Picture auditing method and device, medium and equipment

Abstract

The invention relates to the technical field of picture auditing, and particularly provides a picture auditing method and device, a medium and equipment, and the method can comprise the steps: carrying out the preliminary auditing of a to-be-audited picture in a current task, and obtaining a re-checked picture; determining an auditing threshold value based on the target recall rate corresponding to the current task; taking the auditing threshold value as a condition, utilizing a plurality of pre-trained picture reasoning models to recheck the rechecked picture, and obtaining a picture auditing result; wherein the picture auditing result represents whether the re-checked picture is abnormal or not. According to the embodiment of the invention, the picture auditing precision can be improved.

Description

Technical Field

[0001] This application relates to the field of image review technology, and more specifically, to a method, apparatus, medium, and device for image review. Background Technology

[0002] UGC platforms, or User Generated Content platforms, are online platforms that provide users with a space for content display, exchange, and interaction, centered around content created, published, and shared by ordinary internet users. As the core carrier of user-generated content, image review is a crucial aspect of content compliance management on UGC platforms. Image review is the process of verifying the compliance, security, and suitability of image content through technical means or manual methods.

[0003] Currently, UGC platforms typically use pre-trained review models for automated image review. However, in actual review scenarios, the proportion of inappropriate images is very low, and these images differ significantly from normal images. This leads to an imbalance in the ratio of positive to negative samples during model training, preventing the model from learning positive features and resulting in low review accuracy.

[0004] Therefore, how to provide a technical solution for image review with high accuracy has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of some embodiments of this application is to provide a method, apparatus, medium and device for image review, and the technical solutions of the embodiments of this application can improve the accuracy of image review.

[0006] In a first aspect, some embodiments of this application provide a method for image review, including: performing a preliminary review on images to be reviewed in the current task to obtain review images; determining a review threshold based on the target recall rate corresponding to the current task; using the review threshold as a condition, reviewing the review images using multiple pre-trained image inference models to obtain image review results; wherein, the image review results characterize whether the review images are abnormal.

[0007] Some embodiments of this application obtain review images after a preliminary review of the images to be reviewed. Then, a review threshold is determined based on the target recall rate, and the review images are reviewed based on multiple image inference models and the review threshold to obtain the image review result. These embodiments of the application can detect images to be reviewed from multiple perspectives through preliminary and review checks, improving the image review accuracy of the UGC platform.

[0008] In some embodiments, the plurality of image inference models are obtained through the following steps: constructing a training dataset and a validation dataset from the collected image sample data; wherein, the image sample data includes multiple image samples and the annotation results of each of the multiple images; the annotation results indicate whether each image is a normal image or an abnormal image; training a pre-constructed neural network model using the training dataset to obtain multiple neural network models to be validated; and selecting the plurality of image inference models from the multiple neural network models to be validated based on the validation dataset.

[0009] Some embodiments of this application train a pre-built neural network model by constructing a training dataset and a validation dataset, thereby determining multiple image reasoning models and providing model support for subsequent image review.

[0010] In some embodiments, after selecting the plurality of image inference models from the plurality of neural network models to be validated based on the validation dataset, the method further includes: obtaining an image classification threshold for each of the plurality of image inference models; calculating the model recall rate of each image inference model under the image classification threshold using the validation dataset; and generating a recall rate mapping table based on the image classification threshold and the model recall rate.

[0011] Some embodiments of this application construct a recall mapping table using the image classification threshold and model recall of each image inference model to provide data support for subsequent model inference.

[0012] In some embodiments, the preliminary review of the images to be reviewed in the current task to obtain review images includes: identifying the text content in the images to be reviewed; and using images in the blacklist that do not contain content matching the text content as the review images.

[0013] Some embodiments of this application identify the text content in the image to be reviewed and match it with a blacklist to determine the image to be reviewed, thereby achieving effective initial review of the image and providing input support for subsequent visual review by the model.

[0014] In some embodiments, determining the review threshold based on the target recall rate corresponding to the current task includes: when it is confirmed that the target recall rate exists in the recall rate mapping table corresponding to the plurality of image inference models, obtaining the review threshold corresponding to the target recall rate; when it is confirmed that the target recall rate does not exist in the recall rate mapping table, calculating the review threshold corresponding to the target recall rate using linear interpolation and the recall rate mapping table.

[0015] Some embodiments of this application determine the review threshold of multiple image reasoning models by using the target recall rate corresponding to the current task, thereby providing standardized data support for subsequent model reasoning.

[0016] In some embodiments, the step of reviewing the verification image using multiple pre-trained image inference models based on the review threshold to obtain image review results includes: inputting the verification image into each of the multiple image inference models to obtain each inference result output by each image inference model; comparing each inference result with the review threshold to obtain each review result of the verification image; and obtaining the image review result by analyzing the various review results.

[0017] Some embodiments of this application determine various inference results by reviewing images and various image inference models, and then compare each inference result with the review threshold to obtain each review result; the final image review result is selected from the various review results, thereby improving the accuracy of image review.

[0018] In some embodiments, obtaining the image review result by analyzing the various review results includes: determining the image review result from the various review results using a majority voting mechanism.

[0019] Some embodiments of this application use a majority voting mechanism to determine the final image review result, thereby ensuring the accuracy of image review.

[0020] Secondly, some embodiments of this application provide an image review apparatus, comprising: a preliminary review module for performing a preliminary review of images to be reviewed in the current task to obtain review images; a strategy module for determining a review threshold based on the target recall rate corresponding to the current task; and a review module for reviewing the review images using multiple pre-trained image inference models with the review threshold as a condition to obtain image review results; wherein the image review results characterize whether the review images are abnormal.

[0021] Thirdly, some embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any embodiment of the first aspect.

[0022] Fourthly, some embodiments of this application 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, when executing the program, can implement the method as described in any embodiment of the first aspect.

[0023] Fifthly, some embodiments of this application provide a computer program product, the computer program product including a computer program, wherein the computer program, when executed by a processor, can implement the method described in any embodiment of the first aspect. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of some embodiments of this application, the accompanying drawings used in some embodiments of this application will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 System diagrams for image review provided for some embodiments of this application; Figure 2 Flowcharts of a method for obtaining an image reasoning model provided for some embodiments of this application; Figure 3 A schematic diagram of the neural network model structure provided for some embodiments of this application; Figure 4 One of the flowcharts for image review methods provided in some embodiments of this application; Figure 5 The review threshold-recall mapping curves and dynamic inverse threshold calculation diagrams for different models provided for some embodiments of this application; Figure 6 Flowchart 2 of the image review method provided for some embodiments of this application Figure 7 Block diagrams of the apparatus for image verification provided for some embodiments of this application; Figure 8 A schematic diagram of an electronic device provided for some embodiments of this application. Detailed Implementation

[0026] The technical solutions of some embodiments of this application will now be described with reference to the accompanying drawings.

[0027] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] In related technologies, image review on UGC platforms generally includes both automated and manual review. In actual review scenarios, the proportion of violating images (positive examples) is extremely low, and they differ significantly from normal images. This results in insufficient features in the constructed training set, making it difficult for the model to learn positive example features during subsequent model training. Furthermore, automated review is affected by imbalanced samples and overfitting, leading to unstable accuracy and recall rates. Manual review is then necessary, but it is slow. Therefore, current technologies cannot simultaneously achieve both accuracy and efficiency in image review. Existing technologies typically only review visual features, lacking mechanisms for reviewing text content within images and a two-tiered linkage mechanism between visual review and text recognition, making it impossible to efficiently detect both text and visual violations simultaneously. Other technologies have proposed using multi-model fusion, but this approach is prone to fluctuations in voting results due to differences in classification thresholds. Moreover, it often uses simple averaging or direct voting, resulting in very low positive example recall rates, failing to guarantee the accuracy of image review.

[0029] As can be seen from the above-mentioned technologies, existing technologies cannot simultaneously address both visual and text content verification of images, and the accuracy of image verification is relatively low.

[0030] In view of this, some embodiments of this application provide a method for image review. This method first performs a preliminary review of the image to be reviewed at the text level to obtain images requiring further review. Then, after determining a review threshold based on the target recall rate of the current task, multiple image inference models and the review threshold are used to review the images again, resulting in an image review result. Some embodiments of this application can simultaneously address both visual and text content review of images, improving the accuracy of image review and demonstrating high practicality.

[0031] The following is in conjunction with the appendix Figure 1 The overall structure of the image review system provided by some embodiments of this application is illustrated by way of example.

[0032] like Figure 1 As shown in the figure, some embodiments of this application provide a system diagram for image review. The image review system may include a terminal 100 and a review server 200. The terminal 100 can communicate with the review server 200. The terminal 100 can send images to be reviewed for the current task to the review server 200; the images to be reviewed can be one or more. After receiving the images to be reviewed, the review server 200 first performs a preliminary review at the text level to obtain review images; the review images can be one or more. Then, the review server 200 determines a review threshold based on the target recall rate set for the current task; finally, the review server 200 uses multiple pre-deployed image inference models to infer the review images, and then combines this with the review threshold to determine the final image review result.

[0033] In some embodiments of this application, terminal 100 can be a mobile terminal or a non-portable computer terminal; review server 200 can be a local server or a cloud server; the embodiments of this application are not specifically limited here.

[0034] To achieve accurate image verification from a visual perspective, it is first necessary to train multiple image reasoning models. The following section combines... Figure 2 The present application provides an exemplary embodiment of the implementation process of the image acquisition reasoning model executed by the review server 200.

[0035] Please see the appendix Figure 2 , Figure 2 A flowchart illustrating a method for obtaining an image reasoning model, provided for some embodiments of this application.

[0036] S210, construct a training dataset and a validation dataset from the collected image sample data; wherein, the image sample data includes multiple image samples and the annotation results of each of the multiple images; the annotation results indicate whether each image is a normal image or an abnormal image.

[0037] For example, in a specific embodiment of this application, multiple image samples with timestamps are periodically extracted from the historical review data of the UGC platform; and a time decay factor is used as the sampling weight, with a higher sampling weight for times closer to the present. In actual collection, the sampling weight tends to be between 0.001 and 0.1 to ensure that the sampled data is recent and relatively new. Sampling weight = e (-衰减率×天数差) The attenuation rate can be set to 0.01. The image samples are obtained after preprocessing; the preprocessing operations include standardization, color space processing, and image adjustment. Specifically, the image resolution of all image samples is standardized to 108×192; the color space is unified to RGB; random rotation ±5°, brightness adjustment ±10%, and cropping range 0.9~1.0 are applied to improve robustness.

[0038] Next, multiple image samples are automatically or manually labeled to obtain the labeling results. A labeling result of 0 indicates that the image is normal (positive sample); a labeling result of 1 indicates that the image is in violation of regulations (abnormal image) (negative sample). The ratio of positive to negative samples can be set as needed, such as 1:1. Automatic labeling can be implemented using a pre-trained labeling model, which is not specifically limited in this embodiment. Finally, the labeled image sample data is divided to determine the training dataset, validation dataset, and test dataset (optional).

[0039] S220, the pre-built neural network model is trained using the training dataset to obtain multiple neural network models to be verified.

[0040] For example, in a specific embodiment of this application, a pre-built neural network model is as follows: Figure 3 As shown. Figure 3 The neural network model shown includes: an input layer, convolutional layers, flattened fully connected layers, and an output layer; the convolutional layers include convolutional kernels, normalization layers, pooling layers, and forgetting layers. The input layer has a size of 108×192×3; the convolutional kernels in each layer increase in size according to the filter order of 64, 128, 256, and 512, with kernel=3×3, stride=1, and padding='same'. Each layer is connected to a normalization layer (BN), a pooling layer (MaxPooling (2×2)), and a forgetting layer (Dropout (0.3)).

[0041] In training the neural network model using the training dataset, half-precision training and Dropout are employed to prevent overfitting. Half-precision training is an optimized training strategy that can significantly reduce resource consumption and improve training speed while maintaining performance. During training, each image is input into the neural network model to obtain the output result. The confidence and loss value of the neural network model are determined by combining the output result with the corresponding image annotation result. Using the optimizer ADMA, the weights of positive and negative samples are adjusted and the model is optimized based on the confidence and loss value to iterate the training of the neural network model again until an iteration threshold is reached (e.g., the iteration threshold is 50) or the monitoring metric reaches a preset condition (e.g., the confidence is greater than a set threshold), resulting in multiple neural network models to be validated.

[0042] S230, Based on the verification dataset, select the plurality of image reasoning models from the plurality of neural network models to be verified.

[0043] For example, in a specific embodiment of this application, after reaching the iteration threshold, the AUC index of multiple neural network models to be verified is calculated using a verification dataset. AUC is a core indicator for evaluating the performance of binary classification models; the closer the AUC value is to 1, the stronger the model's ability to distinguish between positive and negative samples; an AUC value around 0.5 indicates that the model has no actual distinguishing ability. Each image in the verification dataset is input into the neural network model to be verified, and its output is obtained. The AUC value of each neural network model to be verified is then determined by combining this with the annotation results in the verification dataset. Based on the AUC values, each neural network model to be verified is sorted from largest to smallest, and the top few (e.g., the top five) models are selected as multiple image inference models for subsequent review (e.g., selecting five image inference models).

[0044] In some embodiments of this application, after executing S230, the method for obtaining the image inference model further includes: obtaining the image classification threshold of each image inference model among the plurality of image inference models; calculating the model recall of each image inference model under the image classification threshold using the validation dataset; and generating a recall mapping table based on the image classification threshold and the model recall.

[0045] For example, in a specific embodiment of this application, multiple points are selected at equal intervals from 0 to 1 as candidate thresholds for each image inference model. The recall rate is calculated for each candidate threshold. If the probability of image prediction is greater than or equal to the candidate threshold, the image is predicted as positive; if the probability of image prediction is less than the threshold, the image is predicted as negative. The recall rate is calculated using the confusion matrix formula. The confusion matrix, also known as the classification result matrix, has four values: True Positive (TP): The model correctly predicts a positive sample as positive; False Positive (FP): The model incorrectly predicts a negative sample as positive (Type I error); True Negative (TN): The model correctly predicts a negative sample as negative; False Negative (FN): The model incorrectly predicts a positive sample as negative (Type II error); Recall rate = TP / (TP + FN).

[0046] Based on the above thresholds and recall rates, a recall rate mapping table is constructed, as shown in Table 1: Table 1

[0047] The following is in conjunction with the appendix Figure 4 The present application provides an exemplary embodiment of the image review process performed by the review server 200.

[0048] Please see the appendix Figure 4 , Figure 4 A flowchart of an image review method is provided for some embodiments of this application. The image review method may include: S410: Perform a preliminary review of the images pending review in the current task and obtain the images for verification.

[0049] For example, in a specific embodiment of this application, OCR technology or other text extraction models are used to review the text in the image to be reviewed in order to determine the image to be reviewed.

[0050] In some embodiments of this application, S410 may include: identifying the text content in the image to be reviewed; and using images in the blacklist that do not contain content matching the text content as the review images.

[0051] For example, in a specific embodiment of this application, OCR technology is used to extract text information (i.e., text content) from the image to be reviewed. The text information is then matched against a blacklist of violations. If it matches the blacklist, it is directly marked as a violation; otherwise, it is used as a review image. Alternatively, text detection technology can be used to detect whether the document information contains any sensitive or illegal text. If it does, it is directly marked as a violation; otherwise, it is used as a review image.

[0052] S420, Determine the review threshold based on the target recall rate corresponding to the current task.

[0053] For example, in a specific embodiment of this application, a target recall rate (e.g., 95%) can be set for the current task according to actual needs, or a target recall rate matching the task strategy of the current task can be selected. Different task strategies correspond to different target recall rates. The review threshold can be determined through the target recall rate.

[0054] In some embodiments of this application, S420 may include: when it is confirmed that the target recall rate exists in the recall rate mapping table corresponding to the plurality of image inference models, obtaining the review threshold corresponding to the target recall rate.

[0055] For example, in a specific embodiment of this application, the existence of an audit threshold corresponding to the target recall rate is searched in the recall rate mapping table constructed above. If it exists, the corresponding value is used as the audit threshold.

[0056] In some other embodiments of this application, S420 may include: when it is confirmed that the target recall rate does not exist in the recall rate mapping table, calculating the review threshold corresponding to the target recall rate using linear interpolation and the recall rate mapping table.

[0057] For example, in a specific embodiment of this application, if the recall rate mapping table does not contain an approval threshold corresponding to the target recall rate, then linear interpolation is needed to dynamically calculate the threshold under the target recall rate in conjunction with the recall rate mapping table. Figure 5 The diagram shows the review threshold (hereinafter referred to as threshold)-recall mapping curves and dynamic inverse threshold calculation for different models provided in this application. The diagram shows the thresholds for three models (i.e., Model 1, Model 2, and Model 3) at a recall rate of 88%, and the mapping curves they follow. The review threshold corresponding to the target recall rate can be solved using linear interpolation based on the curves in this diagram.

[0058] Since the target recall rate must have two adjacent sets of data points in the recall rate mapping table, namely (r1, t1) and (r2, t2), where r1 and r2 represent the recall rate and t1 and t2 represent the review thresholds corresponding to the recall rate, the target recall rate is calculated using the following formula based on linear interpolation. Corresponding review threshold ,Right now:

[0059] For example, When the recall rate is 0.85, according to the recall rate mapping table shown in Table 1, r1=0.88, r2=0.75, t1=0.4, t2=0.6, which can be calculated using the above formula. =0.446. This review threshold serves as the decision boundary for the model's inference in the current round, ensuring consistency in review thresholds across multiple models. This "dynamic threshold alignment" mechanism reduces voting oscillations, improves overall decision stability, and also ensures that the review threshold corresponds to recalls with business implications.

[0060] It is understood that, in addition to linear interpolation, other statistical algorithms can be used for calculation depending on the actual application scenario, and the embodiments of this application are not limited to this.

[0061] S430, using the review threshold as a condition, multiple pre-trained image reasoning models are used to review the review image to obtain the image review result; wherein, the image review result indicates whether the review image is abnormal.

[0062] For example, in a specific embodiment of this application, the review threshold is used as a reference standard for multiple image inference models to review the images, thereby achieving standard uniformity among multiple image inference models.

[0063] In some embodiments of this application, S430 may include: inputting the review image into each of the plurality of image inference models to obtain each inference result output by each image inference model; comparing each inference result with the review threshold to obtain each review result of the review image; and obtaining the image review result by analyzing the review results.

[0064] For example, in a specific embodiment of this application, the review image is input into the five image inference models obtained above, and the inference results output by each of these five image inference models are obtained. Each inference result represents the score of the review image. Then, each inference result is compared with the review threshold. If the score is higher than the review threshold, the review result is considered normal; otherwise, it is considered abnormal. Finally, the cases of normal or abnormal review results are analyzed to obtain the final image review result.

[0065] In some embodiments of this application, S430 may include: determining the image review result from the various review results using a majority voting mechanism.

[0066] For example, in a specific embodiment of this application, the five image reasoning models described above can produce five review results; two of these results indicate that the reviewed image is a normal image, and three results indicate that the reviewed image is an abnormal image (i.e., a violation image). In this case, using a majority voting mechanism can confirm that the image review result indicates that the reviewed image is abnormal. Alternatively, each review result can be manually reviewed to determine the final image review result.

[0067] Subsequently, the image review results can be stored in the corresponding result output module, which can then be used for manual review or for further training and optimization of the image reasoning model to improve the model's reasoning accuracy.

[0068] The following is in conjunction with the appendix Figure 6 The specific process of image review provided by some embodiments of this application is illustrated by way of example.

[0069] Please see the appendix Figure 6 , Figure 6 A flowchart illustrating an image review method provided for some embodiments of this application.

[0070] The above process is illustrated below by example.

[0071] S610 identifies the text content in the images to be reviewed in the current task.

[0072] S620: Determine if the text content matches the blacklist. If yes, execute S621; otherwise, execute S630.

[0073] S621 marks the image to be reviewed as a violation image.

[0074] S630, the image to be reviewed is used as the image for verification.

[0075] S640, determine the review threshold based on the target recall rate corresponding to the current task.

[0076] S650: Input the verification image into each of the multiple image reasoning models, and obtain the reasoning results output by each image reasoning model.

[0077] S660 compares each reasoning result with the review threshold to obtain the review results of each image to be reviewed.

[0078] S670 uses a majority voting mechanism to determine the image review result from the various review results.

[0079] Understandably, multiple image inference models are developed through... Figure 2 The method embodiments provided are used to obtain the data; the specific implementation process of S610~S670 can be referred to the method embodiments provided above. To avoid repetition, detailed descriptions are omitted here.

[0080] Please refer to Figure 7 , Figure 7 The diagram shows a block diagram of an image review apparatus provided in some embodiments of this application. It should be understood that this image review apparatus corresponds to the method embodiments described above and is capable of performing the various steps involved in the method embodiments. The specific functions of this image review apparatus can be found in the description above; detailed descriptions are omitted here to avoid repetition.

[0081] Figure 7 The image review device includes at least one software functional module that can be stored in a memory or embedded in the image review device in the form of software or firmware. The image review device includes: a preliminary review module 710, used to perform a preliminary review on the images to be reviewed in the current task and obtain review images; a strategy module 720, used to determine a review threshold based on the target recall rate corresponding to the current task; and a review module 730, used to review the review images using multiple pre-trained image inference models based on the review threshold and obtain image review results; wherein, the image review results indicate whether the review images are abnormal.

[0082] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the aforementioned method, and will not be elaborated further here.

[0083] Some embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can perform the operation of any of the methods corresponding to the methods provided in the above embodiments.

[0084] Some embodiments of this application also provide a computer program product, which includes a computer program, wherein when the computer program is executed by a processor, it can implement the operation of any of the methods corresponding to the above embodiments provided in the above embodiments.

[0085] like Figure 8 As shown, some embodiments of this application provide an electronic device 800, which includes a memory 810, a processor 820, and a computer program stored in the memory 810 and executable on the processor 820. When the processor 820 reads the program from the memory 810 via a bus 830 and executes the program, it can implement the methods of any of the above embodiments.

[0086] Processor 820 can process digital signals and can include various computing architectures. Examples include complex instruction set computer architectures, reduced instruction set computer architectures, or architectures that implement multiple instruction set combinations. In some examples, processor 820 can be a microprocessor.

[0087] The memory 810 can be used to store instructions executed by the processor 820 or data related to the execution of instructions. These instructions and / or data may include code for implementing some or all of the functions of one or more modules described in the embodiments of this application. The processor 820 of this disclosure embodiment can be used to execute the instructions in the memory 810 to implement the methods shown above. The memory 810 includes dynamic random access memory, static random access memory, flash memory, optical memory, or other memories well known to those skilled in the art.

[0088] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for image review, characterized in that, include: Perform a preliminary review of the images pending review in the current task, and obtain images for final review; Based on the target recall rate corresponding to the current task, determine the review threshold; Using the aforementioned review threshold as a condition, multiple pre-trained image reasoning models are used to review the review images to obtain image review results; wherein, the image review results indicate whether the review images are abnormal.

2. The method as described in claim 1, characterized in that, The multiple image inference models were obtained through the following steps: For the collected image sample data, a training dataset and a validation dataset are constructed; wherein, the image sample data includes multiple image samples and the annotation results of each of the multiple images; the annotation results indicate whether each image is a normal image or an abnormal image; The pre-built neural network model is trained using the training dataset to obtain multiple neural network models to be verified. Based on the verification dataset, the multiple image reasoning models are selected from the multiple neural network models to be verified.

3. The method as described in claim 2, characterized in that, After selecting the plurality of image inference models from the plurality of neural network models to be validated based on the validation dataset, the method further includes: Obtain the image classification threshold of each image inference model among the multiple image inference models; The model recall rate of each image inference model at the image classification threshold is calculated using the validation dataset. Based on the image classification threshold and the model recall rate, a recall rate mapping table is generated.

4. The method according to any one of claims 1-3, characterized in that, The preliminary review of the images to be reviewed in the current task, and the acquisition of images for further review, includes: Identify the text content in the image to be reviewed; Images that do not match the text content in the blacklist are used as the verification images.

5. The method as described in claim 3, characterized in that, The step of determining the review threshold based on the target recall rate corresponding to the current task includes: When it is confirmed that the target recall rate exists in the recall rate mapping table corresponding to the multiple image inference models, the review threshold corresponding to the target recall rate is obtained.

6. The method as described in claim 3, characterized in that, The step of determining the review threshold based on the target recall rate corresponding to the current task includes: When it is confirmed that the target recall rate does not exist in the recall rate mapping table, the review threshold corresponding to the target recall rate is calculated using linear interpolation and the recall rate mapping table.

7. The method according to any one of claims 1-3 and 5-6, characterized in that, The step of reviewing the images using multiple pre-trained image reasoning models based on the review threshold to obtain image review results includes: The verification image is input into each of the multiple image reasoning models to obtain the reasoning results output by each image reasoning model. Each reasoning result is compared with the review threshold to obtain the review results of the reviewed image; The image review result is obtained by analyzing each of the review results.

8. An image verification device, characterized in that, include: The preliminary review module is used to conduct a preliminary review of the images to be reviewed in the current task and obtain images for further review; The strategy module is used to determine the review threshold based on the target recall rate corresponding to the current task; The review module is used to review the review image using multiple pre-trained image inference models based on the review threshold, and obtain the image review result; wherein the image review result indicates whether the review image has any abnormalities.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is executed by a processor to perform the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the computer program is executed by the processor to perform the method as described in any one of claims 1-7.

Images