Text search image method and device based on user feedback, electronic equipment and storage medium

Abstract

The application is suitable for the field of image retrieval technology, and provides a text-to-image method based on user feedback, comprising: obtaining a target query text of a query user and a first image query result of the target query text, and obtaining at least one query intention group of the user, the query intention group comprising a semantic center vector and a latest classification model; extracting a target feature vector corresponding to the target query text, and calculating the similarity between the target feature vector and the semantic center vector in the query intention group; if there is a target query intention group with a similarity greater than a preset similarity threshold, calling the latest classification model in the target query intention group to classify and predict the first image query result, and obtaining a final image query result of the target query text; if there is no target query intention group with a similarity greater than the preset similarity threshold, taking the first image query result as the final image query result of the target query text. The application can improve the query accuracy.

Description

Technical Field

[0001] This application belongs to the field of image retrieval technology, and in particular relates to a text-to-image search method, apparatus, electronic device and storage medium based on user feedback. Background Technology

[0002] With the rapid development of deep learning technology, multimodal and cross-modal retrieval technology has become a core technology in the field of image search. Among them, the CLIP model, with its advantage of large-scale image-text pair comparison learning, can map text and images to the same semantic space, achieving efficient text-to-image retrieval and is widely used in scenarios such as searching for things or people in a sentence. Currently, the typical retrieval process for existing text-to-image searches is as follows: user inputs query text → system extracts text feature vectors through the CLIP model → calculates similarity with pre-stored image feature vectors in the image library → returns search results in order of similarity. The CLIP model is a general pre-trained model whose semantic understanding is based on a large-scale public dataset. However, it does not consider the personalized search preferences of specific users, leading to discrepancies between search results and users' true intentions when using the same search term, resulting in a poor user experience. Furthermore, the positive and negative sample data manually selected and labeled by users is not effectively utilized, preventing continuous optimization of the retrieval model through user feedback and the formation of a closed loop of "retrieval-feedback-optimization." Optimizing the CLIP model using traditional fine-tuning methods requires retraining deep neural networks, resulting in high computational costs, long training cycles, and high resource consumption, failing to meet the business requirements of online real-time optimization. Treating the CLIP feature vector of the query text as a fixed value means it does not dynamically adjust with multiple rounds of user feedback, preventing the system from gradually approaching the user's true search intent and limiting optimization effectiveness. Therefore, there is an urgent need for a text search graph method based on user feedback that can achieve lightweight online optimization, efficient reuse of user feedback, and dynamic evolution of query semantics without altering the main structure of the CLIP model or involving specific model training algorithms. This would address the problems of insufficient personalized adaptation, inability to reuse user feedback, high model optimization costs, and lack of dynamic evolution of query semantics in existing methods. Summary of the Invention

[0003] This application provides a text search image method based on user feedback, which can solve the problems of insufficient personalization adaptation, non-reusability of user feedback, high model optimization cost, and inability to dynamically evolve query semantics in existing methods. By extracting the target feature vector corresponding to the query user's target query text and calculating the similarity between the target feature vector and the semantic center vector in the query intent group, if a target query intent group with a similarity greater than a preset similarity threshold exists, the latest classification model in the target query intent group is called to classify and predict the first image query result, obtaining the final image query result of the target query text. If no target query intent group with a similarity greater than the preset similarity threshold exists, the first image query result is used as the final image query result of the target query text, and the final image query result is returned to the query user. This solves the problems of insufficient personalization adaptation, non-reusability of user feedback, high model optimization cost, and inability to dynamically evolve query semantics in existing methods, thereby improving query accuracy.

[0004] In a first aspect, embodiments of this application provide a text-to-image search method based on user feedback, the method comprising the following steps:

[0005] Obtain the target query text of the querying user and the first image query result of the target query text, and obtain at least one query intent group of the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the historical multiple image query results of the training user.

[0006] Extract the target feature vector corresponding to the target query text, and calculate the similarity between the target feature vector and the semantic center vector in the query intent group;

[0007] If there exists a target query intent group with a similarity greater than a preset similarity threshold, then the latest classification model in the target query intent group is invoked to classify and predict the first image query result, thereby obtaining the final image query result of the target query text.

[0008] If there is no target query intent group with a similarity greater than the preset similarity threshold, then the first image query result is taken as the final image query result of the target query text.

[0009] The final image query result is returned to the querying user.

[0010] Optionally, obtaining at least one query intent group of the user includes:

[0011] Obtain the initial query text of the training user and the second image query result of the initial query text, wherein the second image query result includes image data retrieved based on the initial query text;

[0012] The system receives a first feedback result from the user regarding the second image query result. The first feedback result includes a first positive sample image and a first negative sample image selected by the user. The first positive sample image is image data in the second image query result that matches the intent of the training user, and the first negative sample image is image data in the second image query result that does not match the intent of the training user.

[0013] An initial classification model is trained based on the first positive sample image and the first negative sample image;

[0014] An initial center vector is generated based on the feature vector of the first positive sample image and the feature vector of the initial query text.

[0015] Based on the initial classification model and the initial center vector, construct the user's query intent group.

[0016] Optionally, obtaining at least one query intent group of the user further includes:

[0017] Obtain the subsequent query text of the training user and the third image query result of the subsequent query text, wherein the third image query result includes image data retrieved based on the subsequent query text;

[0018] The system receives a second feedback result from the user regarding the third image query result. The second feedback result includes a second positive sample image and a second negative sample image selected by the user. The second positive sample image is image data in the third image query result that matches the intent of the training user, and the second negative sample image is image data in the third image query result that does not match the intent of the training user.

[0019] Based on the similarity between the feature vector of the subsequent query text and the semantic center vector in the existing query intent group, the query intent group to be updated corresponding to the subsequent query text is determined.

[0020] Based on the second positive sample image and the second negative sample image, as well as the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the classification model obtained from the previous training is trained to obtain the updated classification model.

[0021] Based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated, an updated center vector is generated.

[0022] The user's query intent group is updated based on the updated classification model and the updated center vector.

[0023] Optionally, the step of training the previously trained classification model based on the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, to obtain the updated classification model, includes:

[0024] In the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the most recent preset number of positive sample images and negative sample images are taken as training data in chronological order;

[0025] The previously trained classification model is trained using the training data to obtain an updated classification model.

[0026] Optionally, generating the updated center vector based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated includes:

[0027] Obtain the cumulative training count of the query intent group to be updated;

[0028] Based on the cumulative training times of the query intent group to be updated, the semantic center vector of the query intent group to be updated is fused with the feature vector of the subsequent query text to obtain the updated center vector.

[0029] Optionally, after obtaining the latest query intent group, the method further includes:

[0030] The latest classification model is serialized and persistently stored on a distributed file storage medium.

[0031] Write the latest query intent group's metadata into the database. The metadata includes: query intent group ID, query semantic center vector, cumulative training times, number and label of the current positive and negative sample images, storage path of the classification model, training time, last call time, similarity matching threshold, classification weight configuration, and system user ID of the training user.

[0032] Cache the query semantic center vector to the cache database.

[0033] Optionally, the step of calling the latest classification model in the target query intent group to classify and predict the first image query result to obtain the final image query result of the target query text includes:

[0034] The latest classification model in the target query intent group is invoked to classify and predict the first image query result, thereby obtaining the confidence level of each image data in the first image query result;

[0035] Based on the confidence level, the image data in the first image query result are filtered and sorted to obtain the final image query result of the target query text.

[0036] Secondly, embodiments of this application provide a text search image device based on user feedback, the text search image device based on user feedback comprising:

[0037] The acquisition module is used to acquire the target query text of the querying user and the first image query result of the target query text, and to acquire at least one query intent group of the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the training user's historical multiple image query results.

[0038] The first processing module is used to extract the target feature vector corresponding to the target query text and calculate the similarity between the target feature vector and the semantic center vector in the query intent group;

[0039] The classification prediction module is used to, if there is a target query intent group with a similarity greater than a preset similarity threshold, call the latest classification model in the target query intent group to classify and predict the first image query result, so as to obtain the final image query result of the target query text.

[0040] The second processing module is used to take the first image query result as the final image query result of the target query text if there is no target query intent group with a similarity greater than a preset similarity threshold.

[0041] The return module is used to return the final image query result to the querying user.

[0042] Thirdly, embodiments of the present invention provide an electronic 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 steps in the text search image method based on user feedback provided in embodiments of the present invention.

[0043] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps in the text search image method based on user feedback provided in the embodiments of the present invention.

[0044] The above-described solution of this application has the following beneficial effects: It obtains the target query text and the first image query result of the target query text from the querying user, and obtains at least one query intent group from the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the user's historical multiple image query results. It extracts the target feature vector corresponding to the target query text and calculates the similarity between the target feature vector and the semantic center vector in the query intent group. If there is a target query intent group with a similarity greater than a preset similarity threshold, it calls the latest classification model in the target query intent group to perform classification prediction on the first image query result, obtaining the final image query result of the target query text. If there is no target query intent group with a similarity greater than the preset similarity threshold, it uses the first image query result as the final image query result of the target query text. Finally, it returns the final image query result to the querying user. This invention extracts the target feature vector corresponding to the target query text of the querying user and calculates the similarity between the target feature vector and the semantic center vector in the query intent group. When there is a target query intent group with a similarity greater than a preset similarity threshold, the latest classification model in the target query intent group is called to classify and predict the first image query result, thus obtaining the final image query result of the target query text. When there is no target query intent group with a similarity greater than the preset similarity threshold, the first image query result is used as the final image query result of the target query text, thereby returning the final image query result to the querying user. This solves the problems of insufficient personalized adaptation, inability to reuse user feedback, high model optimization cost, and inability to dynamically evolve query semantics in existing methods, thereby improving query accuracy.

[0045] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art 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.

[0047] Figure 1 A flowchart illustrating a text search image method based on user feedback, provided as an embodiment of this application;

[0048] Figure 2 This is a schematic diagram of the structure of a text search image device based on user feedback provided in an embodiment of this application;

[0049] Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0050] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0051] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0052] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0053] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0054] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0055] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0056] like Figure 1 As shown, Figure 1 This is a flowchart of a text search image method based on user feedback provided in an embodiment of the present invention. The text search image method based on user feedback includes the following steps:

[0057] 101. Obtain the target query text of the querying user and the first image query result of the target query text, and obtain at least one query intent group of the user.

[0058] In this embodiment of the invention, the above-described text-to-image search method based on user feedback can be applied to a data retrieval platform. This data retrieval platform can be built on a server-based or distributed architecture and includes a data interface (for sensor or user uploads), a knowledge database, and a knowledge database construction program. The data interface can be used to obtain the user's target query text and a first image query result of the target query text, and to obtain at least one set of user query intent. The knowledge database construction program can be used to construct the knowledge database, which is specifically designed to provide additional relational information for the identified data entities, thereby enhancing the data recognition system's understanding of the content.

[0059] The users who can perform the above query can be the users currently performing text-to-image search.

[0060] The target query text mentioned above can be the text content entered by the user in the text search image, such as a white puppy rolling in the grass or a large ship at sea.

[0061] The first image query result mentioned above can be a preliminary list of images retrieved based on the target query text. The CLIP pre-trained model can be used to extract the feature vector of the target query text, then the similarity between the feature vector and the feature vectors of all images in the image library can be calculated, and the images can be sorted from high to low similarity to return the first image query result.

[0062] The above-mentioned at least one can be understood as a quantity greater than or equal to 1, that is, 1 or more.

[0063] The aforementioned query intent group includes a semantic center vector and a state-of-the-art classification model. The semantic center vector represents the overall semantic direction of the query intent group. The classification model can be a classification model built based on deep learning or machine learning, such as CLIP or SVM. CLIP can be used to achieve zero-shot image classification by mapping images and text to a shared semantic space; it could be CLIP-ViT-B / 32, which uses VisionTransformer (ViT) as the image encoder and is jointly trained with a text Transformer to achieve image-text cross-modal alignment and zero-shot recognition capabilities. The core idea of ​​SVM (Support Vector Machine) is to find an optimal hyperplane that maximizes the margin between the two classes of samples as the decision boundary, thus maximizing the generalization ability of the classification.

[0064] The aforementioned latest classification model can be the latest classification model trained based on user feedback.

[0065] The query intent group is constructed and updated based on feedback from the training user's historical image query results. The training user can be a user who participates in image queries and provides training data by annotating the query results. The historical image query results can be a preliminary list of search results returned after the training user entered different query texts in the past. The feedback results can be images manually annotated by the user regarding whether the current search result matches their current search intent, distinguishing between images that match and do not. The construction process involves creating a new query intent group based on the feedback provided by the training user for a specific query semantics; this new query intent group does not exist in the previous query intent group. The update process involves modifying the semantic center vector of an existing query intent group based on the feedback provided by the training user, making it more accurately converge to the user's true intent.

[0066] 102. Extract the target feature vector corresponding to the target query text, and calculate the similarity between the target feature vector and the semantic center vector in the query intent group.

[0067] In this embodiment of the invention, the target feature vector can be a feature vector obtained by performing feature extraction processing on the target query text through a classification model.

[0068] Cosine similarity can be used to calculate the similarity between the target feature vector and the semantic center vector in the query intent group. The cosine similarity measures the similarity between two vectors by measuring the cosine of the angle between them. The closer the cosine similarity between the target feature vector and the semantic center vector in the query intent group is to 1, the more similar they are. Conversely, the closer the cosine similarity is to -1, the less similar they are.

[0069] 103. If there is a target query intent group with a similarity greater than the preset similarity threshold, the latest classification model in the target query intent group is called to classify and predict the first image query result, and the final image query result of the target query text is obtained.

[0070] In this embodiment of the invention, the aforementioned similarity may be the similarity between the target feature vector and the semantic center vector in the query intent group.

[0071] The aforementioned preset similarity threshold can be a pre-set similarity threshold used to determine whether the semantic similarity between the feature vector of the target query text and the semantic center vector in the query intent group is sufficient. It can be understood that when the similarity between the feature vector of the target query text and the semantic center vector in the query intent group is greater than the preset similarity threshold, the target query text and the query intent group can be classified into the same query intent group. When the similarity between the feature vector of the target query text and the semantic center vector in the query intent group is greater than the preset similarity threshold, a target query intent group exists.

[0072] The aforementioned target query intent group can be a query intent group whose similarity to the feature vector of the target query text and the semantic center vector in the query intent group is greater than a preset similarity threshold.

[0073] The above classification prediction can be the process of calling the latest classification model in the target query intent group to perform classification prediction on the first image query result and output the final image query result.

[0074] The final image query result can be the list of images returned to the user after calling the latest classification model in the target query intent group to perform classification prediction on the first image query result.

[0075] It should be noted that when there is a target query intent group with a similarity greater than the preset similarity threshold, the latest classification model in the target query intent group is called to classify and predict the first image query result, so as to obtain the final image query result of the target query text, which can improve the retrieval accuracy and user satisfaction.

[0076] 104. If there is no target query intent group with a similarity greater than the preset similarity threshold, the first image query result will be used as the final image query result of the target query text.

[0077] In this embodiment of the invention, when there is no target query intent group with a similarity greater than a preset similarity threshold, the first image query result can be used as the final image query result of the target query text, without affecting the user's normal use, and ensuring the continuity and stability of the retrieval service.

[0078] 105. Return the final image query results to the querying user.

[0079] In this embodiment of the invention, the final image query result can be returned to the querying user.

[0080] Specifically, the final image query results can be transmitted over the network to the query user's client in a form that the query user can receive and perceive, and then presented to the query user by the client.

[0081] In this embodiment of the invention, the target query text and the first image query result of the target query text are obtained from the querying user, and at least one query intent group of the user is obtained. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the user's historical multiple image query results. The target feature vector corresponding to the target query text is extracted, and the similarity between the target feature vector and the semantic center vector in the query intent group is calculated. If there is a target query intent group with a similarity greater than a preset similarity threshold, the latest classification model in the target query intent group is called to perform classification prediction on the first image query result to obtain the final image query result of the target query text. If there is no target query intent group with a similarity greater than the preset similarity threshold, the first image query result is used as the final image query result of the target query text. The final image query result is returned to the querying user. This invention extracts the target feature vector corresponding to the target query text of the querying user and calculates the similarity between the target feature vector and the semantic center vector in the query intent group. When there is a target query intent group with a similarity greater than a preset similarity threshold, the latest classification model in the target query intent group is called to classify and predict the first image query result, thus obtaining the final image query result of the target query text. When there is no target query intent group with a similarity greater than the preset similarity threshold, the first image query result is used as the final image query result of the target query text, thereby returning the final image query result to the querying user. This solves the problems of insufficient personalized adaptation, inability to reuse user feedback, high model optimization cost, and inability to dynamically evolve query semantics in existing methods, thereby improving query accuracy.

[0082] It is understood that in the specific implementation of this application, data such as query data, image data, intent data, and feature data are involved. When the embodiments in this application are applied to specific products or technologies, user permission or consent is required. Furthermore, the collection, use, and processing of related data, as well as the training, deployment, and invocation of algorithm models, must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0083] Optionally, in the step of obtaining at least one query intent group of a user, the user's initial query text and the second image query result of the initial query text can be obtained; the user's first feedback result for the second image query result can be received; an initial classification model can be trained based on the first positive sample image and the first negative sample image; an initial center vector can be generated according to the feature vector of the first positive sample image and the feature vector of the initial query text; and the user's query intent group can be constructed according to the initial classification model and the initial center vector.

[0084] In this embodiment of the invention, the aforementioned training user can be a user who participates in image querying and performs annotation operations based on the image query results, thereby providing training data.

[0085] The initial query text mentioned above can be the text content that the training user initially entered in the text search image, such as "ship" or "sofa".

[0086] The second image query result mentioned above includes image data retrieved based on the initial query text.

[0087] The aforementioned first feedback result includes the first positive sample image and the first negative sample image selected by the user. The first positive sample image is the image data in the second image query result that matches the training user's intent, and the first negative sample image is the image data in the second image query result that does not match the training user's intent.

[0088] The initial classification model described above can be obtained by training a pre-trained classification model using the first positive sample image and the first negative sample image. The pre-trained classification model can be a classification model built based on deep learning or machine learning, such as CLIP or SVM. CLIP, for example, can achieve zero-shot image classification by mapping images and text to a shared semantic space. The training can be supervised training, which uses labeled training data to learn the mapping relationship from input to output, enabling the model to make accurate predictions or classifications for new samples. The core idea of ​​SVM is to find an optimal hyperplane that maximizes the margin between the two classes of samples as the decision boundary, thus maximizing the generalization ability of the classification.

[0089] The feature vector of the first positive sample image mentioned above can be a semantic vector extracted from image data that matches the training user's intent.

[0090] The feature vector of the initial query text mentioned above can be a semantic vector extracted from the initial query text of the training user.

[0091] The aforementioned initial center vector can be obtained by weighted averaging of the feature vector of the initial query text and the feature vector of the first positive sample image.

[0092] In one possible implementation, for example, if the number of first positive samples is n, then the feature vectors of the n first positive samples are summed, and the feature vector Q1 of the initial query text is added. The sum is then divided by (n+1) to obtain the average value, which is the initial semantic center vector T1.

[0093] The aforementioned query intent group can be a user's query intent group constructed based on the initial classification model and the initial center vector.

[0094] It should be noted that the number of samples labeled at one time can be a minimum of 5 and a maximum of 200. Considering performance and user experience, it can be limited to 200 to ensure sample quality and labeling efficiency.

[0095] Optionally, in the step of obtaining at least one query intent group of the user, the following steps can be taken: obtaining the user's subsequent query text and the third image query result of the subsequent query text; receiving the user's second feedback result for the third image query result; determining the query intent group to be updated corresponding to the subsequent query text based on the similarity between the feature vector of the subsequent query text and the semantic center vector in the existing query intent group; training the previously trained classification model based on the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, to obtain an updated classification model; generating an updated center vector based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated; and updating the user's query intent group based on the updated classification model and the updated center vector.

[0096] In this embodiment of the invention, the subsequent query text can be the query text that the user is trained to enter again for searching. Specifically, it can be the query text under the same or similar semantic topics relative to the initial query text, such as sailboats, beanbag chairs, etc.

[0097] The third image query result mentioned above includes image data retrieved based on subsequent query text.

[0098] The second feedback result mentioned above includes the second positive sample image and the second negative sample image selected by the user. The second positive sample image is the image data in the third image query result that matches the training user's intention, and the second negative sample image is the image data in the third image query result that does not match the training user's intention.

[0099] Cosine similarity can be used to calculate the similarity between the feature vector of the subsequent query text and the semantic center vector in the existing query intent group. Query intent groups with similarity greater than a preset similarity threshold are identified as query intent groups to be updated corresponding to the subsequent query text.

[0100] The query intent groups mentioned above that need to be updated are the query intent groups that need to be updated.

[0101] The classification model obtained from the previous training can be an existing classification model trained before subsequent text queries.

[0102] The above training can be a process of training the classification model obtained from the previous training using the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated.

[0103] It should be noted that the training process can be executed using an asynchronous task queue, without blocking the main retrieval link, ensuring that users can use other functions normally. The default timeout for asynchronous processing can be set to 60 seconds, etc., after which it will automatically retry. A training completion notification will be automatically pushed to the user upon completion.

[0104] The updated classification model described above can be obtained by training the previously trained classification model using the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated.

[0105] The feature vector of the subsequent query text can be the feature vector extracted from the subsequent query text.

[0106] The semantic center vector of the above-mentioned query intent group to be updated can be the semantic center feature extracted from the query intent to be updated.

[0107] The updated center vector can be a new query semantic center vector generated based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated.

[0108] Furthermore, the user's query intent group can be updated based on the updated classification model and the updated center vector. The updated classification model and the updated center vector are written into the corresponding query intent group, replacing the original classification model and semantic center vector of the query intent group, so that the semantic matching and classification of the query intent group can be more accurate in subsequent searches.

[0109] It should be noted that the classification model file obtained from training is not bound to a single user, but is shared by all users, which can improve model utilization. At the same time, the training users of the model are tracked, and the system user ID of the training users is recorded. The user ID is generated automatically by the database when the user is created and is used for subsequent model tracking.

[0110] Optionally, in the step of training the previously trained classification model based on the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the most recent preset number of positive sample images and negative sample images can be taken in chronological order from the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated as training data; the previously trained classification model can be trained using the training data to obtain the updated classification model.

[0111] In this embodiment of the invention, the aforementioned historical positive sample images may be images from the query intent group to be updated, which were labeled by the training users in past feedback results as matching the user's search intent.

[0112] The aforementioned historical negative sample images can be images from the query intent group to be updated, which were labeled by the training users in past feedback results as not conforming to the user's search intent.

[0113] The above time order can be sorted from newest to oldest samples by timestamp. This time order represents the chronological order in which users labeled the positive and negative sample images, with samples labeled with later timestamps considered the newest. Specifically, a timestamp can be recorded for each user feedback result, and the newness of the samples can be determined based on the order of the timestamps: the later the timestamp, the newer the sample image; the earlier the timestamp, the older the sample image.

[0114] The aforementioned preset number can be a pre-defined number of samples, used to limit the number of positive and negative sample images that can be used when training the classification model. This can be a minimum of 5, a maximum of 200, etc. When the number of sample images exceeds the preset number, the latest labeled sample image is used, and older labeled sample images are discarded, thus balancing storage, computational costs, and model timeliness.

[0115] The classification model obtained from the previous training can be an existing classification model trained before subsequent text queries.

[0116] The most recent preset number of positive and negative sample images can be taken from the second positive sample image, the second negative sample image, and the historical positive and negative sample images corresponding to the query intent group to be updated, in chronological order, as training data. The classification model obtained from the previous training can be trained using the training data to obtain the updated classification model.

[0117] The above training can be a process of incremental training or retraining the classification model obtained from the previous training using training data. Incremental training involves training the model using newly added sample images from the training data, rather than retraining all data from scratch, to adapt to new information while retaining old knowledge. Retraining involves merging historical samples with new samples from the training data and training the previously trained classification model from scratch, without relying on the parameters of the previously trained model.

[0118] The updated classification model described above can be a new classification model obtained by training the previously trained classification model using training data.

[0119] It should be noted that when updating the classification model, instead of using all new positive and second sample images and historical samples indefinitely, a predetermined number of positive and negative sample images are selected in chronological order from the second positive and second negative sample images, as well as historical positive and negative sample images, as training data to train the classification model. This can effectively adapt to intent drift and maintain training efficiency.

[0120] Optionally, in the step of generating the updated center vector based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated, the cumulative training times of the query intent group to be updated can be obtained; based on the cumulative training times of the query intent group to be updated, the semantic center vector of the query intent group to be updated is fused with the feature vector of the subsequent query text to obtain the updated center vector.

[0121] In this embodiment of the invention, the cumulative number of training iterations can be the total number of effective training iterations that have been completed from the start of the query intent group construction to the current moment.

[0122] The above fusion can be a process of merging the semantic center vector of the query intent group to be updated with the feature vector of the subsequent query text according to certain weighting rules into an updated center vector.

[0123] The semantic center vector of the query intent group to be updated can be fused with the feature vector of the subsequent query text using an online fusion formula to obtain the updated center vector T. n The fusion formula is: T n = (T n ₋1 × (n−1) + Q n ) / n

[0124] Among them, T n For the updated center vector, T n ₋1 represents the query semantic center vector of the query intent group to be updated, Q n Let be the feature vector of the subsequent query text, and n be the cumulative training iterations of the query intent group to be updated, where n ≥ 2. This formula enables the dynamic convergence of the semantic center vector of the query text, gradually approximating the user's true search intent.

[0125] The aforementioned online fusion can be a process in which the feature vectors of subsequent query texts are dynamically merged with the semantic center vectors of the query intent group to be updated according to an online fusion formula during real-time operation, thereby updating the semantic center vectors of the query intent group to be updated.

[0126] Understandably, the cumulative training count of the query intent group to be updated can be obtained, and based on the cumulative training count of the query intent group to be updated, the semantic center vector of the query intent group to be updated can be fused with the feature vector of the subsequent query text to obtain the updated center vector. Through the online vector fusion mechanism, the query semantic center vector gradually converges with multiple rounds of user feedback, effectively solving the problem that the query vector is fixed in the existing technology and cannot adapt to the user's real intent, thus improving the personalized adaptation capability.

[0127] Optionally, after obtaining the latest query intent group, the latest classification model can be serialized and persistently stored in a distributed file storage medium; the metadata of the latest query intent group can be written to the database, including: query intent group ID, query semantic center vector, cumulative training times, number and identifier of current positive and negative sample images, storage path of the classification model, training time, last call time, similarity matching threshold, classification weight configuration, and system user ID of the training user; the query semantic center vector can be cached in the cache database.

[0128] In this embodiment of the invention, the latest classification model can be the latest classification model trained based on user feedback.

[0129] The serialization process described above can be the conversion of the latest classification model into a storable or transmissible format, such as using PKL format or byte streams. By serializing the latest classification model, it can be persistently saved to a distributed file system, shared across services, and deserialized and restored when needed.

[0130] The aforementioned distributed file storage carrier is a storage system that distributes files across multiple servers, providing highly available, highly reliable, and scalable persistent storage for the latest classification model. The distributed file storage carrier is used to store the serialized latest classification model, allowing multiple retrieval service instances to share the latest classification model and ensuring that the model is not lost during system restarts, upgrades, or expansions.

[0131] The aforementioned database is used to persistently store metadata for query intent groups, ensuring data reliability, manageability, and high performance for online retrieval.

[0132] The query intent group ID mentioned above can be a globally unique identifier for the query intent group.

[0133] The aforementioned query semantic center vector can be the latest query semantic center vector corresponding to the query intent group.

[0134] The cumulative training count mentioned above can be the total number of times the query intent group has been trained / updated from its initial construction to the present.

[0135] The number and identifier of the current positive and negative sample images mentioned above can be the number of positive and negative samples currently in use, as well as the identifiers of the positive and negative sample images, such as a list of image_ids or storage paths.

[0136] The storage path of the above classification model can be the complete path of the latest classification model in the distributed storage carrier.

[0137] The training time mentioned above can be the timestamp of the current training or update completion.

[0138] The last call time mentioned above could be the time when the query intent group was last matched and used in a query.

[0139] The aforementioned similarity matching threshold is used to determine whether the feature vector of the subsequent query text matches the similarity threshold of the semantic center vector in the query intent group.

[0140] The classification weight configuration described above can be a set of parameters used to control the degree of influence of the classification model. These classification weights can be used to assign different penalty coefficients to positive and negative samples during model training. The classification weight configuration can be personalized for the feedback quality and maturity of different intent groups, improving retrieval performance and user experience.

[0141] The user ID of the aforementioned training user can be a unique identifier for the training user.

[0142] Cache the query semantic center vector to a cache database to improve the efficiency of subsequent query matching.

[0143] The aforementioned cache database is used to store the semantic center vectors of query intent groups, providing data access with extremely low latency, thereby reducing the query pressure on the main database and improving the response speed of online retrieval.

[0144] It should be noted that the classification model adopts a permanent validity mechanism. Unless the total number of samples in the model exceeds the preset number, or manual intervention is performed to delete or invalidate the model, it will always remain in a valid state.

[0145] Optionally, in the step of calling the latest classification model in the target query intent group to classify and predict the first image query result and obtain the final image query result of the target query text, the latest classification model in the target query intent group can be called to classify and predict the first image query result and obtain the confidence score of each image data in the first image query result; based on the confidence score, each image data in the first image query result is filtered and sorted to obtain the final image query result of the target query text.

[0146] In this embodiment of the invention, the above classification prediction may be a process of calling the latest classification model in the target query intent group, analyzing each image in the first image query result, and outputting a confidence level indicating whether the image matches the user's search intent.

[0147] The confidence level mentioned above can be a quantitative value representing the degree of certainty by which the latest classification model determines that each image in the first image query result matches the user's search intent. The value can range from 0 to 1, and the higher the confidence level, the more the image matches the user's search intent.

[0148] The filtering and sorting described above can be a process of selecting and rearranging images based on their confidence levels. Filtering can involve discarding images with confidence levels below a preset confidence threshold and retaining only those with confidence levels equal to or higher than the preset threshold. This preset confidence threshold can be a pre-defined value, such as 0.8. Sorting can involve rearranging the retained images from highest to lowest confidence level.

[0149] The final image query result can be a list of images returned to the querying user after filtering and sorting the image data in the first image query result according to the confidence level. Alternatively, the final image query result can be a list of images that have been removed from the latest classification model (images with confidence levels below a preset confidence threshold that do not meet the user's intent), and images with confidence levels above or equal to the preset confidence threshold are arranged in descending order of confidence level.

[0150] It should be noted that by calling the latest classification model in the target query intent group to classify and predict the first image query result, the confidence level of each image data in the first image query result is obtained. Based on the confidence level, each image data in the first image query result is filtered and sorted to obtain the final image query result of the target query text, thereby realizing personalized and accurate search result display.

[0151] In this embodiment of the invention, if the querying user is still not satisfied with the categorized search results, the method of the invention can be re-executed to continue labeling the current positive and negative samples, updating the query semantic center vector, calling the existing training algorithm to generate a new classification model, and achieving multiple rounds of iterative optimization until the search results meet the querying user's needs.

[0152] like Figure 2 As shown, this embodiment of the invention provides a text-to-image search device based on user feedback, which includes:

[0153] The acquisition module 201 is used to acquire the target query text of the querying user and the first image query result of the target query text, and to acquire at least one query intent group of the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the training user's historical multiple image query results.

[0154] The first processing module 202 is used to extract the target feature vector corresponding to the target query text and calculate the similarity between the target feature vector and the semantic center vector in the query intent group;

[0155] The classification prediction module 203 is used to, if there is a target query intent group with a similarity greater than a preset similarity threshold, call the latest classification model in the target query intent group to perform classification prediction on the first image query result, so as to obtain the final image query result of the target query text;

[0156] The second processing module 204 is used to take the first image query result as the final image query result of the target query text if there is no target query intent group with a similarity greater than a preset similarity threshold.

[0157] The return module 205 is used to return the final image query result to the querying user.

[0158] Optionally, the acquisition module 201 is further configured to acquire the initial query text of the training user and the second image query result of the initial query text, the second image query result including image data retrieved based on the initial query text; receive a first feedback result from the user regarding the second image query result, the first feedback result including a first positive sample image and a first negative sample image selected by the user, the first positive sample image being image data in the second image query result that conforms to the training user's intent, and the first negative sample image being image data in the second image query result that does not conform to the training user's intent; train an initial classification model based on the first positive sample image and the first negative sample image; generate an initial center vector based on the feature vector of the first positive sample image and the feature vector of the initial query text; and construct the user's query intent group based on the initial classification model and the initial center vector.

[0159] Optionally, the acquisition module 201 is further configured to acquire the subsequent query text of the training user and the third image query result of the subsequent query text, the third image query result including image data retrieved based on the subsequent query text; receive a second feedback result from the user regarding the third image query result, the second feedback result including a second positive sample image and a second negative sample image selected by the user, the second positive sample image being image data in the third image query result that conforms to the training user's intent, and the second negative sample image being image data in the third image query result that does not conform to the training user's intent; determine the query intent group to be updated corresponding to the subsequent query text based on the similarity between the feature vector of the subsequent query text and the semantic center vector in the existing query intent group; and train the previously trained classification model based on the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, to obtain an updated classification model.

[0160] Based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated, an updated center vector is generated; based on the updated classification model and the updated center vector, the user's query intent group is updated.

[0161] Optionally, the acquisition module 201 is further configured to take the most recent preset number of positive and negative sample images in chronological order from the second positive sample image, the second negative sample image, and the historical positive and negative sample images corresponding to the query intent group to be updated, as training data; and train the previously trained classification model using the training data to obtain the updated classification model.

[0162] Optionally, the acquisition module 201 is further configured to acquire the cumulative training times of the query intent group to be updated; based on the cumulative training times of the query intent group to be updated, the semantic center vector of the query intent group to be updated is fused with the feature vector of the subsequent query text to obtain the updated center vector.

[0163] Optionally, the device is further configured to serialize the latest classification model and persistently store it in a distributed file storage medium; write the metadata of the latest query intent group into a database, the metadata including: query intent group ID, query semantic center vector, cumulative training times, number and identifier of the current positive and negative sample images, storage path of the classification model, training time, last call time, similarity matching threshold, classification weight configuration, and system user ID of the training user; and cache the query semantic center vector in a cache database.

[0164] Optionally, the classification prediction module 202 is further configured to call the latest classification model in the target query intent group to perform classification prediction on the first image query result, and obtain the confidence level of each image data in the first image query result; based on the confidence level, filter and sort each image data in the first image query result to obtain the final image query result of the target query text.

[0165] like Figure 3 As shown, this embodiment of the invention also provides an electronic device, including a processor, which can execute any of the above-described text search image methods based on user feedback.

[0166] Specifically, it includes a processor 301 and a memory 302, as well as a computer program stored in the memory 302 and capable of running on the processor 301, which executes a text search method based on user feedback, wherein:

[0167] The processor 301 executes the calculator program based on the user feedback-based text search method stored in the memory 302, and performs the following steps:

[0168] Obtain the target query text of the querying user and the first image query result of the target query text, and obtain at least one query intent group of the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the historical multiple image query results of the training user.

[0169] Extract the target feature vector corresponding to the target query text, and calculate the similarity between the target feature vector and the semantic center vector in the query intent group;

[0170] If there exists a target query intent group with a similarity greater than a preset similarity threshold, then the latest classification model in the target query intent group is invoked to classify and predict the first image query result, thereby obtaining the final image query result of the target query text.

[0171] If there is no target query intent group with a similarity greater than the preset similarity threshold, then the first image query result is taken as the final image query result of the target query text.

[0172] The final image query result is returned to the querying user.

[0173] Optionally, the process of obtaining at least one group of query intents of the user executed by processor 301 includes:

[0174] Obtain the initial query text of the training user and the second image query result of the initial query text, wherein the second image query result includes image data retrieved based on the initial query text;

[0175] The system receives a first feedback result from the user regarding the second image query result. The first feedback result includes a first positive sample image and a first negative sample image selected by the user. The first positive sample image is image data in the second image query result that matches the intent of the training user, and the first negative sample image is image data in the second image query result that does not match the intent of the training user.

[0176] An initial classification model is trained based on the first positive sample image and the first negative sample image;

[0177] An initial center vector is generated based on the feature vector of the first positive sample image and the feature vector of the initial query text.

[0178] Based on the initial classification model and the initial center vector, construct the user's query intent group.

[0179] Optionally, the process of obtaining at least one group of query intents of the user executed by processor 301 further includes:

[0180] Obtain the subsequent query text of the training user and the third image query result of the subsequent query text, wherein the third image query result includes image data retrieved based on the subsequent query text;

[0181] The system receives a second feedback result from the user regarding the third image query result. The second feedback result includes a second positive sample image and a second negative sample image selected by the user. The second positive sample image is image data in the third image query result that matches the intent of the training user, and the second negative sample image is image data in the third image query result that does not match the intent of the training user.

[0182] Based on the similarity between the feature vector of the subsequent query text and the semantic center vector in the existing query intent group, the query intent group to be updated corresponding to the subsequent query text is determined.

[0183] Based on the second positive sample image and the second negative sample image, as well as the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the classification model obtained from the previous training is trained to obtain the updated classification model.

[0184] Based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated, an updated center vector is generated.

[0185] The user's query intent group is updated based on the updated classification model and the updated center vector.

[0186] Optionally, the processor 301 executes the step of training the previously trained classification model based on the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, to obtain an updated classification model, including:

[0187] In the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the most recent preset number of positive sample images and negative sample images are taken as training data in chronological order;

[0188] The previously trained classification model is trained using the training data to obtain an updated classification model.

[0189] Optionally, the step of generating an updated center vector based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated, performed by processor 301, includes:

[0190] Obtain the cumulative training count of the query intent group to be updated;

[0191] Based on the cumulative training times of the query intent group to be updated, the semantic center vector of the query intent group to be updated is fused with the feature vector of the subsequent query text to obtain the updated center vector.

[0192] Optionally, after obtaining the latest query intent group, the method executed by the processor 301 further includes:

[0193] The latest classification model is serialized and persistently stored on a distributed file storage medium.

[0194] Write the latest query intent group's metadata into the database. The metadata includes: query intent group ID, query semantic center vector, cumulative training times, number and label of the current positive and negative sample images, storage path of the classification model, training time, last call time, similarity matching threshold, classification weight configuration, and system user ID of the training user.

[0195] Cache the query semantic center vector to the cache database.

[0196] Optionally, the processor 301 executes the invocation of the latest classification model in the target query intent group to classify and predict the first image query result, thereby obtaining the final image query result of the target query text, including:

[0197] The latest classification model in the target query intent group is invoked to classify and predict the first image query result, thereby obtaining the confidence level of each image data in the first image query result;

[0198] Based on the confidence level, the image data in the first image query result are filtered and sorted to obtain the final image query result of the target query text.

[0199] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the user-feedback-based text search image method provided in this invention and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0200] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A text-to-image search method based on user feedback, characterized in that, The method includes the following steps: Obtain the target query text of the querying user and the first image query result of the target query text, and obtain at least one query intent group of the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the historical multiple image query results of the training user. Extract the target feature vector corresponding to the target query text, and calculate the similarity between the target feature vector and the semantic center vector in the query intent group; If there exists a target query intent group with a similarity greater than a preset similarity threshold, then the latest classification model in the target query intent group is invoked to classify and predict the first image query result, thereby obtaining the final image query result of the target query text. If there is no target query intent group with a similarity greater than the preset similarity threshold, then the first image query result is taken as the final image query result of the target query text. The final image query result is returned to the querying user.

2. The text search image method based on user feedback as described in claim 1, characterized in that, The step of obtaining at least one query intent group of the user includes: Obtain the initial query text of the training user and the second image query result of the initial query text, wherein the second image query result includes image data retrieved based on the initial query text; The system receives a first feedback result from the user regarding the second image query result. The first feedback result includes a first positive sample image and a first negative sample image selected by the user. The first positive sample image is image data in the second image query result that matches the intent of the training user, and the first negative sample image is image data in the second image query result that does not match the intent of the training user. An initial classification model is trained based on the first positive sample image and the first negative sample image; An initial center vector is generated based on the feature vector of the first positive sample image and the feature vector of the initial query text. Based on the initial classification model and the initial center vector, construct the user's query intent group.

3. The text search image method based on user feedback as described in claim 2, characterized in that, The step of obtaining at least one query intent group of the user further includes: Obtain the subsequent query text of the training user and the third image query result of the subsequent query text, wherein the third image query result includes image data retrieved based on the subsequent query text; The system receives a second feedback result from the user regarding the third image query result. The second feedback result includes a second positive sample image and a second negative sample image selected by the user. The second positive sample image is image data in the third image query result that matches the intent of the training user, and the second negative sample image is image data in the third image query result that does not match the intent of the training user. Based on the similarity between the feature vector of the subsequent query text and the semantic center vector in the existing query intent group, the query intent group to be updated corresponding to the subsequent query text is determined. Based on the second positive sample image and the second negative sample image, as well as the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the classification model obtained from the previous training is trained to obtain the updated classification model. Based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated, an updated center vector is generated. The user's query intent group is updated based on the updated classification model and the updated center vector.

4. The text search image method based on user feedback as described in claim 3, characterized in that, The step of training the previously trained classification model based on the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, to obtain the updated classification model, includes: In the second positive sample image, the second negative sample image, and the historical positive sample image and historical negative sample image corresponding to the query intent group to be updated, the most recent preset number of positive sample images and negative sample images are taken as training data in chronological order; The previously trained classification model is trained using the training data to obtain an updated classification model.

5. The text search image method based on user feedback as described in claim 3, characterized in that, The step of generating the updated center vector based on the feature vector of the subsequent query text and the semantic center vector of the query intent group to be updated includes: Obtain the cumulative training count of the query intent group to be updated; Based on the cumulative training times of the query intent group to be updated, the semantic center vector of the query intent group to be updated is fused with the feature vector of the subsequent query text to obtain the updated center vector.

6. The text search image method based on user feedback as described in claim 1, characterized in that, After obtaining the latest query intent group, the method further includes: The latest classification model is serialized and persistently stored on a distributed file storage medium. Write the latest query intent group's metadata into the database. The metadata includes: query intent group ID, query semantic center vector, cumulative training times, number and label of the current positive and negative sample images, storage path of the classification model, training time, last call time, similarity matching threshold, classification weight configuration, and system user ID of the training user. Cache the query semantic center vector to the cache database.

7. The text search image method based on user feedback as described in any one of claims 1 to 6, characterized in that, The step of calling the latest classification model in the target query intent group to classify and predict the first image query result to obtain the final image query result of the target query text includes: The latest classification model in the target query intent group is invoked to classify and predict the first image query result, thereby obtaining the confidence level of each image data in the first image query result; Based on the confidence level, the image data in the first image query result are filtered and sorted to obtain the final image query result of the target query text.

8. A text-to-image search device based on user feedback, characterized in that, The text search image device based on user feedback includes: The acquisition module is used to acquire the target query text of the querying user and the first image query result of the target query text, and to acquire at least one query intent group of the user. The query intent group includes a semantic center vector and the latest classification model. The query intent group is constructed and updated based on the feedback results of the training user's historical multiple image query results. The first processing module is used to extract the target feature vector corresponding to the target query text and calculate the similarity between the target feature vector and the semantic center vector in the query intent group; The classification prediction module is used to, if there is a target query intent group with a similarity greater than a preset similarity threshold, call the latest classification model in the target query intent group to classify and predict the first image query result, so as to obtain the final image query result of the target query text. The second processing module is used to take the first image query result as the final image query result of the target query text if there is no target query intent group with a similarity greater than a preset similarity threshold. The return module is used to return the final image query result to the querying user.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the text search image method based on user feedback as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the text search image method based on user feedback as described in any one of claims 1 to 7.

Images