within a batch. The student feature matrix of the current training batch Teacher feature matrix of the current training batch , It is the Frobenius norm. This represents the number of samples.

[0026] (4) Training and Deployment The AdamW optimizer is used for end-to-end training of parameters θ (or weight matrix W); after training, the student model is frozen and integrated into an edge device or API service; during the inference phase, arbitrary teacher features are input. Output 640-dimensional L2 normalized student features It is used for subsequent retrieval, clustering or classification tasks.

[0027] The model can be trained using knowledge distillation as described above, or it can be trained using online learning. S104: Using the trained student model, convert all global feature vectors of the first preset dimension in the corresponding group into compressed feature vectors of the second preset dimension, and store them in the corresponding vector database. The trained student model is used to convert all 1280-dimensional vectors for the corresponding group into 640-dimensional vectors. Multiple student models can be trained according to different needs, and the 640-dimensional vectors converted by each student model are saved in a separate table, associated with species information via image ID.

[0028] S105, using the pre-trained visual basic model and the trained student model, extract the compressed feature vector of the second preset dimension of the image to be identified, perform similarity retrieval in the corresponding vector database, and determine the recognition result based on the retrieval result; S105 specifically includes: S51, based on the extracted compressed feature vector of the second preset dimension of the image to be identified, perform a similarity search in the corresponding vector database to obtain the search results; S52, Based on the search results, construct a candidate result set; the candidate result set includes multiple candidate vectors and their corresponding similarity scores; and arrange the similarity scores in descending order; Among them, the candidate result set for In the formula, Let i be the i-th candidate vector. Calculate its similarity score with the query vector, and satisfy the following conditions: .

[0029] S53, determine the highest similarity value in the candidate result set; make This represents the highest similarity value in the candidate result set; S54, Based on the highest similarity value, a subset of candidates whose similarity scores meet the preset screening conditions is selected from the candidate result set; by screening and suppressing low-confidence matching items, high consistency results with non-linear correlation with the best matching are retained, thereby improving the robustness of subsequent identification or decision-making. Among them, candidate subset for: ; S55, Calculate the percentage of candidates corresponding to each category label in the candidate subset; S56, the classification label with the highest proportion is determined as the recognition result of the image to be recognized, and the corresponding highest similarity value is used as the confidence level of the recognition result.

[0030] S106, when the external data management platform generates new labeled image data, it automatically uses the pre-trained visual basic model with frozen parameters and the trained student model to perform feature extraction and vector transformation on the new labeled image data, and stores the obtained compressed feature vectors and corresponding labels in the vector database.

[0031] The external data management platform is used to mark newly added labeled image data as research-level data after the data review mechanism has approved it. Feature extraction and vector transformation are performed on newly added labeled image data marked as research-level data. Image labels are revised through the identification process, ultimately achieving automatic optimization of the recognition results without the need for manual intervention or repeated training.

[0032] In this application, the parameters of the pre-trained visual base model are frozen during the training process and it is used only as a feature extractor, thus avoiding the computational overhead of repeatedly training large models.

[0033] Compared with the prior art, this application has the following advantages: (1) Break through the bottleneck of repeated training: Use the pre-trained visual base model (DINOv3) with frozen parameters as a fixed feature source to avoid training a large model from scratch in each iteration, which significantly reduces the consumption of computing resources. (2) Achieve automatic and continuous optimization: Through the linkage between image management and vector search system, the recognition accuracy and species coverage can be automatically improved without human intervention; (3) Strong compatibility: The student model can be flexibly replaced and is suitable for deployment on edge devices; (4) High data utilization efficiency: Newly labeled data can be incorporated into vector retrieval in real time, accelerating the model's learning of new species.

[0034] In one exemplary embodiment, a self-optimizing image recognition device that links feature distillation with data management is provided, comprising: The feature extraction unit integrates a pre-trained visual basic model with frozen parameters, which is used to extract features from the input target image and output a global feature vector of the first preset dimension. According to the image recognition purpose, the global feature vectors in the vector database are grouped based on classification labels to obtain a labeled feature dataset. The distillation training unit is configured with a student network structure to receive a labeled feature dataset and obtain a trained student model through knowledge distillation training; the student model is used to map the global feature vector of the first preset dimension to a compressed feature vector of the second preset dimension. A vector management unit is used to store the global feature vector and the compressed feature vector, and to provide a retrieval service based on vector similarity. The recognition service unit is used to respond to recognition requests, call the feature extraction unit and the trained student model to obtain the compressed feature vector of the image to be recognized, and perform similarity retrieval in the vector management unit to determine the recognition result; The data management unit communicates with an external data management platform. When the external data management platform generates new labeled image data, it automatically triggers the feature extraction unit and the trained student model to perform feature extraction and vector transformation on the new labeled image data, and writes the obtained compressed feature vector and corresponding classification label into the vector management unit to achieve automatic updating of recognition capabilities.

[0035] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a self-optimizing image recognition method that integrates feature distillation and data management.

[0036] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0037] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0038] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0039] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0040] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0041] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0042] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A self-optimizing image recognition method that links feature distillation with data management, characterized in that, include: The target image is used to extract features using a pre-trained visual basic model with frozen parameters to obtain a global feature vector of a first preset dimension; and the global feature vector is stored in a vector database. Based on the image recognition objective, the global feature vectors in the vector database are grouped according to the classification labels to obtain a labeled feature dataset. Based on the labeled feature dataset, a student model is trained using knowledge distillation; the student model is used to map the global feature vector of the first preset dimension to the compressed feature vector of the second preset dimension; the second preset dimension is smaller than the first preset dimension; Using the trained student model, all global feature vectors of the first preset dimension in the corresponding group are converted into compressed feature vectors of the second preset dimension and stored in the corresponding vector database; The compressed feature vector of the second preset dimension of the image to be identified is extracted using a pre-trained visual basic model and a trained student model. Similarity is then searched in the corresponding vector database, and the identification result is determined based on the search results. When an external data management platform generates new labeled image data, it automatically uses a pre-trained visual basic model with frozen parameters and a trained student model to extract features and transform vectors from the new labeled image data, and stores the obtained compressed feature vectors and corresponding labels in the vector database.

2. The self-optimizing image recognition method linking feature distillation and data management according to claim 1, characterized in that, The step of extracting features from the target image using a pre-trained visual base model with frozen parameters to obtain a global feature vector of a first preset dimension specifically includes: Obtain the target image from the external data management platform; The target image is preprocessed; the preprocessing operations include image scaling, center cropping, pixel value normalization, and standardization. The pre-processed image is input into the pre-trained visual base model with frozen parameters, and then the encoder is propagated forward. Obtain the feature output corresponding to the zeroth CLS marker in the sequence output by the pre-trained visual base model; The feature output corresponding to the zeroth CLS marker in the sequence is used as the global feature vector of the first preset dimension.

3. The self-optimizing image recognition method linking feature distillation and data management according to claim 1, characterized in that, The step of grouping the global feature vectors in the vector database based on classification labels according to the image recognition purpose to obtain a labeled feature dataset specifically includes: Obtain the multi-level classification label information corresponding to each target image; A grouping strategy is determined based on the preset recognition purpose, and the global feature vector is grouped according to the classification level corresponding to the grouping strategy. The number of global feature vectors in each group is counted, and groups whose number of global feature vectors meets the preset threshold are selected as valid groups. The global feature vectors and corresponding multi-level classification label information in the effective groupings constitute the labeled feature dataset.

4. The self-optimizing image recognition method linking feature distillation and data management according to claim 1, characterized in that, Knowledge distillation training is optimized using a joint loss function; the joint loss function is: ; in, For the joint loss function, This is the feature alignment loss, used to constrain the consistency of student features and teacher features in direction. , The total number of samples, For the first Teacher feature vectors for each sample, Let T represent the student features of the i-th sample, and let T denote the transpose. This is a similarity distillation loss, used to maintain the relative similarity structure between samples within a batch. The student feature matrix of the current training batch Teacher feature matrix of the current training batch , It is the Frobenius norm. This represents the number of samples.

5. The self-optimizing image recognition method linking feature distillation and data management according to claim 1, characterized in that, The process of extracting a compressed feature vector of the second preset dimension of the image to be identified using a pre-trained visual basic model and a trained student model, performing a similarity search in the corresponding vector database, and determining the recognition result based on the search results specifically includes: Based on the compressed feature vector of the second preset dimension of the extracted image to be identified, a similarity search is performed in the corresponding vector database to obtain the search results; Based on the search results, a candidate result set is constructed; the candidate result set includes multiple candidate vectors and their corresponding similarity scores; and the similarity scores are arranged in descending order. Determine the highest similarity value in the candidate result set; Based on the highest similarity value, a subset of candidates whose similarity scores meet the preset filtering conditions is selected from the candidate result set; The percentage of candidates corresponding to each category label in the candidate subset is calculated. The category label with the highest proportion is determined as the recognition result of the image to be recognized, and the corresponding highest similarity value is used as the confidence level of the recognition result.

6. The self-optimizing image recognition method linking feature distillation and data management according to claim 1, characterized in that, The external data management platform is used to mark newly added labeled image data as research-level data after the data review mechanism has approved it. Feature extraction and vector transformation are performed on newly added labeled image data marked as research-level data.

7. A self-optimizing image recognition device that links feature distillation with data management, characterized in that, include: The feature extraction unit integrates a pre-trained visual basic model with frozen parameters, which is used to extract features from the input target image and output a global feature vector of the first preset dimension. According to the image recognition purpose, the global feature vectors in the vector database are grouped based on classification labels to obtain a labeled feature dataset. The distillation training unit is configured with a student network structure to receive a labeled feature dataset and obtain a trained student model through knowledge distillation training; the student model is used to map the global feature vector of the first preset dimension to a compressed feature vector of the second preset dimension. A vector management unit is used to store the global feature vector and the compressed feature vector, and to provide a retrieval service based on vector similarity. The recognition service unit is used to respond to recognition requests, call the feature extraction unit and the trained student model to obtain the compressed feature vector of the image to be recognized, and perform similarity retrieval in the vector management unit to determine the recognition result; The data management unit communicates with an external data management platform and is used to automatically trigger the feature extraction unit and the trained student model to perform feature extraction and vector transformation on the newly added labeled image data when the external data management platform generates new labeled image data. The obtained compressed feature vector and corresponding classification label are written into the vector management unit to realize automatic updating of recognition capabilities.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the self-optimizing image recognition method of feature distillation and data management linkage as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the self-optimizing image recognition method that links feature distillation and data management as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the self-optimizing image recognition method that links feature distillation and data management as described in any one of claims 1-6.