An AI label confidence interval threshold-based hot-updatable image machine review method
Patent Information
- Application Number
- CN202611065593.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-08-18
AI Technical Summary
[0003]1、规则与代码强耦合，规则调整需要发版上线，这样会使规则发版更新延迟大，规则一旦写错只能等下次发版回滚，运营响应慢；
[0045] The beneficial effects of this invention are: it not only greatly improves the efficiency of querying and response, but also effectively improves the accuracy of judgment and significantly reduces misjudgments.
Smart Images
Figure CN122594286A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of AI multimodal analysis and review technology, and in particular to a hot-updateable image review method based on AI tag confidence interval thresholds for compliance review of qualification image assessment materials, images and other images. Background Technology
[0002] Currently, similar image compliance review methods in the industry generally suffer from the following technical deficiencies:
[0003] 1. The rules are tightly coupled with the code. Rule adjustments require a release and deployment, which causes a large delay in rule release and updates. If a rule is written incorrectly, it can only be rolled back on the next release, resulting in slow operational response.
[0004] 2. Using a single threshold for confidence results in all labels sharing the same cutoff point. However, single-threshold segmentation leads to both high-confidence false positives and low-confidence false negatives, resulting in a high false positive rate.
[0005] 3. The lack of a unified aggregation strategy when an image contains multiple different risk labels leads to inconsistencies between machine review and manual review conclusions.
[0006] 4. Rule-based direct database access (refers to directly retrieving and locking / matching real-time inventory records in the database based on preset business rules (such as first-in-first-out, batch priority, etc.), skipping intermediate calculations or manual intervention steps / the system automatically applies the rules to the underlying data and directly returns or locks specific inventory rows). High QPS (the system's ability to respond to external query requests per unit of time) puts a heavy burden on the database.
[0007] The aforementioned technical deficiencies directly lead to a high misjudgment rate in image compliance review, long rule adjustment cycles, and fluctuating database load, making it impossible to meet the business needs of high-concurrency image material compliance verification in fund matching scenarios. Summary of the Invention
[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0009] A method for machine-based image review that can be hot-updated is provided based on AI label confidence interval thresholds, the steps of which include:
[0010] S1. Create an audit rule configuration table in the database and cache the audit rule configuration table to generate a cache table;
[0011] S2. Receive the review task and the corresponding materials to be reviewed uploaded by the client, and put the review task into the message queue; after receiving the review task, the unified asynchronous listening entry triggers the main review process to review them one by one.
[0012] S3. Load the cache table and query the cache table according to the type of the material to be reviewed:
[0013] If the type of the material to be reviewed can be located in the cache table and the corresponding rule type can be found, then proceed directly to step S4;
[0014] If the type of the material to be reviewed cannot be located in the cache table, the review rule configuration table is read, the rule type of the material to be reviewed is added to the review rule configuration table, and the new rule type is written to the cache table.
[0015] S4. Obtain cloud review information for the materials to be reviewed. The cloud review information includes the cloud judgment result, the current risk label, and the current confidence level value.
[0016] S5. Make a judgment based on the cloud review information and return the machine review result;
[0017] S5.1 If the cloud-based assessment result is "pass", then the material to be reviewed is directly deemed "passed".
[0018] S5.2 If the cloud-based judgment result is a violation but there is no specific risk label in the cloud-based review information, then the material to be reviewed needs to be manually reviewed.
[0019] S5.3 If the cloud-based judgment result is a violation and has a risk label, then obtain all risk labels in the cloud-based review information and their corresponding current confidence level data, and judge each risk label one by one to obtain the final machine-based judgment result of the materials to be reviewed;
[0020] S5.3.1 If the current risk label cannot be located in the cache table, the current risk label is considered risk-free and is directly determined to be passed;
[0021] S5.3.2 If the current risk label can be located in the cache table, then obtain the partition threshold corresponding to the current risk label from the cache table and make a judgment on the current risk label;
[0022] S5.3.2.1 When there is only one current risk label in the material to be reviewed, the current confidence level is directly compared with the partition threshold to determine the status of the current risk label;
[0023] S5.3.2.2 When the materials to be reviewed have at least two current risk labels, the current confidence level of the current risk label is compared with the corresponding partition threshold one by one to determine the review status of the materials to be reviewed. According to the principle of strict deliberation, the judgment result is monotonically increased along the risk level of interception > review > pass to obtain the final machine review judgment result.
[0024] S5.4 generates machine review result information for the materials to be reviewed and saves the machine review result information to the database.
[0025] In a preferred embodiment of the present invention, the audit rule configuration table includes:
[0026] (a) Rule type;
[0027] (b) Rule label coding: There is a one-to-one correspondence between rule label codes and risk labels to establish a mapping from rule label codes to risk labels;
[0028] (c) Partition thresholds: Each rule label code corresponds to a set of partition thresholds. Each set of partition thresholds includes a pass threshold, a review threshold, and an interception threshold. The three thresholds divide the continuous range of the confidence of the risk label into four partitions: machine review pass zone, manual review zone, machine review interception zone, and abnormal pending zone. The risk label judgments corresponding to the four partitions are pass, review, interception, and abnormal pending, respectively.
[0029] (d) Rule name: used to indicate the audit result.
[0030] In a preferred embodiment of the present invention, the audit rule configuration table presets at least one set of basic partition thresholds, which include the values of basic pass threshold, basic review threshold and basic block threshold.
[0031] In a preferred embodiment of the present invention, the cache table includes at least rule type information that is in an enabled state.
[0032] In a preferred embodiment of the present invention, the access address of the material to be reviewed is first automatically rewritten according to the test and production environment, and then the access address is sent to the cloud review service system.
[0033] In a preferred embodiment of the present invention, in step S5.3.1, a rule tag code corresponding to the current risk tag is added to the audit rule configuration table, and the new rule tag code is written to the cache table; the validity period of the rule tag code is set; wherein, the interval threshold of the rule tag code adopts the basic partition threshold.
[0034] In a preferred embodiment of the present invention, in S5.3.2.2, while determining the current risk label, the final machine review result of the materials to be reviewed is revised in real time according to the principle of strict deliberation. The specific steps include:
[0035] (a) If the current confidence level value is in the abnormal pending area: the final machine review result of the material to be reviewed is abnormal pending, and no further judgment of other risk labels will be performed;
[0036] (b) If the current confidence level value is in the machine review interception zone, the final machine review result of the material to be reviewed will be changed to interception and marked as suspected violation;
[0037] (c) If the current confidence level value is in the manual review zone: if there has been no previous interception judgment, the final machine review judgment result of the material to be reviewed shall be changed to review; if there has been a previous interception judgment, the final machine review judgment result of the material to be reviewed shall remain as interception.
[0038] (d) If the current confidence level is in the machine review pass zone: if there has been no previous interception or re-review, the final machine review result of the material to be reviewed will be changed to pass; if there has been previous interception or re-review, the final machine review result of the material to be reviewed will be changed to the judgment result with the highest risk level.
[0039] In a preferred embodiment of the present invention, in S5.3.2.2, the determination of all current risk labels is completed first, and then the final institutional review result of the materials to be reviewed is determined according to the principle of strict deliberation. The specific steps include:
[0040] (1) Determine each current risk label individually:
[0041] If the current confidence level value is in the abnormal pending area, the current risk label is determined to be in an abnormal pending state; if the current confidence level value is in the machine review interception area, the current risk label is determined to be in an interception state; if the current confidence level value is in the manual review area, the current risk label is determined to be in a review state; if the current confidence level value is in the machine review passed area, the current risk label is determined to be in a passed state.
[0042] (2) Determine the final review result of the materials to be reviewed based on the principle of strict deliberation: obtain the judgment result of each current risk label, select the judgment result with the highest risk level, and take the judgment result as the final review result of the materials to be reviewed.
[0043] In a preferred embodiment of the present invention, in step S5.4, the machine review result information includes: the final machine review result of the material to be reviewed, the suspected violation note, the rule name, the display text, and the cloud review result.
[0044] In a preferred embodiment of the present invention, the historical judgment status of each risk label is obtained and viewed according to a preset period, and the value of the partition threshold is adjusted according to the actual usage requirements and saved to the cache table; the adjusted value is verified and stored in the audit rule configuration table.
[0045] The beneficial effects of this invention are: it not only greatly improves the efficiency of querying and response, but also effectively improves the accuracy of judgment and significantly reduces misjudgments. Attached Figure Description
[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:
[0047] Figure 1 This is a schematic diagram of the three-threshold machine review decision logic of a preferred embodiment of a hot-updateable image machine review method based on AI tag confidence interval threshold of the present invention;
[0048] Figure 2 This is a diagram illustrating the rule caching and hot update data flow of the present invention. Detailed Implementation
[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0050] Please see Figure 1-2 The embodiments of the present invention include:
[0051] A hot-updateable image review method based on AI label confidence interval thresholds addresses three core technical pain points: high misjudgment rate with single threshold, slow rule adjustment, and inconsistent decision-making criteria with multiple labels. It includes the following steps.
[0052] S1. Rule Model Design: Create an audit rule configuration table in the database and cache the audit rule configuration table to generate a cache table.
[0053] Each rule in the audit rule configuration table describes how a risk label is determined.
[0054] The cache table only records rule type information that is in an enabled state.
[0055] Furthermore, the audit rules configuration table is cached using Redis.
[0056] The audit rules configuration table includes the following:
[0057] (a) Rule type (type of materials to be reviewed): Rule types mainly include dynamic text, dynamic images, comment text, nickname text, avatar images, etc. By presetting different rule types, the same set of rule models can be reused in multiple scenarios and isolated from each other.
[0058] (b) Rule tag coding: The audit rule configuration table has multiple rule tag codes. Each rule tag code corresponds one-to-one with the risk tag returned by the cloud-based artificial intelligence detection service system to establish a mapping from rule tag code to risk tag.
[0059] (d) Partition thresholds: Each rule label code corresponds to a set of partition thresholds. Each set of partition thresholds includes a pass threshold, a review threshold, and an interception threshold. The three thresholds divide the continuous confidence interval of the risk label into four partitions: “machine review pass zone”, “manual review zone”, “machine review interception zone”, and “abnormal pending zone”. The judgments corresponding to the four partitions are pass, review, interception, and abnormal pending, respectively.
[0060] The continuous confidence interval includes the minimum confidence value and the maximum confidence value.
[0061] For example, when the confidence level of a risk label is in the continuous range of [0, 100], the pass threshold is 20, the review threshold is 60, and the interception threshold is 80, then the range of the machine review pass area is [0, 20), the range of the manual review area is [20, 60), the range of the machine review interception area is [60, 80), and the range of the abnormal pending area is [80, 100].
[0062] Furthermore, the review rule configuration table can also preset at least one set of basic values for the pass threshold, review threshold, and interception threshold, so that they can be directly called when loading new rule tag codes later.
[0063] For example, the base values for the threshold, review threshold, and interception threshold are 30, 60, and 80, respectively.
[0064] (c) Rule Name: Used to display the review results. Generally, the rule name is in Chinese, such as "Violation" or "Approved".
[0065] (e) Enable switch: Supports rule-based grayscale online and offline.
[0066] (f) Processing method: used for discretization of text-based rules.
[0067] Furthermore, all entries in the rule review rule configuration table are logically deleted to prevent the loss of historical data.
[0068] S2. Receive the review task and the corresponding materials to be reviewed uploaded by the client, and put the review task into the message queue; after receiving the review task, the unified asynchronous listening entry triggers the main review process to review them one by one, so that the process of submitting materials to be reviewed by the user is not blocked by the time-consuming cloud review.
[0069] Because image review requires access to a cloud-based AI service system, with each request typically taking hundreds of milliseconds, review tasks are uniformly triggered asynchronously via a message queue to avoid blocking the main user submission process. The system uses a unified asynchronous listening entry to receive review tasks and distribute them to the corresponding review workflows. It categorizes tasks into basic image review and headshot review based on business needs. Then, it sets corresponding detection scenario parameters according to the priority of each review task and passes these parameters to the cloud-based AI service system to detect different risk tags.
[0070] The cloud-based artificial intelligence service system can adopt existing technologies in this field. This application only calls this function and does not involve any improvement to it. Therefore, its principle and structure will not be described in detail here.
[0071] S3. Load / retrieve the cache table and query the cache table according to the type of the material to be reviewed (automatically generated when the user uploads the material):
[0072] If the type of the material to be reviewed can be located in the cache table and the corresponding rule type can be found, then proceed directly to step S4;
[0073] If the type of the material to be reviewed cannot be located in the cache table, the source (database) is retrieved to read the review rule configuration table, a new rule type for the material to be reviewed is created / loaded in the review rule configuration table, and the new rule type is written to the cache table.
[0074] S4. Rewrite the materials to be reviewed to an address accessible by the cloud review service system (the address is automatically rewritten based on the test and production environments), call the cloud review service system, and obtain the cloud review information of the materials to be reviewed. The cloud review information includes the cloud judgment result (violation, approval), the current risk label, and the current confidence level value.
[0075] The cloud-based review information is generated and returned by the cloud-based artificial intelligence service system, and this application only obtains and calls upon it.
[0076] S5. Perform an overall short circuit based on the cloud review information (only the overall judgment result needs to be obtained, and the specific judgment steps are not required) and return the machine review result.
[0077] S5.1 If the cloud-based assessment result is "pass", then the material to be reviewed is directly deemed "passed".
[0078] S5.2 If the cloud-based judgment result is a violation but there is no specific risk label in the cloud-based review information, then according to the security principle of "if the risk point cannot be located, manual review is required", the material to be reviewed needs to be manually reviewed, and the review material is sent directly to the reviewer.
[0079] S5.3. For all other cases except S5.1 and S5.2 (where the cloud-based assessment result is a violation and carries a risk label), all risk labels in the cloud-based review information and their corresponding current confidence level data are obtained. Each risk label is then assessed to obtain the final machine-based assessment result of the materials to be reviewed.
[0080] S5.3.1 If the current risk label cannot be located in the cache table for the corresponding rule label code, the current risk label is considered to be risk-free and is directly determined to be passed.
[0081] Create / load the current risk label in the audit rule configuration table, rebuild the mapping, write the new rule label code to the cache table, and set the validity period of the rule label code to achieve lazy loading and automatic backfilling, thereby reducing the pressure of rule queries on the database.
[0082] Furthermore, the validity period of the rule label code can be set and modified as needed. For example, the validity period of the rule label code can be set to one day.
[0083] Furthermore, the threshold values for the partitions corresponding to the newly created rule label codes can initially be the preset base values in the audit rule configuration table, and the specific values of the three thresholds can be adjusted as needed later.
[0084] S5.3.2 If the current risk label of the material to be reviewed can be located in the cache table, the corresponding rule label code is obtained from the cache table based on the current risk label, and the current risk label is judged.
[0085] S5.3.2.1) When there is only one current risk label in the material to be reviewed, the current confidence level is directly compared with the partition threshold to determine the status of the current risk label.
[0086] S5.3.2.2) When the material to be reviewed has at least two current risk labels, compare the current confidence level of each current risk label with the corresponding partition threshold to determine the review status of the material to be reviewed. Under the principle of strictest consensus, the judgment result is monotonically increased in the direction of risk level "interception > review > pass" through status guard. The final conclusion is the strictest judgment among all labels.
[0087] Specifically:
[0088] Once a risk label is determined to be blocked, the current determination of the material to be reviewed is blocked, and the status will not be downgraded in the future. That is, the determination results of subsequent risk labels will no longer affect the determination result of "blocked". Therefore, the final determination of the material to be reviewed is blocked.
[0089] Once a risk label is determined to be under review, the current determination of the material to be reviewed is under review, and the subsequent "review" determination can only be overridden by the "block" determination; if no subsequent risk label is blocked, the final determination of the material to be reviewed is under review.
[0090] In some embodiments of this application, the final institutional review result of the materials to be reviewed can be revised in real time according to the principle of strict deliberation when the current risk label is determined one by one. The specific steps include:
[0091] (a) If the current confidence level is higher than the interception threshold and less than or equal to the maximum confidence level, the current risk label will be classified into the abnormal pending area. The final machine judgment result of the material to be reviewed will be abnormal pending, and no further judgments of other risk labels will be executed. That is, the current judgment cannot be rewritten to prevent abnormal noise labels from covering the previously effective judgments.
[0092] (b) If the current confidence level is between the review threshold and the interception threshold, the current risk label will be placed in the machine review interception area, and the final machine review result of the materials to be reviewed will be interception and marked as suspected violation.
[0093] (c) If the current confidence level is between the pass threshold and the review threshold: the current risk label is assigned to the manual review area: if there has been no previous interception judgment, the final machine review result of the material to be reviewed is review; if there has been a previous interception judgment, the final machine review result of the material to be reviewed is still interception, so as to avoid downgrading a more severe judgment.
[0094] (d) If the current confidence level is lower than the pass threshold but greater than or equal to the minimum confidence level: the current risk label is assigned to the machine review pass zone: if there has been no previous interception or re-examination, the final machine review result of the material to be reviewed is pass; if there has been previous interception or re-examination, the final machine review result of the material to be reviewed will still maintain the highest risk level judgment result.
[0095] In other embodiments of this application, after all current risk labels have been determined, the final institutional review result of the materials to be reviewed can be determined according to the principle of rigorous deliberation. Specific steps include:
[0096] (1) Determine each current risk label individually:
[0097] (a) If the current confidence level is higher than the interception threshold and less than or equal to the maximum confidence level, then the current risk label is classified into the abnormal pending area and the current risk label is determined to be in an abnormal pending state.
[0098] (b) If the current confidence level is between the review threshold and the interception threshold: the current risk label is classified into the machine review interception area and the current risk label is determined to be in an interception state;
[0099] (c) If the current confidence level is between the pass threshold and the review threshold: the current risk label is assigned to the manual review area and the current risk label is determined to be in review status;
[0100] (d) If the current confidence level is lower than the pass threshold but greater than or equal to the minimum confidence level, then the current risk label is classified into the machine review pass zone and the current risk label is determined to be in the pass status.
[0101] (2) Determine the final review result of the materials to be reviewed based on the principle of strict deliberation: obtain the judgment result of each current risk label, select the judgment result with the highest risk level, and take the judgment result as the final review result of the materials to be reviewed.
[0102] S5.4 generates and returns the machine review result information for the material to be reviewed. The machine review result information includes: the final machine review result of the material to be reviewed, the suspected violation remarks, the rule name, the display text, and the cloud review result. Among them, the display text is concatenated according to the "risk label (confidence level)" format.
[0103] The machine review results are written back to the business image table and saved to the database. This allows the downstream review platform to simultaneously present the final judgment, the hit tags, and the corresponding confidence levels, facilitating subsequent offline parameter adjustment.
[0104] S6. Hot rule update: The system retrieves and views the historical judgment status of each risk label according to a preset period, adjusts the value of the partition threshold according to actual usage needs, and saves it to the cache table (replacing the original value); the system verifies and stores the audit rule configuration table in the database, so that the next audit request will automatically return to the source and fill in the new rule. The whole process takes effect in seconds and does not require restarting the service.
[0105] For example, when verifying the adjusted partition thresholds, the system can perform verification from the following aspects (at least one of them must be verified):
[0106] (a) Determine whether each threshold is within a preset confidence level continuous interval;
[0107] (b) Whether the newly added risk label is legal;
[0108] (c) Whether the partition threshold value, newly added risk label, and newly added rule label code are duplicated.
[0109] When reviewing historical judgments, we mainly look at four indicators: consistency rate, misjudgment rate, missed judgment rate, and review rate. This can guide threshold adjustments in reverse, forming a closed loop of "online data - offline evaluation - online hot update".
[0110] For example, when operations finds that marketing ad tags trigger too many interceptions, the original review threshold is raised from 60 to 70. Furthermore, rule changes during hot rule updates employ a "write-after-expires" strategy. After operations modify and save the rule in the backend, the system first verifies the rule's existence and writes it directly to the database. Then, it only clears the corresponding cache without rebuilding the cache during the write process. The system only triggers a back-to-origin request and cache regeneration when the next review request arrives. This strategy ensures eventual consistency between the database and cache at a low cost, avoids the risk of inconsistency due to double writes, and allows rule adjustments to take effect without a restart.
[0111] Example 1
[0112] Step 1: The user uploads an ID photo on the client, and the business system puts the review task into the message queue; after receiving the task, the unified asynchronous listening entry triggers the main review process, so that the user submission process is not blocked by the time-consuming cloud review.
[0113] Step 2: Enter the image review process and load the rule set for the image scenario (directly read from the cache if a match is found). In this example, the rules cover three types of risk tags: marketing advertisements, brand logos, and QR codes. The corresponding partition thresholds (pass threshold, review threshold, and intercept threshold) for the three types of risk tags in the cache table are configured as follows: marketing advertisements (20, 60, 85), brand logos (10, 50, 80), and QR codes (20, 40, 70).
[0114] Step 3: Rewrite the image to a cloud-accessible address (automatically adapting to differences between testing and production environments), call the cloud audit service, and obtain an overall judgment of violation, returning three risk labels and their confidence levels: marketing advertisement 78.5, brand logo 65.2, and QR code 30.1.
[0115] Step 4: Initialize the state to "Pass" and evaluate the three labels sequentially:
[0116] (a) Marketing Ad 78.5 falls between its review threshold of 60 and interception threshold of 85, so it is set to "intercepted" and marked as suspected of violating the rules, and the text is displayed as "Marketing Ad (78.50)";
[0117] (b) Brand logo 65.2 falls between its review threshold of 50 and interception threshold of 80. It should have been judged as "intercepted" and is currently "intercepted". The status remains unchanged.
[0118] (c) QR code 30.1 falls between its pass threshold 20 and review threshold 40. It should be judged as "manual review", but the status guard recognizes that it is now "intercepted" in a stricter way, so it is not downgraded.
[0119] Step 5, the final conclusion is "interception", along with the suspected violation notes, three hit rules, the display text "Marketing Ads (78.50) Brand Logo (65.20) QR Code (30.10)" and the original cloud results, all of which are written back to the business image table; the downstream review platform can then simultaneously present the final judgment, hit tags and confidence level.
[0120] Step 6, rule hot update: The operations team found that the marketing and advertising tags triggered too many interceptions, so they raised the review threshold from 60 to 70 and saved it. After the system verifies and enters the data into the database, it clears the cache for this scenario. The next review request will automatically return to the source and fill in the new rule. The whole process takes effect in seconds and does not require restarting the service.
[0121] Step 7, Offline parameter adjustment: Based on historical review results, different threshold combinations are batch-fed back, and four indicators—human review consistency rate, misjudgment rate, missed judgment rate, and re-review rate—are calculated offline to guide threshold adjustment in reverse, forming a closed loop of "online data—offline evaluation—online hot update".
[0122] The beneficial effects of the hot-updateable image machine review method based on AI tag confidence interval threshold of the present invention are:
[0123] 1. By caching the audit rule configuration table, the QPS of rule queries no longer hits the database, and the average query time is reduced from 50~100ms to less than 10ms, which greatly improves the efficiency of queries;
[0124] 2. By setting up four-segment partitions to finely segment and judge different risk labels, the accuracy of the judgment can be effectively improved and the false judgment can be significantly reduced compared with a single threshold. Furthermore, the "strictest" strategy of multi-label deliberation has unified the output of machine review, and the consistency rate with manual review has been improved by about 25%.
[0125] 3. By implementing a rule-based hot update strategy, the effective latency of operational threshold adjustments has been reduced from "requiring a new release (several hours)" to "seconds," thus improving response efficiency;
[0126] 4. The original confidence level and the final conclusion are stored in the database simultaneously, providing complete backfeed data for offline threshold parameter tuning, thereby improving the accuracy and reliability of subsequent adjustments.
[0127] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for machine-based review of hot-updable images based on AI tag confidence interval thresholds, characterized in that the steps include: include: S1. Create an audit rule configuration table in the database and cache the audit rule configuration table to generate a cache table; S2. Receive the review task and the corresponding materials to be reviewed uploaded by the client, and put the review task into the message queue. A unified asynchronous listening entry point receives the review task and triggers the main review process to review each task one by one. S3. Load the cache table and query the cache table according to the type of the material to be reviewed: If the type of the material to be reviewed can be located in the cache table and the corresponding rule type can be found, then proceed directly to step S4; If the type of the material to be reviewed cannot be located in the cache table, the review rule configuration table is read, the rule type of the material to be reviewed is added to the review rule configuration table, and the new rule type is written to the cache table. S4. Obtain cloud review information for the materials to be reviewed. The cloud review information includes the cloud judgment result, the current risk label, and the current confidence level value. S5. Make a judgment based on the cloud review information and return the machine review result; S5.1 If the cloud-based assessment result is "pass", then the material to be reviewed is directly deemed "passed". S5.2 If the cloud-based judgment result is a violation but there is no specific risk label in the cloud-based review information, then the material to be reviewed needs to be manually reviewed. S5.3 If the cloud-based judgment result is a violation and has a risk label, then obtain all risk labels in the cloud-based review information and their corresponding current confidence level data, and judge each risk label one by one to obtain the final machine-based judgment result of the materials to be reviewed; S5.3.1 If the current risk label cannot be located in the cache table, the current risk label is considered risk-free and is directly determined to be passed; S5.3.2 If the current risk label can be located in the cache table, then obtain the partition threshold corresponding to the current risk label from the cache table and make a judgment on the current risk label; S5.3.2.1 When there is only one current risk label in the material to be reviewed, the current confidence level is directly compared with the partition threshold to determine the status of the current risk label; S5.3.2.2 When the materials to be reviewed have at least two current risk labels, the current confidence level of the current risk label is compared with the corresponding partition threshold one by one to determine the review status of the materials to be reviewed. According to the principle of strict deliberation, the judgment result is monotonically increased along the risk level of interception > review > pass to obtain the final machine review judgment result. S5.4 generates machine review result information for the materials to be reviewed and saves the machine review result information to the database.
2. The method for hot-updating image machine review based on AI tag confidence interval threshold as described in claim 1, characterized in that, The audit rules configuration table includes: (a) Rule type; (b) Rule label coding: There is a one-to-one correspondence between rule label codes and risk labels to establish a mapping from rule label codes to risk labels; (c) Partition thresholds: Each rule label code corresponds to a set of partition thresholds. Each set of partition thresholds includes a pass threshold, a review threshold, and an interception threshold. The three thresholds divide the continuous range of the confidence of the risk label into four partitions: machine review pass zone, manual review zone, machine review interception zone, and abnormal pending zone. The risk label judgments corresponding to the four partitions are pass, review, interception, and abnormal pending, respectively. (d) Rule name: used to indicate the audit result.
3. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, The audit rules configuration table pre-sets at least one set of basic partition thresholds, which include the values of the basic pass threshold, the basic review threshold, and the basic block threshold.
4. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, The cache table should include at least the rule type information that is enabled.
5. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, First, the access address of the materials to be reviewed is automatically rewritten based on the test and production environments, and then the access address is sent to the cloud review service system.
6. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, In step S5.3.1, add the rule tag code corresponding to the current risk tag to the audit rule configuration table, and write the new rule tag code to the cache table; set the validity period of the rule tag code; wherein, the interval threshold of the rule tag code adopts the basic partition threshold.
7. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, In S5.3.2.2, while determining the current risk label, the final machine review result of the materials to be reviewed is revised in real time according to the principle of strict deliberation. The specific steps include: (a) If the current confidence level value is in the abnormal pending area: the final machine review result of the material to be reviewed is abnormal pending, and no further judgment of other risk labels will be performed; (b) If the current confidence level value is in the machine review interception zone, the final machine review result of the material to be reviewed will be changed to interception and marked as suspected violation; (c) If the current confidence level value is in the manual review zone: if there has been no previous interception judgment, the final machine review judgment result of the material to be reviewed shall be changed to review; if there has been a previous interception judgment, the final machine review judgment result of the material to be reviewed shall remain as interception. (d) If the current confidence level is in the machine review pass zone: if there has been no previous interception or re-review, the final machine review result of the material to be reviewed will be changed to pass; if there has been previous interception or re-review, the final machine review result of the material to be reviewed will be changed to the judgment result with the highest risk level.
8. The method for hot-updating image machine review based on AI tag confidence interval threshold as described in claim 1, characterized in that, In S5.3.2.2, the determination of all current risk labels is completed first, and then the final review result of the materials to be reviewed is determined according to the principle of strict deliberation. The specific steps include: (1) Determine each current risk label individually: If the current confidence level value is in the abnormal pending area, the current risk label is determined to be in an abnormal pending state; if the current confidence level value is in the machine review interception area, the current risk label is determined to be in an interception state; if the current confidence level value is in the manual review area, the current risk label is determined to be in a review state; if the current confidence level value is in the machine review passed area, the current risk label is determined to be in a passed state. (2) Determine the final review result of the materials to be reviewed based on the principle of strict deliberation: obtain the judgment result of each current risk label, select the judgment result with the highest risk level, and take the judgment result as the final review result of the materials to be reviewed.
9. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, In step S5.4, the machine review result information includes: the final machine review result of the material to be reviewed, the suspected violation note, the rule name, the display text, and the cloud review result.
10. The method for machine review of hot-updable images based on AI tag confidence interval thresholds according to claim 1, characterized in that, The system retrieves and reviews the historical assessment results of each risk label according to a preset cycle, adjusts the values of the partition thresholds based on actual usage needs, and saves them to the cache table. After the adjusted values are verified, they are stored in the audit rule configuration table.