An artificial intelligence-based data annotation method and system

Abstract

The present invention is applicable to the field of data annotation technology, and provides a data annotation method and system based on artificial intelligence. The method includes: when the proportion of secondary annotation reminder frames triggered in a target time segment exceeds a threshold ratio, analyzing whether the current scene complexity level of the target time segment is higher than the corresponding threshold. If so, obtaining historical video processing records and historical annotation deviation feedback data. By introducing double statistical parameters of the slope mean of the trend curve and the mean of historical annotation deviations, the present invention constructs a confidence correction factor, realizes a dynamic adjustment mechanism for confidence parameters, and can effectively identify and correct the trend misjudgment offset problem accumulated by the system under similar scene complexities.

Description

Technical Field

[0001] The present invention belongs to the technical field of data annotation, and particularly relates to an artificial intelligence-based data annotation method and system. Background Art

[0002] In the existing data annotation technology, during the process of analyzing and processing data such as videos and images by an artificial intelligence model, a confidence parameter is usually generated for each frame to measure the credibility of the system's label recognition result. To improve the data annotation efficiency, the existing system generally sets a fixed confidence threshold. When the confidence of a certain frame is lower than this threshold, a "secondary annotation reminder" is automatically triggered, and this frame is marked as a suspected low-confidence frame that needs further manual confirmation. This method can assist manual verification to a certain extent, avoid obviously incorrect label results from being directly accepted, and thus improve the annotation quality.

[0003] However, the existing technology generally uses a static threshold strategy for judgment and fails to effectively model the law of system confidence fluctuation in different video scenarios. Especially in scenarios with complex structures or many interference factors, the annotation system is prone to the phenomenon of concentrated false alarms. However, the current system cannot dynamically identify whether there is a systematic deviation based on historical processing trends or manual correction feedback. This problem of lacking an adaptive correction mechanism is likely to cause the system to repeatedly trigger redundant reminders in similar scenarios, increasing the burden of manual verification. Moreover, in the case of continuous accumulation of false alarm trends, the accuracy of confidence judgment will also continuously decline, seriously affecting the overall efficiency and consistency of the annotation process. Summary of the Invention

[0004] The purpose of the present invention is to provide an artificial intelligence-based data annotation method and system, aiming to solve the problems raised in the background art.

[0005] The present invention is implemented as follows. An artificial intelligence-based data annotation method, the method includes:

[0006] When the proportion of frames triggering the secondary annotation reminder in the target time segment exceeds the threshold ratio, analyze whether the current scene complexity level of the target time segment is higher than the corresponding threshold. If so, obtain the historical video processing record and the historical annotation deviation feedback data;

[0007] Parse the historical video processing record, and screen out several historical comparison segment records that are in the same level range as the current scene complexity level and have a matching time length from it;

[0008] Parse each historical comparison segment record in turn, calculate the proportion of frames triggering the secondary annotation reminder corresponding to each of them, draw a trend curve in chronological order, and calculate the average slope of the trend curve;

[0009] Analyze the historical annotation deviation feedback data, obtain the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculate the mean value of all deviation values;

[0010] Based on the mean value of the slopes and the mean value of all deviation values, generate a confidence correction factor by weighting, and use it to adjust the confidence parameters of all frames within the target time segment.

[0011] As a further limitation of the technical solution of the embodiment of the present invention, the steps of sequentially analyzing each historical comparison segment record, calculating the proportion of the secondary annotation reminder frames corresponding to each of them, plotting a trend curve in chronological order, and calculating the mean value of the slope of the trend curve include:

[0012] Sequentially analyze each historical comparison segment record, calculate the ratio between the number of frames triggering the secondary annotation reminder in each historical comparison segment record and the corresponding total number of frames, and obtain the proportion of the secondary annotation reminder frames corresponding to each of them;

[0013] Taking the time axis as the horizontal axis and the proportion of the secondary annotation reminder frames as the vertical axis, plot the corresponding trend curve;

[0014] Calculate the mean value of the slope of the trend curve.

[0015] As a further limitation of the technical solution of the embodiment of the present invention, the steps of analyzing the historical annotation deviation feedback data, obtaining the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculating the mean value of all deviation values include:

[0016] Analyze the historical annotation deviation feedback data, and extract the local annotation deviation feedback data corresponding to each historical comparison segment record from it;

[0017] Sequentially analyze each local annotation deviation feedback data, and extract the initial annotation value and the correction result from it, where the initial annotation value refers to all the frames triggering the secondary annotation reminder in each historical comparison segment record, and the correction result refers to the number of frames that finally need to be reminded of the secondary annotation after manual confirmation or feedback;

[0018] Calculate the deviation value between each group of initial annotation values and the corresponding correction results, and statistically analyze the deviation values corresponding to all historical comparison segment records to obtain the mean value of the deviation values.

[0019] As a further limitation of the technical solution of the embodiment of the present invention, the steps of generating a confidence correction factor by weighting based on the mean value of the slopes and the mean value of all deviation values, and using it to adjust the confidence parameters of all frames within the target time segment include:

[0020] Call the confidence correction factor generation formula, and use the mean value of the slope of the trend curve and the mean value of all deviation values as input parameters to perform weighted calculation to generate a confidence correction factor;

[0021] Adjust the initial confidence parameters of all frames within the target time segment based on the confidence correction factor to obtain the corrected confidence parameters;

[0022] Determine whether the corrected confidence parameter of each frame is lower than the preset confidence threshold. If so, retain the secondary annotation reminder for this frame; if not, cancel the original secondary annotation reminder.

[0023] As a further limitation of the technical solution of the embodiment of the present invention, the confidence correction factor generation formula is: , where refers to the confidence correction factor, refers to the average slope of the trend curve, refers to the adjustment weight corresponding to the average slope, refers to the absolute value of the average of all deviation values, refers to the adjustment weight corresponding to the absolute value;

[0024] In the confidence correction factor generation formula: , where refers to the total number of local annotation deviation feedback data, where refers to the correction result of the th local annotation deviation feedback data, refers to the initial annotation value of the th local annotation deviation feedback data, refers to the deviation value between the initial annotation value and the corresponding correction result corresponding to the

[0025] An artificial intelligence-based data annotation system, the system includes: a data acquisition module, a data screening module, a slope calculation module, a deviation value calculation module, and a correction factor generation module, where:

[0026] The data acquisition module is used to analyze whether the current scene complexity level of the target time segment is higher than the corresponding threshold when the proportion of frames triggering secondary annotation reminders in the target time segment exceeds the threshold ratio. If so, obtain the historical video processing record and the historical annotation deviation feedback data;

[0027] The data screening module is used to parse the historical video processing record and screen out several historical comparison segment records that are within the same level range as the current scene complexity level and have a matching time length;

[0028] The slope calculation module is used to parse each historical comparison segment record in turn, calculate the proportion of frames triggering secondary annotation reminders corresponding to each of them, draw a trend curve in chronological order, and calculate the average slope of the trend curve;

[0029] The deviation value calculation module is used to analyze the historical annotation deviation feedback data, obtain the deviation value between the initial annotation value and the correction result corresponding to each historical comparison segment record, and calculate the average value of all deviation values;

[0030] The correction factor generation module is used to generate a confidence correction factor by weighting based on the average slope and the average value of all deviation values, and is used to adjust the confidence parameters of all frames within the target time segment.

[0031] As a further limitation of the technical solution of the embodiment of the present invention, the slope calculation module specifically includes:

[0032] The proportion calculation unit is used to analyze each historical comparison segment record in turn, calculate the proportion between the number of frames triggering the secondary annotation reminder and the total number of frames in each historical comparison segment record, and obtain the proportion of the secondary annotation reminder frames for each;

[0033] The curve drawing unit is used to draw a corresponding trend curve with the time axis as the horizontal axis and the proportion of the secondary annotation reminder frames as the vertical axis;

[0034] The slope calculation unit is used to calculate the average value of the slopes of the trend curves.

[0035] As a further limitation of the technical solution of the embodiment of the present invention, the deviation value calculation module specifically includes:

[0036] The local feedback data acquisition unit is used to analyze the historical annotation deviation feedback data and extract the local annotation deviation feedback data corresponding to each historical comparison segment record therefrom;

[0037] The local feedback data analysis unit is used to analyze each local annotation deviation feedback data in turn and extract the initial annotation value and the correction result therefrom. Among them, the initial annotation value refers to all the frames triggering the secondary annotation reminder in each historical comparison segment record, and the correction result refers to the number of frames that finally need to be reminded of the secondary annotation after manual confirmation or feedback;

[0038] The deviation value average calculation unit is used to calculate the deviation value between each group of initial annotation values and the corresponding correction results, and statistically analyze the deviation values corresponding to all historical comparison segment records to obtain the average value of the deviation values.

[0039] As a further limitation of the technical solution of the embodiment of the present invention, the correction factor generation module specifically includes:

[0040] The confidence correction factor generation unit is used to call the confidence correction factor generation formula, and use the average value of the slopes of the trend curves and the average value of all deviation values as input parameters to perform weighted calculation to generate a confidence correction factor;

[0041] A confidence parameter correction unit, configured to adjust the initial confidence parameters of all frames within a target time segment based on a confidence correction factor to obtain corrected confidence parameters;

[0042] A reminder retention determination unit, configured to determine whether the corrected confidence parameter of each frame is lower than a preset confidence threshold. If so, retain the secondary annotation reminder for that frame; if not, cancel the original secondary annotation reminder.

[0043] As a further limitation of the technical solution of the embodiment of the present invention, the confidence correction factor generation formula is: , where refers to the confidence correction factor, refers to the average slope of the trend curve, refers to the adjustment weight corresponding to the average slope, refers to the absolute value of the average of all deviation values, refers to the adjustment weight corresponding to the absolute value;

[0044] In the confidence correction factor generation formula: , where refers to the total number of local annotation deviation feedback data, where refers to the th correction result of the local annotation deviation feedback data, refers to the th initial annotation value of the local annotation deviation feedback data, refers to the th deviation value between the initial annotation value and the corresponding correction result corresponding to the local annotation deviation feedback data.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] By introducing dual statistical parameters of the average slope of the trend curve and the average historical annotation deviation, the present invention constructs a confidence correction factor, realizes a dynamic adjustment mechanism for confidence parameters, and can effectively identify and correct the trend misjudgment offset problem accumulated by the system under similar scene complexities. Compared with the prior art that only judges based on a static confidence threshold, the present invention models by fusing historical processing records and manual feedback data, improves the system's perception ability of the reasons for confidence fluctuations and the judgment accuracy, significantly reduces redundant reminders, reduces the manual review burden, and enhances the stability and adaptability of the system under complex video structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the method provided by the embodiment of the present invention;

[0048] Figure 2Flowchart for drawing a trend curve and calculating its average slope in the method provided by the embodiments of the present invention;

[0049] Figure 3 Flowchart for calculating the deviation value between the initial annotation value and the corresponding correction result in the method provided by the embodiments of the present invention;

[0050] Figure 4 Flowchart for generating a confidence correction factor in the method provided by the embodiments of the present invention;

[0051] Figure 5 Application architecture diagram of the system provided by the embodiments of the present invention;

[0052] Figure 6 Structural block diagram of the slope calculation module in the system provided by the embodiments of the present invention;

[0053] Figure 7 Structural block diagram of the deviation value calculation module in the system provided by the embodiments of the present invention;

[0054] Figure 8 Structural block diagram of the correction factor generation module in the system provided by the embodiments of the present invention. Detailed implementation manners

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] Figure 1 The flowchart of the method provided by the embodiments of the present invention is shown.

[0057] Specifically, a data annotation method based on artificial intelligence, the method specifically includes the following steps:

[0058] Step S100, when the proportion of the secondary annotation reminder frames triggered in the target time segment exceeds the threshold ratio, analyze whether the current scene complexity level of the target time segment is higher than the corresponding threshold. If so, obtain the historical video processing records and the historical annotation deviation feedback data.

[0059] In the embodiments of the present invention, the target time segment refers to a local continuous frame section delimited by the system according to the triggering situation of the secondary annotation reminder when performing annotation analysis on the current video data, and is used to form a sub-range to be further identified and processed. This time segment usually has a fixed frame length or is dynamically delimited according to the annotation behavior, and its boundary frames are determined based on the frame number dense area where the secondary annotation reminder is triggered.

[0060] The secondary annotation reminder frame is a suspected low-confidence frame determined by the system by judging whether the confidence parameter corresponding to each frame is lower than the preset confidence threshold. During the preliminary annotation stage, the secondary annotation reminder frame is automatically marked as an annotation suspected frame that requires further manual confirmation. Such reminder frames usually originate from the system's preliminary perception of typical low-confidence scenarios such as blurred boundaries of image targets, interrupted action continuity, occlusion and overlap, or uncertain semantic recognition, and serve as important reminder points that need to be manually reviewed again in the annotation process.

[0061] The confidence parameter refers to the numerical value of the credibility generated by the annotation system when processing each frame of video data, based on the results of object recognition, semantic judgment, or behavior analysis of the frame by the artificial intelligence model, and is used to measure the certainty degree of the system's judgment result for the current frame label. This parameter is usually expressed in numerical form. The higher the value, the more confident the system is in the annotation result of the frame; the lower the value, the higher the uncertainty of the system.

[0062] The preset confidence threshold is a judgment boundary value set within the system to determine whether the confidence parameter of the current frame is reliable. When the confidence parameter of a certain frame is lower than this threshold, the system believes that the annotation result of this frame does not have sufficient credibility and needs to trigger a secondary annotation reminder and include it in the scope of manual review; when the confidence parameter is higher than this threshold, the system believes that the annotation result has reasonable accuracy and can directly retain the original label without further processing. This threshold can be adjusted according to the task type, model performance, or scene requirements.

[0063] The threshold ratio is used to measure whether the number of secondary annotation reminder frames in the target time segment has reached the abnormally dense level determined by the system, and is used to judge whether to enter the trend correction process. This threshold ratio can be set according to the system's annotation experience data. For example, in historical tasks, when the proportion of reminder frames in a certain time segment exceeds 15%, it often corresponds to a dense area of model misjudgment. Therefore, the threshold ratio can be set to 15% or dynamically adjusted according to specific application scenarios.

[0064] The scene complexity level refers to the grading result of the comprehensive evaluation of the scene structure complexity by the system based on multiple dimensions such as the spatial structure of the image, target density, difference degree between front and back frames, and occlusion situation within the target time segment. This level is usually represented by a multi-level score, such as from level one (simple) to level ten (complex), and can be automatically generated by the image feature extraction module.

[0065] The reason for setting "whether the current scene complexity level is higher than the corresponding threshold" as the judgment condition for triggering further analysis in the present invention is that when the scene structure of a video segment is complex, the misjudgment probability of the annotation system for object boundaries, action intentions, or semantic judgments increases significantly. Even if there is a brief concentration of annotation reminders in a simple scene, it may not necessarily represent a system deviation. Therefore, it is necessary to use this complexity level as the threshold for trend correction analysis. The corresponding threshold can be set based on historical annotation performance. For example, when the scene complexity level is at level three or above, the confidence fluctuation of the model is significant. Therefore, level three is used as the starting judgment threshold.

[0066] The historical video processing records are derived from the data processing logs recorded during the system's previous execution of annotation tasks, including raw annotation-related data such as feature recognition of video segments, confidence output, label generation and modification, and trigger reminder records. The historical annotation deviation feedback data is derived from the proofreading operation records in the manual annotation stage, including the difference information between the initial annotation and the final correction result in each paragraph, which is an important data basis for the system to construct the trend of deviation recognition and correction.

[0067] Among them, the historical video processing records at least include: video segment number, frame index range, initial annotation label generated by the system, confidence value corresponding to each frame, whether to trigger a secondary annotation reminder flag, summary information of scene structure features, and corresponding video metadata, etc.

[0068] The historical annotation deviation feedback data at least includes: the set of system's original annotation frames, the set of manually finally corrected and confirmed frames, the label difference result between the two, the manual feedback timestamp, the correction confirmation status flag, etc. in each annotation section.

[0069] Furthermore, the artificial intelligence-based data annotation method further includes the following steps:

[0070] Step S200, parse the historical video processing records, and screen out several historical comparison segment records that are within the same level range as the current scene complexity level and have a matching time length.

[0071] In the embodiment of the present invention, screening out several historical comparison segment records that are within the same level range as the current scene complexity level and have a matching time length means that based on the structural attributes of the current target time segment in the obtained historical video processing records, historical segments with similar background complexity and frame number characteristics are screened out as comparison objects to ensure the comparability and reference value of subsequent trend analysis and deviation judgment.

[0072] Specifically, the system first obtains the scene complexity level of the current target time segment and sets a level tolerance range, such as a ±1-level range centered on the current level, as the allowable matching level range. The system then traverses the recorded scene complexity levels of each video segment in the historical video processing records and filters out the segments whose complexity levels fall within this tolerance range.

[0073] On this basis, the system further makes a matching judgment on the time lengths of these candidate segments. The condition for time length matching can be set as the frame number difference not exceeding a preset percentage threshold of the frame number of the current segment (such as ±10%), to ensure that the selected historical control segment records have a similar behavior rhythm and annotation span as the current segment in terms of duration dimension. Finally, the historical segments that meet the above two conditions are determined as the set of historical control segment records required for the current analysis.

[0074] The reason for selecting historical segments with a scene complexity level similar to the current one is that in the annotation system, scene complexity is often closely related to the fluctuations of annotation confidence and the accuracy of model judgment. Video segments with different complexity levels have significantly different target overlap degrees, background dynamic factors, and visual structure characteristics. If the comparative analysis is not within the same level range, it is easy to lead to inaccurate deviation judgments.

[0075] At the same time, limiting the time length matching can ensure that the compared historical segments and the current segment have a structural alignment basis in dimensions such as frame density, action persistence, and label trigger rhythm, so that the subsequent calculation of the mean value of the trend curve slope and the mean value of the deviation value is more objective and stable, avoiding statistical distortion caused by segment dimension differences. Through these two screening criteria, the representativeness and referenceability of the historical control segment records are ensured, providing accurate data support for the subsequent generation of confidence correction factors in the present invention.

[0076] Furthermore, the artificial intelligence-based data annotation method further includes the following steps:

[0077] Step S300, parse each historical control segment record in sequence, calculate the proportion of secondary annotation reminder frames corresponding to each of them, draw a trend curve in chronological order, and calculate the mean value of the slope of the trend curve.

[0078] Specifically, Figure 2 The flowchart of drawing the trend curve and calculating its mean slope value is shown.

[0079] Among them, parsing each historical control segment record in sequence, calculating the proportion of secondary annotation reminder frames corresponding to each of them, drawing a trend curve in chronological order, and calculating the mean value of the slope of the trend curve specifically include the following steps:

[0080] Step S301: Analyze each historical comparison segment record in sequence, calculate the ratio between the number of frames triggering the secondary annotation reminder and the total number of frames in each historical comparison segment record, and obtain the secondary annotation reminder frame proportion for each;

[0081] Step S302: Use the time axis as the horizontal axis and the secondary annotation reminder frame proportion as the vertical axis to draw the corresponding trend curve;

[0082] Step S303: Calculate the average slope of the said trend curve.

[0083] In the embodiment of the present invention, the "analyze each historical comparison segment record, calculate the ratio between the number of frames triggering the secondary annotation reminder and the total number of frames in each historical comparison segment record, and obtain the secondary annotation reminder frame proportion for each" means that for the historical comparison segment records obtained by the aforesaid screening, the system sequentially reads the total number of frames included in each segment record, and the set of frame indices marked by the system as triggering the secondary annotation reminder in that segment. The system calculates the ratio of the number of frames triggering the secondary annotation reminder in each segment to the total number of frames in that segment to obtain a normalized proportion value, which is used to reflect the density of frames identified as low-confidence frames in the historical annotation process for that segment.

[0084] Specifically, the system reads the frame index range of each segment in the historical video processing record and the corresponding secondary annotation reminder mark table, counts the frames with reminder marks, and calculates the total number of frames in combination with the start and end frame numbers of that segment. Dividing the number of reminder frames by the total number of frames can obtain the secondary annotation reminder frame proportion for that segment. This proportion value is used to express the relative density of system uncertainty marks generated during the annotation process for this historical segment, and can reflect the confidence fluctuation of this segment in model recognition.

[0085] The drawing of the said trend curve means that after the system obtains the secondary annotation reminder frame proportions of multiple historical comparison segment records, according to the order of appearance of these historical segments on the time line, it sets the time axis as the horizontal axis and the corresponding proportion values as the vertical axis, and connects these points in a two-dimensional coordinate system to form a continuous trend line. The time order can be the time stamp order of the generation of historical segment records, or the record order in the system historical processing log. The purpose is to establish the change trend of the proportion index through the time dimension.

[0086] Calculating the average slope of this trend curve is to identify whether the annotation system under a historical similar background shows a continuous misjudgment trend. When the average slope is greater than zero, it indicates that under a similar complexity background, the secondary annotation reminder frame proportion of the system has an upward trend, which may mean that the system shows increasing uncertainty in this type of segment; when the average slope is less than zero, it indicates that the reminder proportion gradually decreases, indicating enhanced system processing stability; if the average slope is close to zero, it means that the misjudgment volatility of this type of video segment in history is small and the system judgment is relatively stable.

[0087] Therefore, the formation of the trend misjudgment offset direction is essentially due to the fact that the annotation system fails to dynamically adjust its confidence output strategy when processing video segments with similar levels of structural complexity, resulting in the accumulation of systematic biases during historical processing. Specifically, in this type of complexity background, the error between the original judgment result of the model and the manually corrected result is not timely fed back to the confidence adjustment mechanism, thus triggering a trend offset in the reminder density that gradually becomes higher or lower over time, seriously affecting the confidence judgment accuracy of the system for the current video segment and the reminder control ability. The present invention constructs a confidence correction mechanism with the ability to identify time trends by introducing the mean parameter of the trend curve slope and combining with the statistical analysis of historical correction biases, which can achieve adaptive correction for the above problems and improve the stability and robustness of the annotation system in complex structural scenarios.

[0088] Furthermore, the data annotation method based on artificial intelligence further includes the following steps:

[0089] Step S400, parse the historical annotation deviation feedback data, obtain the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculate the mean of all deviation values.

[0090] Specifically, Figure 3 The flowchart of calculating the deviation value between the initial annotation value and the corresponding correction result is shown.

[0091] Among them, parsing the historical annotation deviation feedback data, obtaining the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculating the mean of all deviation values specifically include the following steps:

[0092] Step S401, parse the historical annotation deviation feedback data, and extract the local annotation deviation feedback data corresponding to each historical comparison segment record from it;

[0093] Step S402, parse each local annotation deviation feedback data in turn, and extract the initial annotation value and the correction result from it. Among them, the initial annotation value refers to all the frame numbers that trigger the secondary annotation reminder in each historical comparison segment record, and the correction result refers to the frame numbers that finally need to be reminded for secondary annotation after manual confirmation or feedback;

[0094] Step S403, calculate the deviation value between each group of initial annotation values and the corresponding correction results, and statistically analyze the deviation values corresponding to all historical comparison segment records to obtain the mean of the deviation values.

[0095] In an embodiment of the present invention, the local annotation deviation feedback data corresponding to the historical comparison segment record is extracted by association through the index relationship established when the historical data is archived. While the system performs the annotation task and records the historical video processing record, it will store the corresponding manual feedback data as supplementary information. Each historical comparison segment record contains unique identification information of the time segment, such as the paragraph number, frame start and end range or timestamp. Based on this type of identification information, the system can locate the corresponding local feedback data from the historical annotation deviation feedback data to ensure the accuracy of the one-to-one correspondence.

[0096] When analyzing each set of local annotation deviation feedback data, the system extracts two key indicators: one is the number of frames that the system believes need to trigger a secondary annotation reminder during the initial judgment stage, as the initial annotation value; the other is the number of frames that are finally confirmed to need a secondary annotation reminder after manual review, as the correction result. After the system completes the traversal of all historical comparison segment records, it calculates the difference between the initial annotation value and the correction result in each set of data one by one, and quantifies these difference results into deviation values.

[0097] Next, the system summarizes and counts the deviation values ​​corresponding to all historical comparison segment records, and finally calculates the average value of these deviation values. This average value represents the average error degree between the system's automatic judgment and manual correction under similar scene complexity, and is an important reference for measuring whether the system's confidence output is stable.

[0098] By calculating the mean of the deviation value, we can quantify the system's misjudgment range in similar backgrounds and reflect the deviation trend of the annotation system's judgment on uncertain areas. If the mean is large, it means that the system has obvious confidence judgment deviation in such scenes and needs to be compensated by a correction mechanism; if the mean is small, it means that the system is consistent with the manual results and the current strategy can be maintained without adjustment. Therefore, as a component of the subsequent generation of confidence correction factors, the mean can improve the system's responsiveness to historical misjudgment trends and achieve more accurate annotation strategy optimization.

[0099] Furthermore, the artificial intelligence-based data labeling method further includes the following steps:

[0100] Step S500, based on the slope mean and the mean of all deviation values, a confidence correction factor is generated by weighting, and is used to adjust the confidence parameters of all frames in the target time segment.

[0101] Specifically, Figure 4 A flow chart for generating a confidence correction factor is shown.

[0102] Among them, based on the average slope and the average of all deviation values, a confidence correction factor is generated by weighting and is used to adjust the confidence parameters of all frames within the target time segment. Specifically, it includes the following steps:

[0103] Step S501: Call the confidence correction factor generation formula, and use the average slope of the trend curve and the average of all deviation values as input parameters to perform weighted calculation to generate the confidence correction factor;

[0104] Step S502: Based on the confidence correction factor, adjust the initial confidence parameters of all frames within the target time segment to obtain the corrected confidence parameters;

[0105] Step S503: Determine whether the corrected confidence parameter of each frame is lower than the preset confidence threshold. If so, retain the secondary annotation reminder for this frame; if not, cancel the original secondary annotation reminder.

[0106] The confidence correction factor generation formula is: , where refers to the confidence correction factor, refers to the average slope of the trend curve, refers to the adjustment weight corresponding to this average slope, refers to the absolute value of the average of all deviation values, refers to the adjustment weight corresponding to this absolute value;

[0107] In the confidence correction factor generation formula: , where refers to the total number of local annotation deviation feedback data, where refers to the th correction result of the local annotation deviation feedback data, refers to the th initial annotation value of the local annotation deviation feedback data, refers to the th deviation value between the initial annotation value and the corresponding correction result of the local annotation deviation feedback data.

[0108] In the embodiments of the present invention, the mean slope of the trend curve and the mean of all deviation values are jointly used as parameter inputs to perform weighted calculation to generate a confidence correction factor. This is because these two types of parameters respectively reflect the stability and deviation intensity of the historical control segment records in the time dimension and the error dimension. The mean slope of the trend curve is used to depict the confidence fluctuation evolution trend of the system in processing scenarios with similar complexity. If this trend continues to rise or fall, it often indicates that the annotation reliability of the system for this type of video segment is in dynamic change; while the mean of all deviation values reflects the overall deviation degree between the system's judgment result and the final manual correction result, and is an important indicator for measuring the stability and consistency of the system. Fusing these two parameters through weighting helps the system accurately identify potential false alarm or missed alarm risks when processing the current target time segment, and achieve directional adjustment and intensity correction of the initial confidence judgment.

[0109] Through the above mechanism, the system can adaptively perform dynamic fine-tuning on the confidence parameters of the current frame segment according to the changes in historical trends and feedback data while keeping the overall recognition strategy unchanged, thereby avoiding the repetitive misjudgment phenomenon caused by the fixed threshold mechanism, effectively enhancing the system's adaptability to diverse video scenarios, and improving the accuracy of recognition results and the flexibility of reminder control.

[0110] During the implementation process, the system can adjust the confidence parameters in various ways. For example, by using the weighted offset method, the confidence correction factor is used as an adjustment increment and directly acts on the initial confidence value; or the confidence correction factor is used as an adjustment ratio to compress or amplify the initial confidence value; it can also map the correction factor to the confidence interval by looking up the correction reference table established based on the historical model calibration results, so as to obtain the corrected confidence parameters. Any of the above methods can achieve confidence optimization based on historical trends.

[0111] Subsequently, the system judges the corrected confidence parameters of each frame in the target time segment one by one. If the confidence value corresponding to a certain frame is still lower than the preset confidence threshold, the system retains its original secondary annotation reminder label; if the confidence value is higher than the threshold, the original reminder mark is automatically cancelled. Through this judgment mechanism, the system can make a more reasonable judgment on the effectiveness of the current reminder label on the basis of absorbing historical deviation experience, thereby effectively reducing invalid or redundant manual review operations, reducing the burden of manual intervention, and improving the accuracy and efficiency of the overall annotation process.

[0112] Furthermore, Figure 5 shows the application architecture diagram of the system provided by the embodiments of the present invention.

[0113] Among them, in another preferred embodiment provided by the present invention, an artificial intelligence-based data annotation system includes:

[0114] The data acquisition module 100 is configured to analyze whether the current scene complexity level of the target time segment is higher than the corresponding threshold when the proportion of the secondary annotation reminder frames triggered in the target time segment exceeds the threshold ratio. If so, the historical video processing records and the historical annotation deviation feedback data are acquired.

[0115] Furthermore, the artificial intelligence-based data annotation system further includes:

[0116] The data screening module 200 is configured to parse the historical video processing records and screen out several historical comparison segment records that are within the same level range as the current scene complexity level and have a matching time length.

[0117] Furthermore, the artificial intelligence-based data annotation system further includes:

[0118] The slope calculation module 300 is configured to sequentially parse each historical comparison segment record, calculate the proportion of the secondary annotation reminder frames corresponding to each of them, draw a trend curve in chronological order, and calculate the mean value of the slope of the trend curve.

[0119] Specifically, Figure 6 FIG. shows the structural block diagram of the slope calculation module 300 in the system provided by the embodiment of the present invention.

[0120] Among them, in the preferred embodiment provided by the present invention, the slope calculation module 300 specifically includes:

[0121] The proportion calculation unit 301 is configured to sequentially parse each historical comparison segment record, calculate the ratio between the number of frames triggering the secondary annotation reminder in each historical comparison segment record and the corresponding total number of frames, and obtain the proportion of the secondary annotation reminder frames corresponding to each of them;

[0122] The curve drawing unit 302 is configured to draw a corresponding trend curve with the time axis as the horizontal axis and the proportion of the secondary annotation reminder frames as the vertical axis;

[0123] The slope calculation unit 303 is configured to calculate the mean value of the slope of the trend curve.

[0124] Furthermore, the artificial intelligence-based data annotation system further includes:

[0125] The deviation value calculation module 400 is configured to parse the historical annotation deviation feedback data, obtain the deviation value between the initial annotation value and the correction result corresponding to each historical comparison segment record, and calculate the mean value of all deviation values.

[0126] Specifically, Figure 7 FIG. shows the structural block diagram of the deviation value calculation module 400 in the system provided by the embodiment of the present invention.

[0127] Among them, in the preferred embodiment provided by the present invention, the deviation value calculation module 400 specifically includes:

[0128] A local feedback data acquisition unit 401, configured to parse historical annotation deviation feedback data, and extract local annotation deviation feedback data corresponding to each historical comparison segment record therefrom;

[0129] A local feedback data parsing unit 402, configured to sequentially parse each local annotation deviation feedback data, and extract an initial annotation value and a correction result therefrom. Among them, the initial annotation value refers to all frames triggering a secondary annotation reminder in each historical comparison segment record, and the correction result refers to the number of frames that finally need to be reminded of secondary annotation after manual confirmation or feedback;

[0130] A deviation value average calculation unit 403, configured to calculate the deviation value between each group of initial annotation values and the corresponding correction results, and statistically analyze the deviation values corresponding to all historical comparison segment records to obtain the average value of the deviation values.

[0131] Further, the artificial intelligence-based data annotation system further includes:

[0132] A correction factor generation module 500, configured to generate a confidence correction factor by weighted calculation based on the average value of the slopes and the average value of all deviation values, and used to adjust the confidence parameters of all frames within the target time segment.

[0133] Specifically, Figure 8 shows the structural block diagram of the correction factor generation module 500 in the system provided by the embodiment of the present invention.

[0134] Among them, in the preferred embodiment provided by the present invention, the correction factor generation module 500 specifically includes:

[0135] A confidence correction factor generation unit 501, configured to call the confidence correction factor generation formula, and use the average value of the slopes of the trend curve and the average value of all deviation values as input parameters to perform weighted calculation to generate a confidence correction factor;

[0136] A confidence parameter correction unit 502, configured to adjust the initial confidence parameters of all frames within the target time segment based on the confidence correction factor to obtain the corrected confidence parameters;

[0137] A reminder retention determination unit 503, configured to determine whether the corrected confidence parameter of each frame is lower than the preset confidence threshold. If so, retain the secondary annotation reminder for this frame. If not, cancel the original secondary annotation reminder.

[0138] The confidence correction factor generation formula is: , where refers to the confidence correction factor, refers to the average slope of the trend curve, refers to the adjustment weight corresponding to the average slope, refers to the absolute value of the average of all deviation values, refers to the adjustment weight corresponding to the absolute value;

[0139] In the confidence correction factor generation formula: , where refers to the total number of local annotation deviation feedback data, where refers to the correction result of the th local annotation deviation feedback data, refers to the initial annotation value of the th local annotation deviation feedback data, refers to the deviation value between the initial annotation value and the corresponding correction result corresponding to the

[0140] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0141] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0142] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0143] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be subject to the appended claims.

[0144] The above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A data annotation method based on artificial intelligence, characterized in that, The method includes: When the proportion of frames triggering secondary annotation reminders in the target time segment exceeds the threshold ratio, analyze whether the current scene complexity level of the target time segment is higher than the corresponding threshold. If so, obtain the historical video processing records and historical annotation deviation feedback data; Parse the historical video processing records, and screen out several historical comparison segment records that are within the same level range as the current scene complexity level and have a matching time length; Parse each historical comparison segment record in sequence, calculate the proportion of frames triggering secondary annotation reminders corresponding to each, draw a trend curve in chronological order, and calculate the average slope of the trend curve; Parse the historical annotation deviation feedback data, obtain the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculate the average value of all deviation values; Based on the average slope and the average value of all deviation values, generate a confidence correction factor through weighting, and use it to adjust the confidence parameters of all frames within the target time segment.

2. The data annotation method based on artificial intelligence according to claim 1, wherein The steps of parsing each historical comparison segment record in sequence, calculating the proportion of frames triggering secondary annotation reminders corresponding to each, drawing a trend curve in chronological order, and calculating the average slope of the trend curve include: Parse each historical comparison segment record in sequence, calculate the ratio between the number of frames triggering secondary annotation reminders in each historical comparison segment record and the corresponding total number of frames, and obtain the proportion of frames triggering secondary annotation reminders corresponding to each; Taking the time axis as the horizontal axis and the proportion of frames triggering secondary annotation reminders as the vertical axis, draw the corresponding trend curve; Calculate the average slope of the trend curve.

3. The data annotation method based on artificial intelligence according to claim 2, wherein The steps of parsing the historical annotation deviation feedback data, obtaining the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculating the average value of all deviation values include: Parse the historical annotation deviation feedback data, and extract the local annotation deviation feedback data corresponding to each historical comparison segment record from it; Parse each local annotation deviation feedback data in sequence, and extract the initial annotation value and the correction result from it. Among them, the initial annotation value refers to all the frames triggering secondary annotation reminders in each historical comparison segment record, and the correction result refers to the number of frames that finally need to trigger secondary annotation reminders after manual confirmation or feedback; Calculate the deviation value between each group of initial annotation values and the corresponding correction results, and count the deviation values corresponding to all historical comparison segment records to obtain the average value of the deviation values.

4. The data annotation method based on artificial intelligence according to claim 3, wherein The steps of generating a confidence correction factor through weighting based on the average slope and the average value of all deviation values, and using it to adjust the confidence parameters of all frames within the target time segment include: Call the confidence correction factor generation formula, and use the average slope of the trend curve and the average value of all deviation values as input parameters, and perform weighted calculation to generate a confidence correction factor; Based on the confidence correction factor, adjust the initial confidence parameters of all frames within the target time segment to obtain the corrected confidence parameters; Judge whether the corrected confidence parameter of each frame is lower than the preset confidence threshold. If so, retain the secondary annotation reminder for this frame. If not, cancel the original secondary annotation reminder.

5. The data annotation method based on artificial intelligence according to claim 4, wherein The confidence correction factor generation formula is as follows: , where refers to the confidence correction factor, refers to the average slope of the trend curve, refers to the adjustment weight corresponding to the average slope, refers to the absolute value of the average of all deviation values, refers to the adjustment weight corresponding to the absolute value; In the confidence correction factor generation formula: , where refers to the total number of local annotation deviation feedback data, where refers to the th correction result of the local annotation deviation feedback data, refers to the th initial annotation value of the local annotation deviation feedback data, refers to the th deviation value between the initial annotation value and the corresponding correction result corresponding to the local annotation deviation feedback data.

6. An artificial intelligence-based data annotation system, characterized in that, The system includes: a data acquisition module, a data screening module, a slope calculation module, a deviation value calculation module, and a correction factor generation module, where: A data acquisition module, which is used to analyze whether the current scene complexity level of the target time segment is higher than the corresponding threshold when the proportion of secondary annotation reminder frames triggered in the target time segment exceeds the threshold ratio. If so, it acquires the historical video processing records and the historical annotation deviation feedback data; A data screening module, which is used to parse the historical video processing records and screen out several historical comparison segment records that are within the same level range as the current scene complexity level and have a matching time length; A slope calculation module, which is used to parse each historical comparison segment record in sequence, calculate the proportion of secondary annotation reminder frames corresponding to each of them, draw a trend curve in chronological order, and calculate the average value of the slopes of the trend curves; A deviation value calculation module, which is used to parse the historical annotation deviation feedback data, obtain the deviation values between the initial annotation values and the correction results corresponding to each historical comparison segment record, and calculate the average value of all deviation values; A correction factor generation module, which is used to generate a confidence correction factor by weighting based on the average value of the slopes and the average value of all deviation values, and is used to adjust the confidence parameters of all frames within the target time segment.

7. The data annotation system based on artificial intelligence according to claim 6, wherein The slope calculation module specifically includes: A proportion calculation unit, which is used to parse each historical comparison segment record in sequence, calculate the ratio between the number of frames triggering secondary annotation reminders in each historical comparison segment record and the corresponding total number of frames, and obtain the proportion of secondary annotation reminder frames corresponding to each of them; A curve drawing unit, which is used to draw a corresponding trend curve with the time axis as the horizontal axis and the proportion of secondary annotation reminder frames as the vertical axis; A slope calculation unit, which is used to calculate the average value of the slopes of the trend curve.

8. The data annotation system based on artificial intelligence according to claim 7, wherein The deviation value calculation module specifically includes: A local feedback data acquisition unit, which is used to parse the historical annotation deviation feedback data and extract the local annotation deviation feedback data corresponding to each historical comparison segment record from it; A local feedback data parsing unit, which is used to parse each local annotation deviation feedback data in sequence and extract the initial annotation value and the correction result from it. Among them, the initial annotation value refers to all the frames triggering secondary annotation reminders in each historical comparison segment record, and the correction result refers to the number of frames that finally need to be reminded of secondary annotation after manual confirmation or feedback; A deviation value average calculation unit, which is used to calculate the deviation values between each group of initial annotation values and the corresponding correction results, and statistically analyze the deviation values corresponding to all historical comparison segment records to obtain the average value of the deviation values.

9. The data annotation system based on artificial intelligence according to claim 8, wherein, The correction factor generation module specifically includes: A confidence correction factor generation unit, which is used to call the confidence correction factor generation formula and use the average value of the slopes of the trend curve and the average value of all deviation values as input parameters to perform weighted calculation to generate a confidence correction factor; A confidence parameter correction unit, which is used to adjust the initial confidence parameters of all frames within the target time segment based on the confidence correction factor to obtain the corrected confidence parameters; A reminder retention determination unit, which is used to determine whether the corrected confidence parameter of each frame is lower than the preset confidence threshold. If so, it retains the secondary annotation reminder for that frame. If not, it cancels the original secondary annotation reminder.

10. The data annotation system based on artificial intelligence according to claim 9, wherein The confidence correction factor generation formula is as follows: , where refers to the confidence correction factor, refers to the average slope of the trend curve, refers to the adjustment weight corresponding to the average slope, refers to the absolute value of the average of all deviation values, refers to the adjustment weight corresponding to the absolute value; In the confidence correction factor generation formula: , where refers to the total number of local annotation deviation feedback data, where refers to the correction result of the th local annotation deviation feedback data, refers to the initial annotation value of the th local annotation deviation feedback data, refers to the deviation value between the initial annotation value corresponding to the th local annotation deviation feedback data and the corresponding correction result.

Images