A contrast method, device, equipment and medium across batch image increments

Abstract

Embodiments of the present application provide a cross-batch image incremental comparison method and device, equipment and medium, relating to the field of image processing, which obtains a plurality of test scene groups and corresponding scene fingerprints by first identifying and processing the batch images to be processed, filters candidate historical scene fingerprints according to the condition that the scene keyword similarity is greater than a first preset threshold and the number of images in the group is the same, then calculates the comprehensive similarity and takes the highest comprehensive similarity to determine whether the test scene group is marked as a matching success, for each test scene group in the batch images to be processed, obtains a structured index based on a visual language model combined with a hierarchical quantization mapping rule, compares the structured index of the test scene group marked as a matching success with the structured index of the matching historical scene group to obtain an incremental comparison result, and relies on the above process to automatically complete the pairing of different batch test scene groups, reducing the consumption of manpower and time, and solving the technical problem of low image comparison efficiency.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to a method, apparatus, device and medium for comparing incremental images across batches. Background Technology

[0002] Imaging equipment firmware is continuously iterated and upgraded. After each version update, regression testing of imaging effect is required. By comparing the images captured in the current batch with historical test samples, the changes in image quality brought about by firmware adjustments can be evaluated. At the same time, the overall evolution of image quality during multiple version iterations can be tracked over a long period of time to accurately identify imaging defect fluctuations. Therefore, comparative analysis of images from different test batches has important engineering application value.

[0003] In existing technologies, cross-batch image comparison is carried out manually. After acquiring images for a single batch, testers manually retrieve the current batch of images and compare them item by item with the image analysis records of historical batches to complete the comparative evaluation of the imaging effect before and after firmware iteration.

[0004] However, existing technologies suffer from low image comparison efficiency. Different batches of test images typically cover various test scenario types. Images of the same test scenario in different batches need to be paired before meaningful comparisons can be made. However, existing manual annotation and pairing methods consume a lot of manpower and time, resulting in low image comparison efficiency. Summary of the Invention

[0005] This application provides a method, apparatus, device, and medium for comparing incremental images across batches, in order to solve the technical problem of low image comparison efficiency in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for comparing image increments across batches, including:

[0007] The acquired batch images to be processed are identified and processed to obtain multiple test scene groups in the batch images to be processed and scene fingerprints corresponding to each test scene group;

[0008] For each test scenario group, a matching candidate historical scenario fingerprint is obtained from the historical scenario fingerprints in the historical batch; wherein, the similarity between the historical scenario keywords in the candidate historical scenario fingerprint and the scenario keywords in the scenario fingerprint is greater than a first preset threshold, and the number of images within the group in the historical scenario fingerprint is the same as the number of images within the group in the scenario fingerprint.

[0009] For each test scenario group, based on the scenario fingerprint and the matched candidate historical scenario fingerprints, a comprehensive similarity is obtained, and the highest comprehensive similarity is obtained from the comprehensive similarity. Based on the highest comprehensive similarity, it is determined whether the test scenario group is successfully matched.

[0010] For each test scene group in the batch of images to be processed, according to the preset level quantization mapping rules and based on the visual language model, the structured indicators corresponding to the test scene group are obtained; wherein, the structured indicators include: the scene identifier, batch code, and the set of indicators for the image quality dimension in the test scene group.

[0011] For test scene groups marked as successfully matched, the structured indicators corresponding to the test scene groups are compared with the structured indicators of historical scene groups that were successfully matched by the test scene groups to obtain the incremental comparison results of the batch images to be processed.

[0012] In one possible implementation, the image quality dimension includes: exposure, white balance, noise, texture, focus and / or a combination thereof, and the set of indicators for the image quality dimension includes: defect level, level quantification score, defect description summary and judgment confidence.

[0013] The step of comparing the structured metrics corresponding to the test scenario group with the structured metrics of historical scenario groups that were successfully matched with the test scenario group to obtain the incremental comparison results of the batch images to be processed includes:

[0014] Based on the image quality dimension index set in the structured indicators corresponding to the test scenario group, the image quality dimension index set in the structured indicators of the historical scenario group that successfully matched the test scenario group is compared dimension by dimension to obtain the change in each image quality dimension; wherein, the structured indicators of the historical scenario group are obtained by reading data from a preset storage terminal.

[0015] In one possible implementation, it also includes:

[0016] Obtain incremental comparison results of multiple batches of images, and structured metrics for each batch of images;

[0017] A time series is constructed based on the structured indicators under each batch of images; wherein, the time series includes: scene mode, image quality dimension and data point list, and the data point list includes: batch code, analysis timestamp, defect level, level quantification score and number of samples;

[0018] Based on the time series, linear regression and significance tests are used to obtain the trend direction, linear fit degree and significance probability of each image quality dimension;

[0019] A comprehensive trend result is obtained based on the incremental comparison results of the multiple batches of images, the trend direction of each image quality dimension, the linear fitting degree, and the significance probability.

[0020] In one possible implementation, the step of obtaining the trend direction, linear fit degree, and significance probability of each image quality dimension based on the time series using linear regression and significance testing includes:

[0021] Based on the time span between the analysis timestamp corresponding to each historical data point in the time series and the analysis timestamp of the batch images to be processed, the time decay weight of each historical data point is determined; wherein, the time decay weight decreases monotonically as the time span increases.

[0022] Historical data points whose time decay weight is less than a preset expiration threshold are marked as expired data.

[0023] Using the time decay weight as a weighting coefficient, a weighted linear regression is performed on the rank quantification scores of each historical data point in the time series that is not marked as expired data to obtain the weighted slope and the linear fit degree.

[0024] Based on the weighted slope and the preset slope threshold, the trend direction of each image quality dimension is obtained;

[0025] When the number of historical data points in the time series that are not marked as expired data is greater than or equal to a preset number, a significance test is performed based on the weighted slope to obtain the significance probability.

[0026] In one possible implementation, the comprehensive trend result includes a comprehensive trend score and a trend insight report. Obtaining the comprehensive trend result based on the incremental comparison results of the multiple batches of images, the trend direction of each image quality dimension, the linear fit degree, and the significance probability includes:

[0027] Based on the trend direction of each image quality dimension, determine the trend score corresponding to each image quality dimension;

[0028] The overall trend score is obtained by multiplying the trend score corresponding to each image quality dimension by its respective preset dimension weight and then summing the results.

[0029] Based on the comparison between the significance probability and the preset significance threshold, the significance level corresponding to each image quality dimension is determined.

[0030] The trend direction of each image quality dimension is classified to obtain a list of improvement dimensions, a list of regression dimensions, and a list of stable dimensions;

[0031] Based on the incremental comparison results of the multiple batches of images, as well as the trend direction, linear fitting degree and significance probability of each image quality dimension, regression alarm flags and regression severity levels are generated.

[0032] The trend score corresponding to each image quality dimension, the overall trend score, the significance level corresponding to each image quality dimension, the list of improvement dimensions, the list of regression dimensions, the list of stable dimensions, the regression alarm flag, and the regression severity level are all input into the large language model to obtain the trend insight report.

[0033] In one possible implementation, it also includes:

[0034] When the structured metrics corresponding to the test scenario group are obtained, the corresponding historical scenario group is determined based on the correspondence established by the successful matching of the test scenario group with the historical scenario group.

[0035] Before adding the structured indicators of the test scenario group as new data points to the time series corresponding to the historical scenario group, the new data points are predicted using a weighted linear regression equation fitted based on historical data points in the historical scenario group that are not marked as expired data.

[0036] If the deviation between the actual quantized score of the new data point and the predicted value exceeds a preset detection threshold, the new data point is marked as a mutation point; wherein, the preset detection threshold is the product of a preset multiple and the residual standard deviation of all historical data points in the time series corresponding to the historical scene group that have not been marked as expired data.

[0037] In one possible implementation, the step of performing recognition processing on the acquired batch of images to be processed, and obtaining multiple test scene groups in the batch of images to be processed and a scene fingerprint corresponding to each test scene group, includes:

[0038] The scene identifiers in the batch of images to be processed are identified to obtain multiple test scene groups; wherein the scene identifiers of the images in each test scene group are the same;

[0039] Extract scene information from each test scene group, and construct a scene fingerprint corresponding to the test scene group based on the scene information;

[0040] The scenario information includes: the file name pattern of the test scenario group, the serial number group identifier of the test scenario group, scenario keywords, the number of images in the group, batch code, and analysis timestamp.

[0041] In one possible implementation, obtaining a comprehensive similarity score for each scene fingerprint corresponding to each test scene group, based on the scene fingerprint and the matched candidate historical scene fingerprints, includes:

[0042] For each test scenario group, the similarity of scenario keywords, file name patterns, and sequence numbers between the scenario fingerprint and each matching candidate historical scenario fingerprint is obtained.

[0043] Based on the similarity of scene keywords, file name patterns, and sequence numbers between the scene fingerprint and each matching candidate historical scene fingerprint, and a preset weight, the comprehensive similarity between the scene fingerprint and each matching candidate historical scene fingerprint is obtained.

[0044] In one possible implementation, it also includes:

[0045] The confidence information of each scene fingerprint is determined based on the highest comprehensive similarity of each scene fingerprint and the preset confidence level.

[0046] The step of obtaining the structured indicators corresponding to the test scenario group according to the preset level quantization mapping rules and based on the visual language model includes:

[0047] In response to the high confidence level information, according to the preset level quantization mapping rules and based on the visual language model, the structured indicators corresponding to the test scenario groups marked as successful matching and with high confidence level information are obtained respectively.

[0048] In response to the confidence information being of medium confidence and the confirmation signal for the test scenario group with medium confidence information, the structured indicators corresponding to the test scenario groups marked as successfully matched and with medium confidence information are obtained according to the preset level quantization mapping rules and based on the visual language model.

[0049] In one possible implementation, determining whether the marker for the test scenario group is a successful match based on the highest overall similarity includes:

[0050] If the highest overall similarity is greater than or equal to the second preset threshold, then the test scenario group is marked as a successful match.

[0051] or,

[0052] If the highest overall similarity is less than the second preset threshold, then the test scenario group is marked as a failed match.

[0053] Secondly, embodiments of this application provide a comparison device for cross-batch image increments, comprising:

[0054] The recognition module is used to perform recognition processing on the acquired batch images to be processed, and to obtain multiple test scene groups in the batch images to be processed and the scene fingerprint corresponding to each test scene group;

[0055] The fingerprint acquisition module is used to acquire matching candidate historical scene fingerprints from historical scene fingerprints in historical batches for each scene fingerprint corresponding to the test scene group; wherein, the similarity between the historical scene keywords in the candidate historical scene fingerprint and the scene keywords in the scene fingerprint is greater than a first preset threshold, and the number of images within the group in the historical scene fingerprint is the same as the number of images within the group in the scene fingerprint.

[0056] The determination module is used to obtain a comprehensive similarity for each scene fingerprint corresponding to each test scene group, based on the scene fingerprint and the matched candidate historical scene fingerprints, and to obtain the highest comprehensive similarity from the comprehensive similarity, and to determine whether the tag of the test scene group is a successful match based on the highest comprehensive similarity;

[0057] The indicator acquisition module is used to acquire the structured indicators corresponding to each test scene group in the batch of images to be processed, according to a preset level quantization mapping rule and based on a visual language model; wherein, the structured indicators include: the scene identifier, batch code, and a set of indicators for the image quality dimension in the test scene group.

[0058] The comparison module is used to compare the structured indicators corresponding to the test scene group marked as a successful match with the structured indicators of the historical scene group that was successfully matched by the test scene group, and obtain the incremental comparison results of the batch images to be processed.

[0059] Thirdly, embodiments of this application provide an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0060] The memory stores computer-executed instructions;

[0061] When the processor executes computer execution instructions stored in the memory, it is used to implement the cross-batch image increment comparison method as described in any of the first aspects.

[0062] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the cross-batch image increment comparison method as described in any of the first aspects.

[0063] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, is used to implement the cross-batch image increment comparison method as described in any of the first aspects.

[0064] This application provides a method, apparatus, device, and medium for incremental comparison of images across batches. First, the batch images to be processed are identified to obtain multiple test scene groups and corresponding scene fingerprints. Candidate historical scene fingerprints are filtered based on conditions such as scene keyword similarity exceeding a first preset threshold and the number of images within the group being the same. Then, a comprehensive similarity is calculated, and the highest comprehensive similarity is used to determine whether the test scene group is marked as a successful match. For each test scene group in the batch images to be processed, a structured index is obtained based on a visual language model combined with a hierarchical quantization mapping rule. For test scene groups marked as successfully matched, the structured index is compared with the structured index of the matching historical scene group to obtain the incremental comparison result. This process automatically completes the pairing of test scene groups from different batches, eliminating manual labeling and pairing operations, reducing manpower and time consumption, and solving the technical problem of low image comparison efficiency in the prior art. Attached Figure Description

[0065] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0066] Figure 1 A schematic diagram illustrating an application scenario of the cross-batch image increment comparison method provided in this application embodiment;

[0067] Figure 2 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 1 ;

[0068] Figure 3 Example flowchart of incremental comparison and structured index reuse provided for embodiments of this application;

[0069] Figure 4 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 2 ;

[0070] Figure 5 Example diagram of trend detection and regression alarm process provided in the embodiments of this application;

[0071] Figure 6 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 3 ;

[0072] Figure 7 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 4 ;

[0073] Figure 8 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 5 ;

[0074] Figure 9 This is an example diagram illustrating the scene fingerprint extraction and scene alignment process provided in an embodiment of this application.

[0075] Figure 10 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 6 ;

[0076] Figure 11 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 7 ;

[0077] Figure 12 Example diagram of a three-layer architecture for cross-batch incremental comparison and trend analysis provided in this application embodiment;

[0078] Figure 13 An example diagram illustrating the extended cross-batch data model provided in this application embodiment;

[0079] Figure 14 A schematic diagram of the structure of the comparison device for cross-batch image increments provided in an embodiment of this application;

[0080] Figure 15 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0081] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0082] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0083] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply difference. It should be noted that in the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more.

[0084] It should be noted that the phrase "at...time" in the embodiments of this application can refer to the instant at which a certain situation occurs, or to a period of time after the occurrence of a certain situation. The embodiments of this application do not specifically limit this. In addition, the comparison method, apparatus, device and medium for cross-batch image increments provided in the embodiments of this application are only examples. A comparison method, apparatus, device and medium for cross-batch image increments may also include more or less content.

[0085] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0086] To clearly understand the technical solution of this application, the solutions of existing technologies will first be described in detail. Imaging device firmware undergoes continuous iteration and upgrades. After each version update, regression testing of the imaging effect is required. By comparing the images captured in the current batch with historical test samples, the changes in image quality caused by firmware adjustments are evaluated. Simultaneously, by tracking the overall evolution of image quality over multiple version iterations, fluctuations in imaging defects can be accurately identified. Therefore, comparative analysis of images from different test batches has significant engineering application value.

[0087] In existing technologies, cross-batch image comparison is conducted manually. After acquiring images for a single batch, testers manually retrieve and compare the images from the current batch with the image analysis records of historical batches to evaluate the imaging effects before and after firmware iteration. However, in existing technologies, test images from different batches typically cover various test scenario types. Images of the same test scenario in different batches need to be paired before meaningful comparisons can be made. Current manual labeling and pairing methods are labor-intensive and time-consuming, resulting in low image comparison efficiency. Therefore, existing technologies suffer from low image comparison efficiency.

[0088] Therefore, to address the low image comparison efficiency in existing technologies, this research found that, to solve this problem: Optionally, the batch images to be processed can be identified to obtain multiple test scene groups and scene fingerprints corresponding to each test scene group. Optionally, after obtaining the scene fingerprints of each test scene group, for each test scene group's scene fingerprint, candidate historical scene fingerprints with a historical scene keyword similarity greater than a first preset threshold and the same number of images within the group can be selected from historical scene fingerprints in historical batches. Then, a comprehensive similarity is calculated based on the scene fingerprint and the candidate historical scene fingerprints, and the highest comprehensive similarity is selected. The labeling of the test scene group as a successful match is determined based on the highest comprehensive similarity. Optionally, for each test scene group in the batch images to be processed, a structured index corresponding to the test scene group can be obtained according to a preset level quantization mapping rule and based on a visual language model. Then, for the test scene group labeled as a successful match, the structured index is compared with the structured index of the historical scene group that successfully matched the test scene group to obtain the incremental comparison result of the batch images to be processed.

[0089] This application provides a method, apparatus, device, and medium for incremental comparison of images across batches. First, the batch images to be processed are identified to obtain multiple test scene groups and corresponding scene fingerprints. Candidate historical scene fingerprints are filtered based on conditions such as scene keyword similarity exceeding a first preset threshold and the number of images within the group being the same. Then, a comprehensive similarity is calculated, and the highest comprehensive similarity is used to determine whether the test scene group is marked as a successful match. For each test scene group in the batch images to be processed, a structured index is obtained based on a visual language model combined with a hierarchical quantization mapping rule. For test scene groups marked as successfully matched, the structured index is compared with the structured index of the matching historical scene group to obtain the incremental comparison result. This process automatically completes the pairing of test scene groups from different batches, eliminating manual labeling and pairing operations, reducing manpower and time consumption, and solving the technical problem of low image comparison efficiency in the prior art.

[0090] Based on the above-mentioned inventive discovery, the technical solution of this application is proposed.

[0091] The following describes the application scenarios of the cross-batch image increment comparison method provided in the embodiments of the present invention. Figure 1 This is a schematic diagram illustrating an application scenario of the cross-batch image increment comparison method provided in this application embodiment. For example... Figure 1 As shown, this application scenario includes a mobile terminal 101 and a server 102. The mobile terminal 101 collects batch images to be processed and sends them to the server 102. The server 102 performs recognition processing on the acquired batch images to be processed, obtaining multiple test scene groups in the batch images to be processed and the scene fingerprint corresponding to each test scene group. For the scene fingerprint corresponding to each test scene group, the server 102 obtains matching candidate historical scene fingerprints from historical scene fingerprints in historical batches. For the scene fingerprint corresponding to each test scene group, the server 102 obtains a comprehensive similarity based on the scene fingerprint and the matching candidate historical scene fingerprints, and obtains the highest comprehensive similarity from the comprehensive similarity. Based on the highest comprehensive similarity, it determines whether the test scene group is marked as a successful match. For each test scene group in the batch images to be processed, the server 102 obtains the structured indicators corresponding to the test scene group according to a preset level quantization mapping rule and based on a visual language model. For the test scene groups marked as successfully matched, the server 102 compares the structured indicators corresponding to the test scene groups with the structured indicators of historical scene groups that have successfully matched the test scene groups, and obtains the incremental comparison results of the batch images to be processed.

[0092] Figure 2 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 1 .like Figure 2 As shown, the cross-batch image increment comparison method provided in this embodiment includes the following steps:

[0093] Step S201: Perform recognition processing on the acquired batch images to be processed, and obtain multiple test scene groups in the batch images to be processed and the scene fingerprint corresponding to each test scene group.

[0094] In this embodiment, the acquired batch of images to be processed can be parsed by filename parsing. The images are divided into different test scenario groups according to the different values ​​of the scene identifier field in the filename. All images in the same test scenario group have the same value for the scene identifier field. Then, for each test scenario group, scene keywords, filename patterns, and sequence group identifiers are extracted from the filenames of the images in that group. These are combined with the number of images in the group, the batch code, and the analysis timestamp to form the scene fingerprint corresponding to that test scenario group. This step is used to automatically group the batch of images to be processed according to test scenarios and construct a structured scene fingerprint for each scene group, providing a directly comparable data basis for matching candidate historical scene fingerprints from historical scene fingerprints.

[0095] Alternatively, scene matching based on image content can be used instead of scene fingerprinting based on filenames. By extracting global feature vectors of images (such as CLIP embedding) and calculating cosine similarity across batches of images for scene matching, this approach can handle cases where filenames do not contain scene information, but it increases the computational cost of feature extraction, and the matching accuracy for the same scene from different shooting angles is not as stable as filename matching.

[0096] Step S202: For the scene fingerprint corresponding to each test scene group, obtain matching candidate historical scene fingerprints from the historical scene fingerprints in the historical batch; wherein, the similarity between the historical scene keywords in the candidate historical scene fingerprint and the scene keywords in the scene fingerprint is greater than the first preset threshold, and the number of images within the group in the historical scene fingerprint is the same as the number of images within the group in the scene fingerprint.

[0097] In this embodiment, for each test scenario group, all historical scenario fingerprints in the historical batch can be traversed. For each historical scenario fingerprint, the Jaccard similarity between the set of historical scenario keywords in the historical scenario fingerprint and the set of scenario keywords in the scenario fingerprint can be calculated. The number of images within the group in the historical scenario fingerprint is checked to see if it is equal to the number of images within the group in the scenario fingerprint. Historical scenario fingerprints with a Jaccard similarity greater than a first preset threshold and an equal number of images within the group are selected as candidate historical scenario fingerprints for matching. This step is used to filter out historical scenario groups that are consistent with the test scenario group in terms of scenario semantics and sample size from all historical scenario fingerprints, thus narrowing the scope of subsequent comprehensive similarity calculation.

[0098] Step S203: For each test scenario group, based on the scene fingerprint and the matched candidate historical scene fingerprint, obtain the comprehensive similarity, and obtain the highest comprehensive similarity from the comprehensive similarity. Based on the highest comprehensive similarity, determine whether the test scenario group is successfully matched.

[0099] In this embodiment, for each test scenario group, the scene fingerprint corresponding to that test scenario group is sequentially traversed through each matching candidate historical scene fingerprint. For the currently traversed candidate historical scene fingerprint, the Jaccard similarity of the current scene fingerprint with the candidate historical scene fingerprint on the scene keyword set, the edit distance similarity on the file name pattern, and the equality score on the sequence number group identifier are calculated. These three similarity scores are multiplied by their respective preset weight coefficients and then added together to obtain the comprehensive similarity between the test scenario group and the candidate historical scene fingerprint. The comprehensive similarity of all candidate historical scene fingerprints is compared, and the one with the largest value is selected as the highest comprehensive similarity. The highest comprehensive similarity is compared with a preset matching threshold. If the highest comprehensive similarity is greater than or equal to the matching threshold, the test scenario group is marked as a successful match. If the highest comprehensive similarity is less than the matching threshold, the test scenario group is marked as a failed match. This step is used to merge the multidimensional matching degree between the current test scenario group and multiple candidate historical scene fingerprints into a single comparable comprehensive value, select the most suitable matching result, and determine whether the test scenario group has successfully matched with the historical scene.

[0100] A successful match indicates that the test scenario group has a corresponding historical scenario in the historical batch, while an unsuccessful match indicates that the test scenario group does not have a corresponding historical scenario in the historical batch.

[0101] Step S204: For each test scene group in the batch of images to be processed, according to the preset level quantization mapping rules and based on the visual language model, obtain the structured indicators corresponding to the test scene group; wherein, the structured indicators include: the scene identifier of the test scene group, the batch code, and the set of indicators for the image quality dimension in the test scene group.

[0102] In this embodiment, for each test scene group in the batch of images to be processed, the images in the test scene group, along with preset analysis prompts, are input into a visual language model to obtain the natural language analysis results of the test scene group on each image quality dimension output by the visual language model. Then, the natural language analysis results, along with preset gradation quantization mapping rules, are input into the language model, instructing the language model to convert the natural language analysis results into a structured format output according to the gradation quantization mapping rules. This structured format includes the scene identifier, batch code, and indicator set corresponding to each image quality dimension of the test scene group. The indicator set for each image quality dimension includes defect level, gradation quantization score, defect description summary, and judgment confidence. This step is used to convert the natural language analysis conclusions generated by the visual language model into standardized indicator data that can be quantified, stored, and directly compared with historical data.

[0103] Optionally, the structured metrics include the scene identifier (scene_id), batch code (batch_code), and a set of metrics for the image quality dimension within the test scene group. The metric set includes defect grade (grade), grade quantification score (grade_score), defect description summary (description), and confidence level (confidence). The preset grade quantification mapping rules are: Blocker=0 points, Critical=25 points, Major=50 points, Minor=75 points, None=100 points. Regular expressions and rule templates can be used to replace LLM for structured metric extraction. Grade keywords for each dimension are extracted from the VLM analysis results through pattern matching. This solution does not rely on additional LLM calls and has lower token costs, but it is more dependent on the VLM output format and cannot handle non-standard expressions.

[0104] Structured metrics are stored in binary JSON format fields in a relational database. Analysis prompts require the language model's output to conform to a predefined JSON Schema structure, including all required dimensions and their ranking and score fields. The language model's output is then validated.

[0105] (a) Format validation: Check that the output is compliant JSON and contains all required dimension keys.

[0106] (b) Value range verification: Check whether the grade of each dimension is one of the predefined grade enumeration values ​​and whether the grade_score is within the valid range.

[0107] (c) Consistency check: Check whether the grade and score conform to the grade quantification mapping rules.

[0108] If the verification fails, the corrected prompt will be used to re-request (up to a maximum of the preset number of retries). If the retries still fail, the metric for that scenario will be marked as extraction failed, the incremental comparison for that scenario will be skipped, and an alarm log will be recorded.

[0109] In addition, the image quality dimensions include: exposure (AE), white balance (AWB), noise (ISP_noise), texture (ISP_texture), focus (AF), and overall. Overall reflects the overall image quality and perceived performance of the test scene group beyond individual dimensions such as exposure, white balance, noise, texture, and focus. Unlike other dimensions that focus on specific technical indicators, the overall dimension focuses on the overall visual perception level of the image, such as the clarity, color saturation, overall balance between detail retention and noise suppression, and the subjective perception of "image quality" by the human eye. The overall dimension, together with the other five individual dimensions, constitutes a complete six-dimensional evaluation system. Structured indicator extraction maintains >90% consistency with human judgment: Language model-driven structured indicator extraction achieved a 91.5% consistency rate in grade judgment on 200 manually labeled test sets, with a 96% consistency rate for Blocker and Critical grades.

[0110] In existing technologies, the image analysis results output by visual language models are presented in the form of natural language text, which cannot be directly converted into traceable numerical trend indicators. This results in a lack of quantifiable comparison criteria when comparing across batches. This application converts the natural language analysis results output by the visual language model for each image quality dimension into structured indicators containing defect levels and level quantification scores through a preset level quantification mapping rule. This allows the analysis results for each image quality dimension to be stored in a preset storage terminal in the form of numerical scores from 0 to 100. Subsequent batches can directly read these values ​​for dimensional quantification comparison and trend tracking, thereby solving the problem that natural language results cannot be quantified and tracked.

[0111] Step S205: For the test scene group marked as successfully matched, compare the structured indicators corresponding to the test scene group with the structured indicators of the historical scene group that was successfully matched by the test scene group to obtain the incremental comparison results of the batch images to be processed.

[0112] In this embodiment, for each test scene group marked as successfully matched, the historical structured indicators of the historical scene groups that were successfully paired with the test scene group can be read from the database. The level quantization score of each image quality dimension in the structured indicators of the test scene group is subtracted one by one from the level quantization score of the corresponding dimension in the historical structured indicators to obtain the score difference of each dimension. For each dimension, the change direction label of the dimension is determined according to the preset difference range in which the score difference of the dimension falls. The change direction labels of all dimensions are summarized into the incremental comparison result of the test scene group. This step is used to directly compare the quantitative performance of each test scene group in the current batch in each image quality dimension with the quantitative performance of the successfully paired historical scene group in the same dimension, quantify the specific change direction and magnitude of each dimension, and form a specific comparison conclusion of the current batch relative to the historical batch.

[0113] All test scene groups in the batch of images to be processed undergo structured index extraction. Although test scene groups marked as unsuccessful matches do not undergo incremental comparison, their structured indexes are still extracted and stored as historical baseline data for future batches for reuse in scene matching and incremental comparison. This prevents the scene group from losing the opportunity to participate in subsequent comparisons due to failure to match in the current batch, and allows the index data of each scene group to continuously accumulate as the batches progress.

[0114] For example, Figure 3 Example diagram of incremental comparison and structured index reuse process provided in the embodiments of this application, such as Figure 3 As shown, using a new batch of scenes as input, VLM analysis is performed only on the new batch. Then, LLM is used to extract standardized structured indicators containing dimensions, levels, and scores and store them in the analysis record. Subsequently, it is determined whether there is a matching historical scene. If there is no matching historical scene, the comparison step is skipped. If there is a matching historical scene, the corresponding historical scene structured indicators are retrieved from the database. The difference between the new and old indicators is calculated dimension by dimension to complete the incremental comparison. Based on the difference, the image quality improvement, stagnation, and regression are determined and the corresponding confidence level is labeled. The complete single-round cross-batch incremental comparison results are output. The entire process only performs model parsing and reuses the stored historical indicators for the new batch of images, thereby reducing the overhead of repeated model calls and realizing automated cross-batch image quality incremental quantitative comparison.

[0115] In this embodiment, multiple test scene groups and corresponding scene fingerprints are obtained by first identifying and processing the batch images to be processed. Candidate historical scene fingerprints are filtered based on the conditions that the similarity of scene keywords is greater than a first preset threshold and the number of images in the group is the same. Then, the comprehensive similarity is calculated and the highest comprehensive similarity is taken to determine whether the test scene group is marked as a successful match. For each test scene group in the batch images to be processed, a structured index is obtained based on a visual language model combined with a graded quantization mapping rule. For the test scene group marked as a successful match, the structured index is compared with the structured index of the matching historical scene group to obtain an incremental comparison result. The pairing of different batches of test scene groups is automatically completed by the above process, eliminating the need for manual labeling and pairing operations, reducing manpower and time consumption, and solving the technical problem of low image comparison efficiency in the prior art.

[0116] Furthermore, in the above Figure 2 Based on the illustrated embodiment, the image quality dimensions include: exposure, white balance, noise, texture, focus, and / or a combination thereof. The set of indicators for the image quality dimensions includes: defect level, level quantization score, defect description summary, and judgment confidence level. Therefore, step S205 is specifically implemented as follows:

[0117] Step S2051: Based on the image quality dimension index set in the structured indicators corresponding to the test scene group, compare it dimension by dimension with the image quality dimension index set in the structured indicators of the historical scene group that was successfully matched with the test scene group, and obtain the change in each image quality dimension; wherein, the structured indicators of the historical scene group are obtained by reading data from the preset storage terminal.

[0118] In this embodiment, for each test scene group marked as successfully matched, each quality dimension among exposure, white balance, noise, texture, focus, and overall quality can be sequentially traversed. For the currently traversed quality dimension, the level quantization score corresponding to that dimension is read from the structured index of that test scene group. The level quantization score of the same quality dimension is read from the structured index of successfully matched historical scene groups. The level quantization score of the current batch is subtracted from the level quantization score of the historical batch to obtain the score difference of that dimension. The score difference is recorded as the change amount of that quality dimension, corresponding to the name of that quality dimension. This step is used to decompose the comparison of multiple image quality dimensions within the same test scene group into independent item-by-item comparisons for each dimension, and obtain the change direction and change magnitude of exposure, white balance, noise, texture, focus, and overall quality respectively.

[0119] The structured metrics of historical scene groups are obtained by reading data from a pre-set storage source. This means that during cross-batch comparisons, the structured metrics of successfully matched historical scene groups are not obtained by re-analyzing the images in the historical scene group using the visual language model. Instead, they are directly read from the pre-set storage source (e.g., a database or caching system) from the structured metric records that were extracted and stored during previous batch analyses. This reuse mechanism avoids repeatedly performing visual language model analysis on historical images during each cross-batch comparison, reducing the number of model calls and token consumption in the cross-batch comparison stage, and lowering the computational cost and time overhead of image comparison.

[0120] For example, for each aligned scene pair (new_scene, hist_scene), the structured metrics (new_metrics) of the currently matched image group in the current batch are read, and the structured metrics (hist_metrics) of the historical scenes are read directly from the database without calling the API. new_scene refers to the test scene group in the current batch of images that has been marked as "successfully matched" after scene matching, and hist_scene refers to the historical scene group that successfully paired with the aforementioned new_scene during the historical scene fingerprint matching process.

[0121] Dimensionally compare the changes in scores, and calculate the change in score as delta = new_metrics[dim].grade_score - hist_metrics[dim].grade_score, according to the following rules:

[0122] delta>TH_IMPROVE_MAJOR: Significant improvement (↑).

[0123] delta>TH_IMPROVE_MINOR: Slight improvement (↗).

[0124] -TH_REGRESS_MINOR≤delta≤TH_IMPROVE_MINOR: Balanced (→).

[0125] delta<-TH_REGRESS_MINOR: Mild regression (↘).

[0126] delta <- TH_REGRESS_MAJOR: Significant regression (↓).

[0127] Where TH_IMPROVE_MAJOR is the threshold for significant improvement of the fractional change delta, TH_IMPROVE_MINOR is the threshold for slight improvement of delta, TH_REGRESS_MINOR is the threshold for slight regression of delta, and TH_REGRESS_MAJOR is the threshold for significant regression of delta. The values ​​of TH_IMPROVE_MAJOR and TH_REGRESS_MAJOR range from 12 to 18, and the values ​​of TH_IMPROVE_MINOR and TH_REGRESS_MINOR range from 3 to 7. We can set TH_IMPROVE_MAJOR=15, TH_IMPROVE_MINOR=5, TH_REGRESS_MINOR=5, and TH_REGRESS_MAJOR=15.

[0128] If any dimension shows significant regression, a regression alert is triggered. If the number of dimensions improved exceeds the number of regression dimensions, the overall situation improves; otherwise, the overall situation regresses; otherwise, the situation remains basically unchanged.

[0129] In existing technologies, each time a new batch arrives, all historical images are re-input into the visual language model for analysis along with the images from the new batch. This causes the token cost to increase exponentially with the number of batches, i.e., O(N). 2 The increasing number of batches of images makes full re-analysis of token costs unacceptable. This application adopts an incremental comparison strategy, that is, in the process of cross-batch image comparison, only the visual language model is called for analysis of the images of the new batch to obtain the structured indicators of the new batch. The structured indicators of the historical batches have been extracted and stored in the database in the previous analysis, and can be directly read from the database for reuse during comparison. There is no need to call the visual language model to analyze the historical images again. Visual language model analysis is only performed once for the images of the new batch, so that the token cost increases linearly with the number of batches, that is, O(N), thus solving the problem of unacceptable token cost for full re-analysis. Based on the above incremental comparison strategy, for each test scene group that has completed scene matching, a dimension-by-dimensional comparison operation is performed to obtain the specific changes in each quality dimension. The token cost advantage of the incremental comparison strategy can be described as follows:

[0130]

[0131] This table compares the token costs of the full re-analysis scheme and the incremental comparison scheme of this invention at different batch sizes. For single-batch processing, both schemes consume 200,000 tokens. As the batch size increases, the cumulative token consumption of the full re-analysis scheme decreases by O(N). 2The total number of tokens increased from 2,200,000 in 10 batches to 25,500,000 in 50 batches. In contrast, the incremental comparison scheme in this application only analyzes new batches and reuses historical structured indicators, resulting in an O(N) growth rate for the cumulative tokens: 2,000,000 in 10 batches and 10,000,000 in 50 batches. Compared to full reanalysis, this application reduces costs by 9% in 10 batches and by 61% in 50 batches. In the 100-batch scenario, the cost reduction can reach over 80%.

[0132] Alternatively, VLM can be used to directly compare images across batches (by simultaneously inputting images from both new and old batches into VLM) instead of structured metric comparison. This approach leverages VLM's cross-image understanding capabilities for more nuanced comparisons, but the token cost is significantly higher than structured metric comparison (which requires resending historical images), and historical analysis results cannot be reused.

[0133] In this embodiment, the specific types of image quality dimensions and the constituent elements of each dimension's indicator set are defined. Based on this, the structured indicator comparison is limited to a dimension-by-dimensional comparison, obtaining the change in each image quality dimension. Through the establishment of the above-mentioned multi-dimensional indicator system and the dimension-by-dimensional quantitative comparison, cross-batch image comparison is no longer a general overall comparison, but a refined comparison of each quality dimension, such as exposure, white balance, noise, texture, focus, and overall quality. Each dimension has a graded quantitative score that can intuitively reflect the direction and magnitude of change, thereby accurately locating the specific changes of each image quality dimension before and after firmware iteration, improving the precision and practicality of cross-batch image comparison results.

[0134] Figure 4 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 2 In the above Figure 2 Based on the illustrated embodiment, the method may further include:

[0135] Step S401: Obtain the incremental comparison results of multiple batches of images and the structured indicators under each batch of images.

[0136] In this embodiment, for each test scenario group, the structured indicator records stored in all historical batches of the test scenario group can be read from the database. At the same time, the change direction label records of each dimension generated in each incremental comparison of the test scenario group can be read. These records are arranged in ascending order of batch code as the structured indicators and incremental comparison results of the test scenario group in multiple batches. This step is used to collect the quantitative indicators and comparison conclusions of the same test scenario group in multiple historical batches from the database, and prepare data materials for constructing the time series sequence of the scenario group.

[0137] Step S402: Construct a time series based on the structured indicators under each batch of images; wherein, the time series includes: scene mode, image quality dimension and data point list, and the data point list includes: batch code, analysis timestamp, defect level, level quantification score and number of samples.

[0138] In this embodiment, the structured indicators of each batch can be aggregated according to three dimensions: "scene mode × image quality dimension × analysis timestamp" to construct a TimelineSeries. For each test scene group, the scene identifier of the test scene group is used as the value of the scene mode field. Each quality dimension in exposure, white balance, noise, texture, focus, and overall is traversed in turn. For the currently traversed quality dimension, the batch code, analysis timestamp, defect level, level quantification score, and sample number corresponding to that dimension are extracted from the structured indicators of each batch under that test scene group. This information is combined into a data point. For the same quality dimension, the data points of each batch are arranged from early to late according to the analysis timestamp to form a complete list of data points for that quality dimension under that scene mode on the time axis. This reorganizes the structured indicators scattered in each batch into a sequence in chronological order. The scene mode, quality dimension name, and data point list are combined to obtain the TimelineSeries of that test scene group under that quality dimension. After the above aggregation, each test scene group corresponds to an independent TimelineSeries in six quality dimensions: exposure, white balance, noise, texture, focus, and overall performance, which prepares input data for subsequent linear regression calculations.

[0139] Here, the sample size refers to the total number of images contained within a specific test scenario group when performing visual language model analysis. In structured metrics, the sample size for each test scenario group records the number of images from that group used in the analysis, reflecting the sample size for that batch and scenario group. In time series, the sample size, as one of the components of the data point list, records the number of images corresponding to historical data points, and can be used as a reference for weighting or determining data reliability in linear regression and significance tests. The larger the sample size, the more representative the data point is.

[0140] Step S403: Based on the time series, use linear regression and significance test to obtain the trend direction, linear fit degree and significance probability of each image quality dimension.

[0141] In this embodiment, based on the time series, the batch code sequence number of each data point in the data point list is used as the independent variable x, and the grade quantization score of each data point is used as the dependent variable y. The least squares method is used to fit a linear regression equation to the data points in the time series to obtain the slope value and the coefficient of determination value. The slope value is compared with multiple preset slope thresholds, and the trend direction of the image quality dimension is determined according to the threshold interval in which the slope value falls. The coefficient of determination value is used as the linear fit of the trend direction. At the same time, it is checked whether the number of data points in the time series has reached a preset number. If the preset number has been reached, a t-test is performed on the slope value to obtain the significance probability of the slope value. This step is used to convert the grade quantization scores accumulated in different batches into trend direction, linear fit, and significance probability, providing basic data for the generation of comprehensive trend results.

[0142] For example, we can use batch_index(0, 1, 2, ...) as the independent variable x and grade_score as the dependent variable y, fitting the equation y = slope·x + intercept. batch_index is the batch code of each data point in the time series, sorted by the analysis timestamp from earliest to latest. grade_score is the grade quantification score corresponding to a specific image quality dimension in the structured index, calculated using the following formula:

[0143]

[0144]

[0145] Where slope is the gradient. Let f(i) be the batch_index of the i-th data point in the time series. The arithmetic mean of all data points at batch_index. Let grade_score be the i-th data point in the time series. The grade_score is the arithmetic mean of all data points, and the intercept is the cutoff.

[0146] Determine the trend direction of each image quality dimension based on the slope:

[0147] slope>TH_SLOPE_MAJOR: Continuous improvement.

[0148] slope>TH_SLOPE_MINOR: Slow improvement.

[0149] -TH_SLOPE_MINOR≤slope≤TH_SLOPE_MINOR: Stable.

[0150] slope<-TH_SLOPE_MINOR: Slow regression.

[0151] slope<-TH_SLOPE_MAJOR: Continuous regression.

[0152] Wherein, TH_SLOPE_MAJOR is the continuous improvement threshold of slope, TH_SLOPE_MINOR is the slow improvement threshold of slope, -TH_SLOPE_MINOR is the slow regression threshold of slope, and -TH_SLOPE_MAJOR is the continuous regression threshold of slope, and TH_SLOPE_MAJOR=3 and TH_SLOPE_MINOR=1 can be taken.

[0153] Confidence calculation for each image quality dimension: through R 2 The coefficient of determination evaluates the degree of linear fitting of the trend, R 2 >TH_R2_HIGH is high confidence, R 2 >TH_R2_LOW is medium confidence, R 2 ≤TH_R2_LOW is low confidence, TH_R2_HIGH is the threshold for determining high confidence, TH_R2_LOW is the demarcation threshold for determining medium confidence and low confidence, and TH_R2_HIGH=0.7 and TH_R2_LOW=0.4 can be taken.

[0154] Significance test for each image quality dimension: when the number of batches ≥ 3, a t-test is performed, p<TH_P_HIGH is extremely significant, p<TH_P_MID is significant, p<TH_P_LOW is marginally significant, p≥TH_P_LOW is not significant, TH_P_HIGH is the extremely significant threshold, TH_P_MID is the significant threshold, TH_P_LOW is the demarcation threshold between marginal significance and non-significance, and TH_P_HIGH=0.01, TH_P_MID=0.05, TH_P_LOW=0.10 can be taken, through the triple evaluation of linear regression + R 2 coefficient of determination + t-test, the trend judgment not only has a directional conclusion (improvement / regression), but also is supported by confidence. In 10 batches of simulated data, for the truly existing trend (slope=-5.2), R 2 =0.82 (high confidence), p=0.003 (extremely significant).

[0155] Furthermore, while the number of data points in a time series is numerically equal to the number of batches, they are not the same concept. The number of batches represents the actual number of batches for which imaging tests were performed, while the number of valid historical data points is the number of records in the time series that actually participated in the regression calculation after removing expired data. If the structured indicator corresponding to a certain batch is filtered out or marked as expired due to anomalies and not included in the calculation, the number of valid historical data points will be less than the actual number of batches. In this case, the number of historical data points in the time series that were not marked as expired should be used as the criterion for deciding whether to perform a t-test, rather than directly using the original number of test batches.

[0156] Alternatively, nonparametric tests (such as the Mann-Kendall trend test) can be used instead of linear regression for trend detection. The Mann-Kendall test does not require the data to follow a normal distribution and is more robust to outliers, but it cannot provide a quantitative indicator of trend strength such as the slope, and its effectiveness is low with small samples (<5 batches).

[0157] Figure 5 Example diagram of trend detection and regression alarm process provided in the embodiments of this application, such as Figure 5 As shown, using a time-series index sequence aggregated by scenario, dimension, and time as input, a linear regression fitting operation is first performed to obtain the slope and intercept, and then R is calculated. 2 The coefficient of determination is used to assess the degree of linearity of the trend. Then, a t-test is performed to determine the significance of the data. Based on the slope value obtained from the fitting, the trend is classified into five categories: continuous improvement, slow improvement, stable, slow regression, and continuous regression. Then, the regression alarm check is performed. If an alarm is triggered, regression alarm information including level, dimension, score change delta, and slope is output. If no alarm is triggered, the normal trend record is retained. Both types of results are sent to LLM to generate a trend summary report in natural language. This fully realizes the quantitative determination of multi-batch image quality time-series trends, anomaly regression identification, and visualization report output.

[0158] Step S404: Based on the incremental comparison results of multiple batches of images, the trend direction of each image quality dimension, the degree of linear fitting, and the significance probability, obtain the comprehensive trend result.

[0159] In this embodiment, the trend direction of each image quality dimension can be mapped to a corresponding trend score. The trend scores of each dimension are multiplied by their respective dimension weights and then summed to obtain a comprehensive trend score. The significance probability of each image quality dimension is compared with a preset significance threshold, and the significance level of each image quality dimension is determined based on the comparison results. The trend directions of each image quality dimension are classified to obtain an improvement dimension list, a regression dimension list, and a stable dimension list. Based on the incremental comparison results of multiple batches of images, whether a dimension regresses, and the linear fit degree and significance probability of the regressing dimension, a regression alarm flag and a regression severity level are generated. The comprehensive trend score, the significance level of each dimension, the improvement dimension list, the regression dimension list, the stable dimension list, the regression alarm flag, and the regression severity level are input into a large language model, and the large language model outputs a trend insight report in natural language form. This step is used to convert the trend direction, linear fit degree, and significance probability into a comprehensive trend score and a trend insight report.

[0160] For example, the trend levels of each dimension can be mapped to trend scores and then weighted and summed. The score mapping rules for each trend level are as follows: Continuous Improvement = +SCORE_MAJOR, Slow Improvement = +SCORE_MINOR, Stable = 0, Slow Regression = -SCORE_MINOR, Continuous Regression = -SCORE_MAJOR. +SCORE_MAJOR is the positive significant trend score corresponding to continuous improvement, +SCORE_MINOR is the positive slight trend score corresponding to slow improvement, -SCORE_MINOR is the negative slight trend score corresponding to slow regression, and -SCORE_MAJOR is the negative significant trend score corresponding to continuous regression. We can take SCORE_MAJOR = 30 and SCORE_MINOR = 10, with a score range of -100 to +100. The weighted summation formula is:

[0161]

[0162] Wherein, trend_score is the overall trend score, dim_trend_score is the trend level score corresponding to a certain image quality dimension (dim), and dim_weight is the preset weight coefficient corresponding to a certain image quality dimension (dim). The dimension coefficients are determined by the preset WEIGHT dictionary, and can be set to AE=0.2, AWB=0.2, ISP_noise=0.2, ISP_texture=0.15, AF=0.15, overall=0.1, with the sum of all weights being 1.0.

[0163] In addition, a regression alert mechanism can also be set: the trigger condition is that any dimension simultaneously satisfies that the level score drops by more than TH_ALERT_THRESHOLD compared with the previous batch, the linear regression slope < -TH_SLOPE_MINOR, and the significance p < TH_P_LOW (or 2 consecutive drops when the number of batches is insufficient), where TH_ALERT_THRESHOLD is a preset alert trigger threshold. The alert levels include Critical (comprehensive dimension regression), Major (exposure / white balance / noise / texture dimension regression), and Minor (any dimension regression). The regression alert mechanism triggers an alert immediately when a significant regression is detected, including the regression dimension, severity, change trend, and optimization suggestions. In the simulation scenario, the regression alert is about 2 test cycles earlier than manual discovery.

[0164] After obtaining the comprehensive trend result of each test scenario group, the comprehensive trend result can also be input into the large language model, and the large language model generates a trend insight report in the form of natural language based on a preset summary prompt. This calling process only relies on the plain text interface of the large language model, and does not need to attach any image data to the model, so the calling overhead for generating trend insights each time is low.

[0165] In this embodiment, the structured indicators of multiple batches are aggregated into a time series according to the scenario mode, image quality dimension and time dimension, so that the single comparison results scattered in each batch can be organized into an analyzable data sequence in chronological order. On this basis, linear regression is used to fit the level quantification score of each data point in the time series, so as to obtain the trend direction and linear fitting degree of each image quality dimension, to characterize the long-term change trend of this dimension across multiple firmware versions and the linear credibility of the change trajectory. At the same time, a significance test is used to analyze the above trend to obtain the significance probability to judge whether the trend may be caused by random fluctuations. Finally, the incremental comparison results accumulated in multiple batches are combined with the above trend direction, linear fitting degree and significance probability to obtain a comprehensive trend result. Cross-batch image comparison is no longer limited to pairwise comparison between adjacent batches, but can perform quantitative analysis and significance verification of long-term trends based on data of multiple historical batches, thereby providing a more global comprehensive trend judgment than a single incremental comparison result.

[0166] Further, on the basis of the above Figure 4 illustrated embodiment, the specific implementation of step S403 is as follows:

[0167] Step S4031: According to the time span between the analysis timestamp corresponding to each historical data point in the time series and the analysis timestamp of the batch image to be processed, determine the time decay weight of each historical data point respectively; wherein, the time decay weight decreases monotonically as the time span increases.

[0168] In this embodiment, the time span between the analysis timestamp of each historical data point in the time series and the analysis timestamp of the batch image to be processed can be calculated. A decay coefficient is set, and the time decay weight of each historical data point is determined as the result of a power operation with the natural constant e as the base and the negative decay coefficient multiplied by the time span as the exponent. This makes the time decay weight of historical data points with larger time spans smaller. This step is used to control the influence of each historical data point in the subsequent weighted linear regression, so that the historical data points closer to the current batch contribute more to the weighted slope, and the historical data points farther away from the current batch contribute less to the weighted slope.

[0169] Step S4032: Mark historical data points with a time decay weight less than the preset expiration threshold as expired data.

[0170] In this embodiment, a preset value can be used as an expiration threshold. The time decay weight of each historical data point in the time series is compared with the expiration threshold. Historical data points with a time decay weight less than the expiration threshold are marked as expired data. This step is used to remove historical data points with low influence due to being too far removed from the data point list before subsequent weighted linear regression, and retain historical data points with decay weights reaching the expiration threshold to participate in subsequent calculations.

[0171] Step S4033: Using time decay weight as a weighting coefficient, perform weighted linear regression on the rank quantification scores of each historical data point in the time series that has not been marked as expired data, and obtain the weighted slope and linear fit degree.

[0172] In this embodiment, the time decay weight of each historical data point that is not marked as expired can be used as the weighting coefficient of that data point. The batch code sequence number of each data point is used as the independent variable, and the grade quantization score of each data point is used as the dependent variable. In the objective function of the least squares method, the squared error of each data point is multiplied by the corresponding weighting coefficient. The slope and intercept values ​​that minimize the sum of weighted squared errors are solved to obtain the weighted slope value. Then, based on the difference between the actual grade quantization score of each data point and the predicted value given by the weighted linear regression equation, the coefficient of determination is calculated as the linear fit of the regression. This step is used to fit a regression line from the weighted historical data points to describe the overall change trend of these data points on the time axis before determining the trend direction based on the weighted slope.

[0173] Furthermore, in existing cross-batch trend analysis techniques, the image quality analysis results of each historical batch originate from firmware versions from different periods. Different batches are at different stages of firmware evolution, and the reference value of their analysis results for current trend judgment gradually weakens over time. If all historical batch data are treated equally, the analysis results of earlier batches have the same influence on trend judgment as those of recent batches. This may lead to excessive interference from early abnormal batches or outdated data on the current trend, causing the trend direction to deviate from the true path of change. To address the issue of historical data timeliness decay and enable cross-batch trend analysis to adaptively adjust the influence of historical data, a time decay model can be set up:

[0174] Define the exponential decay function as:

[0175]

[0176] in, λ is the decay exponent, λ is the decay rate parameter, Δt is the difference between the current batch time and the historical batch time, and the value of λ determines the half-life of the historical data. λ can be taken as 0.005 (corresponding to a half-life of about 138 days). The characteristic of its value is: when Δt=0 =1.00 (complete trust), Δt is the half-life. ≈0.50, When large enough Approaching 0.

[0177] In linear regression for trend detection, time decay weights can be introduced into the weighted regression calculation. The formula for calculating the weighted slope is:

[0178]

[0179] Where slope1 is the weighted slope, α i Let x be the time decay weight corresponding to the i-th batch of samples. i Let y be the batch number corresponding to the i-th sample. i The image level quantization score corresponding to the i-th sample. This is the weighted average of the batch serial numbers. The weighted mean of the image grade quantization scores. Based on the aforementioned time decay confidence model, this application further introduces an incremental learning mechanism, enabling trend analysis to dynamically update model parameters with the arrival of new batches. This mechanism includes:

[0180] Once the structured metrics for the new batch are extracted, for each test scenario group marked as a successful match, the scenario alignment results obtained from the scenario matching layer are directly reused to obtain the historical scenario group corresponding to the test scenario group. The structured metrics of the new batch are then associated with the historical scenario group as new data points for the test scenario group in the time series.

[0181] Before adding new data points to the time series, calculate the time decay weights for all historical data points corresponding to this scenario group. If a certain historical data point If the data point is below a preset expiration threshold TH_EXPIRED (e.g., TH_EXPIRED = 0.10), it is marked as expired data. This expired data does not participate in subsequent trend calculations but is retained in the database for auditing and tracing. This step ensures that historical data that is too far removed from the present and has too low a decay weight will not interfere with the current trend judgment.

[0182] Before adding the new batch of structured indicators as new data points to the time series, the newly added data points are predicted using the weighted linear regression equation before the update (i.e., the regression equation fitted only based on all historical valid data points of this test scenario group on this quality dimension). Calculate the actual rank quantification score y and the predicted value for this new data point. The residual is then divided by the standard deviation of the residuals of all historical valid data points in the current time series to obtain the standardized deviation. If this standardized deviation exceeds the preset mutation point detection threshold N_SIGMA (for example, N_SIGMA=3), the new data point is marked as a "mutation point," triggering a special alarm, indicating that the quality performance of this dimension has deviated from the historical trend. After this detection is completed, the new data point is added to the time series, and the weighted slope slope1 is recalculated to complete the incremental update of the model. Here, y is the actual level quantization score of a certain test scene group in the current batch on a certain image quality dimension. The predicted level quantification score is calculated by substituting the batch_index of the newly added data points as the independent variable x into the weighted linear regression equation before the update.

[0183] To verify the practical effect of the time decay model and incremental learning mechanism in long-term trend tracking, this embodiment conducted a comparative experiment on 30 batches of firmware iteration test datasets. The equal-weighted comparison scheme without time decay was compared with the scheme incorporating the time decay model. The experimental comparison results can be expressed as follows:

[0184]

[0185] Regarding trend prediction accuracy, the time decay model achieved 85% accuracy with 10 batches, a 7 percentage point improvement over the 78% of the equal-weighted comparison scheme. With 30 batches, the accuracy of the equal-weighted comparison scheme dropped significantly from 78% to 62%, while the time decay model only slightly decreased from 85% to 83%, maintaining a high level of accuracy. This indicates that the time decay of historical data has a significant impact on trend prediction accuracy, and the introduction of time decay weights effectively suppresses interference from distant historical data, ensuring stable trend prediction in long-term tracking scenarios. In terms of false alarm rate, the time decay model achieved 5%, a 7 percentage point reduction compared to the 12% of the equal-weighted comparison scheme. Regarding abrupt change detection rate, the equal-weighted comparison scheme, lacking an anomaly detection mechanism, failed to detect abrupt changes, while the time decay model, combined with the standardized deviation assessment method, achieved a 91% abrupt change detection rate.

[0186] This application determines the time decay weight of each historical data point based on the time span between the analysis timestamp of historical data points and the analysis timestamp of the current batch. This reduces the influence of historical data points with a longer time span from the current batch. Expired data that is too old is marked and excluded from the weighted linear regression calculation, provided that the decay weight is less than a preset expiration threshold. This ensures that the trend judgment results mainly reflect the quality change trend of recent valid batches, avoids interference from early abnormal data on the current trend judgment, and solves the problem of the impact of the time decay of historical data on the accuracy of trend analysis.

[0187] Alternatively, a sliding window strategy can be used instead of the exponential decay model. Only the most recent N batches of data are retained for trend calculation, while earlier batches are completely discarded. This approach is simpler to implement, but hard truncation can lead to the sudden loss of useful information, unlike the smoother soft decay of exponential decay.

[0188] Step S4034: Obtain the trend direction of each image quality dimension based on the weighted slope and the preset slope threshold.

[0189] In this embodiment, a preset first slope threshold and a second slope threshold can be used as the determination boundary. The first slope threshold is greater than the second slope threshold. The weighted slope is compared with the positive first slope threshold, the positive second slope threshold, the negative second slope threshold, and the negative first slope threshold, respectively. When the weighted slope is greater than the positive first slope threshold, the trend direction is determined to be continuous improvement. When the weighted slope is greater than the positive second slope threshold and less than or equal to the positive first slope threshold, the trend direction is determined to be slow improvement. When the weighted slope is greater than or equal to the negative second slope threshold and less than or equal to the positive second slope threshold, the trend direction is determined to be stable. When the weighted slope is greater than or equal to the negative first slope threshold and less than the negative second slope threshold, the trend direction is determined to be slow regression. When the weighted slope is less than the negative first slope threshold, the trend direction is determined to be continuous regression. This step is used to convert the continuous value of the weighted slope into one of five discrete trend direction categories.

[0190] Step S4035: When the number of historical data points in the time series that have not been marked as expired data is greater than or equal to a preset number, a significance test is performed based on the weighted slope to obtain the significance probability.

[0191] In this embodiment, a preset number can be set as the minimum number of data points required for performing credibility determination. When the number of historical data points in the time series that have not been marked as expired data reaches or exceeds the preset number, the value of the weighted slope is compared with a reference distribution constructed based on the dispersion of the quantization scores of each data point around the regression line. According to the tail position of the weighted slope value in the reference distribution, the random fluctuation probability corresponding to the weighted slope is determined. This step is used to determine whether the weighted slope truly reflects the real changes in the image quality dimension among multiple batches, or is only caused by random fluctuations of data points.

[0192] In this embodiment, by introducing a time decay weight in the time series analysis, the influence of historical data points with a longer time span from the current batch on trend determination is reduced. Furthermore, based on this, excessively old historical data is filtered out when the decay weight is less than a preset expiration threshold, preventing it from participating in subsequent weighted linear regression. This allows for the retention of all historical data for audit traceability while only allowing data points with high timeliness to participate in the calculation of trend direction and linear fit, achieving adaptive adjustment of the influence of historical data in trend detection. Simultaneously, since the weighted linear regression uses the time decay weight as the weighting coefficient, the calculated weighted slope and linear fit can more realistically reflect the quality change trend of recent historical batches. Combined with the preset slope threshold, the trend direction determination result is closer to the actual trend of the current firmware iteration stage, avoiding misjudgments of trend direction caused by interference from early abnormal batches or historical data decay. In addition, a significance test based on the weighted slope is only performed when there are sufficient historical data points not marked as expired, improving the accuracy and anti-interference capability of long-term cross-batch trend detection.

[0193] Furthermore, in the above Figure 4 Based on the illustrated embodiment, the comprehensive trend result includes a comprehensive trend score and a trend insight report. Step S404 is specifically implemented as follows:

[0194] Step S4041: Determine the trend score corresponding to each image quality dimension based on the trend direction of each image quality dimension.

[0195] In this embodiment, corresponding trend score values ​​can be set for each of the five trend directions: continuous improvement, slow improvement, stable, slow regression, and continuous regression. Continuous improvement corresponds to the maximum positive value, slow improvement corresponds to the minimum positive value, stable corresponds to zero, slow regression corresponds to the minimum negative value, and continuous regression corresponds to the maximum negative value. The trend direction of each image quality dimension is matched with the above five trend directions, and the trend score corresponding to that dimension is determined based on the matching result. This step is used to convert the discrete category of trend direction into a value that can participate in the weighted summation operation, so that a comprehensive trend score can be obtained by weighted fusion of the trend scores of each dimension.

[0196] Step S4042: Multiply the trend score corresponding to each image quality dimension by its respective preset dimension weight and sum them to obtain the comprehensive trend score.

[0197] In this embodiment, a dimension weight can be set for each image quality dimension, and the sum of the dimension weights of each dimension is equal to one. The trend score corresponding to each image quality dimension is multiplied by the dimension weight corresponding to that dimension to obtain the weighted trend score of each dimension. Then, the weighted trend scores of all dimensions are added together to obtain the comprehensive trend score. This step is used to merge the independent trend scores of each image quality dimension into a single value to reflect the overall direction and magnitude of the overall trend of all dimensions.

[0198] Step S4043: Based on the comparison results between the saliency probability and the preset saliency threshold, determine the saliency level corresponding to each image quality dimension.

[0199] In this embodiment, a first significance threshold, a second significance threshold, and a third significance threshold can be set, wherein the first significance threshold is less than the second significance threshold, and the second significance threshold is less than the third significance threshold. The significance probability of each image quality dimension is compared with the first significance threshold, the second significance threshold, and the third significance threshold, respectively. When the significance probability is less than the first significance threshold, the significance level corresponding to that dimension is determined to be highly significant. When the significance probability is greater than or equal to the first significance threshold and less than the second significance threshold, the significance level corresponding to that dimension is determined to be significant. When the significance probability is greater than or equal to the second significance threshold and less than the third significance threshold, the significance level corresponding to that dimension is determined to be marginally significant. When the significance probability is greater than or equal to the third significance threshold, the significance level corresponding to that dimension is determined to be insignificant. This step is used to convert the significance probabilities of each dimension into discrete significance levels, which are used in subsequent trend insight reports to describe the credibility of the trends of each dimension.

[0200] Step S4044: Classify the trend direction of each image quality dimension to obtain an improvement dimension list, a regression dimension list, and a stable dimension list.

[0201] In this embodiment, a dimension identifier can be configured for each image quality dimension, and three empty lists can be created to store the improvement dimension, regression dimension, and stable dimension, respectively. All image quality dimensions are traversed, and each dimension is classified according to its trend direction: when the trend direction is continuous improvement or slow improvement, the dimension identifier of that dimension is added to the improvement dimension list; when the trend direction is continuous regression or slow regression, the dimension identifier of that dimension is added to the regression dimension list; and when the trend direction is stable, the dimension identifier of that dimension is added to the stable dimension list. This step is used to classify each image quality dimension into three groups of improvement, regression, and stability according to the trend direction, so that the dimensions with different trends can be displayed separately in the subsequent trend insight report.

[0202] Step S4045: Based on the incremental comparison results of multiple batches of images, as well as the trend direction, linear fitting degree and significance probability of each image quality dimension, generate regression alarm signs and regression severity levels.

[0203] In this embodiment, an alarm triggering condition can be set, which requires the simultaneous fulfillment of three sub-conditions: the first sub-condition is that the trend direction of the image quality dimension is the regression direction; the second sub-condition is that the significance probability of the dimension is lower than a preset threshold; and the third sub-condition is that the change direction of the dimension in the latest batch is the regression direction based on the incremental comparison results of multiple batches of images. When an image quality dimension simultaneously meets the above three sub-conditions, a regression alarm flag containing the dimension identifier is generated, and the regression severity level is determined according to the image quality type corresponding to the dimension. This step is used to filter out dimensions from multiple dimensions that show regression in both trend and latest single comparison and have reliable significance, and generate alarm information.

[0204] Step S4046: Input the trend score, overall trend score, significance level, improvement dimension list, regression dimension list, stable dimension list, regression alarm flag, and regression severity level corresponding to each image quality dimension into the large language model to obtain the trend insight report.

[0205] In this embodiment, the trend score, comprehensive trend score, significance level, improvement dimension list, regression dimension list, stability dimension list, regression alarm flag, and regression severity level corresponding to each image quality dimension can be formatted into an input text according to a preset template. This input text is sent as a prompt word to the large language model, and the output format of the large language model is set to natural language paragraph. The text content returned by the large language model is received as a trend insight report. This step is used to convert the trend direction, significance level, classification list, and alarm information of each dimension into a natural language description that can be directly read by humans, avoiding the need for manual review of multiple scattered structured data items.

[0206] In this embodiment, the trend directions of each image quality dimension are mapped to quantifiable trend scores, and then weighted and fused using preset dimension weights. The trend judgment results of multiple dimensions are aggregated into a single comprehensive trend score, allowing the overall quality trend to be represented intuitively with numerical values. Simultaneously, the significance level of each dimension is determined based on the significance probability, and the trend directions of each dimension are classified to obtain a list of improvement / regression / stabilization dimensions. Combined with the incremental comparison results accumulated from multiple batches and the trend directions, linear fitting degree, and significance probability of each dimension, regression alarm signs and severity levels are generated. All the above structured information is input into a large language model, which outputs a trend insight report in natural language form. The trend analysis results are no longer limited to scattered indicators or values, but are transformed into structured data covering overall trends, dimension details, credibility, and alarm information. This data is then interpreted by the large language model into a text report that is easy to read directly, reducing the threshold for manual interpretation of multi-dimensional trend data and providing an intuitive and operable basis for quality trend decisions in firmware iteration.

[0207] Figure 6 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 3 In the above Figure 2 Based on the illustrated embodiment, the method may further include:

[0208] Step S601: When the structured indicators corresponding to the test scenario group are obtained, the historical scenario group corresponding to the test scenario group is determined based on the correspondence established by the successful matching of the test scenario group with the historical scenario group.

[0209] In this embodiment, when a test scenario group is successfully matched, the correspondence between the test scenario group and the successfully matched historical scenario group can be stored as a pairing record. The pairing record contains the scenario identifier of the test scenario group and the scenario identifier of the successfully matched historical scenario group. When the structured index corresponding to the test scenario group is obtained, the scenario identifier of the corresponding historical scenario group is found from the pairing record based on the scenario identifier of the test scenario group. Then, all the data of the corresponding historical scenario group is obtained based on the found scenario identifier of the historical scenario group. This step is used to obtain the historical scenario group corresponding to the test scenario group, so that the data of the historical scenario group can be used for prediction and mutation point determination in subsequent steps.

[0210] Step S602: Before adding the structured indicators of the test scenario group as new data points to the time series corresponding to the historical scenario group, the new data points are predicted using a weighted linear regression equation fitted based on historical data points in the historical scenario group that have not been marked as expired.

[0211] In this embodiment, after obtaining the historical scene group, the slope and intercept values ​​of the weighted linear regression equation corresponding to the historical scene group can be read from the storage. The weighted linear regression equation is obtained by fitting historical data points that have not been marked as expired data in the time series corresponding to the historical scene group. The batch code sequence number of the new data point is substituted into the independent variable of the weighted linear regression equation to calculate the corresponding dependent variable value as the predicted value. This step is used to calculate the expected level quantification score of the new data point based on the regression model fitted by the historical data before adding the new data point to the time series, for use in the subsequent calculation of the deviation between the actual score and the predicted score.

[0212] Here, "new data point" refers to the structured index data of a certain test scene group in a certain image quality dimension within the current batch of images to be processed. It is used as a newly added data record corresponding to that dimension of that batch and scene group in the time series. "New" is relative to the existing historical data points in the time series corresponding to historical scene groups. It represents the structured index data of that test scene group in the current batch of images to be processed in a certain image quality dimension. After this data point is added to the time series as the record corresponding to the latest timestamp, the number of data points in the time series increases by one.

[0213] Step S603: If the deviation between the actual quantification score and the predicted value of the new data point exceeds the preset detection threshold, the new data point is marked as a mutation point; wherein, the preset detection threshold is the product of the preset multiple and the residual standard deviation of all historical data points in the time series corresponding to the historical scene group that have not been marked as expired data.

[0214] In this embodiment, all historical data points that are not marked as expired data can be extracted from the time series corresponding to the historical scene group. The difference between the actual level quantification score of these historical data points and their corresponding predicted values ​​is calculated. Then, the standard deviation of these differences is calculated. The standard deviation is multiplied by a preset factor to obtain a preset detection threshold. The deviation between the actual level quantification score of the new data point and the predicted value is calculated. The deviation is compared with the preset detection threshold. When the deviation is greater than the preset detection threshold, the new data point is marked as a mutation point. This step is used to identify new data points that deviate from the normal fluctuation range of historical data points around the regression line. After the new data point is marked as a mutation point, it is used for manual review of the batch in subsequent processes.

[0215] In this embodiment, historical scene groups are determined by reusing the successfully matched scene correspondences to avoid repeated scene matching. Before adding new data points to the time series, the new data points are predicted using a regression equation fitted based on the unexpired data in the historical scene groups. This makes the prediction independent of the new data points and avoids the bias caused by self-prediction. The preset detection threshold is the product of a preset multiple and the standard deviation of the residuals of the unexpired historical data points. That is, the threshold is adaptively adjusted according to the degree of fluctuation of historical data. When the deviation between the actual score and the predicted value of the new data point exceeds the threshold, it is marked as a mutation point, thus realizing the identification and marking of abnormal deviations in quality of a single batch.

[0216] Figure 7 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 4 In the above Figure 2 Based on the illustrated embodiments, as Figure 7 As shown, the specific implementation method of step S201 is as follows:

[0217] Step S2011: Recognize the scene identifiers in the batch of images to be processed to obtain multiple test scene groups; wherein, the scene identifiers of the images in each test scene group are the same.

[0218] In this embodiment, for the acquired batch of images to be processed, each image in the batch can be traversed, the file name of each image can be read, the file name can be split according to the preset naming delimiter, and the key field representing the scene type can be extracted from the split field as the scene identifier of the image. Image file paths with the same scene identifier can be grouped into the same set, and each set can be used as a test scene group. After all images are processed, multiple test scene groups are obtained. This step is used to automatically classify the batch of images to be processed into their respective scene groups according to the scene information carried in their file names, so that images of the same scene under different shooting conditions are aggregated together, and scene fingerprints are generated for each test scene group to prepare structured image data.

[0219] Step S2012: Extract scene information from each test scene group, and construct the scene fingerprint corresponding to the test scene group based on the scene information.

[0220] The scenario information includes: file name pattern of test scenario group, serial number group identifier of test scenario group, scenario keywords, number of images in group, batch code and analysis timestamp.

[0221] In this embodiment, for each test scenario group, the filename of any image in the test scenario group can be read, and the filename can be split according to a preset naming delimiter. Key fields other than the sequence number and extension can be extracted as a set of scene keywords. At the same time, the common filename string pattern of each image filename in the test scenario group can be extracted as the filename pattern. The sequence number group identifier value can be extracted from the filename as the sequence number group identifier. The total number of images in the test scenario group can be obtained as the number of images in the group. The batch code and analysis timestamp of the batch of images can be recorded together. All the above fields are combined according to the preset fingerprint structure into a structured object containing the sequence number group identifier, filename pattern, scene keyword list, number of images in the group, batch code, and analysis timestamp, which serves as the scene fingerprint of the test scenario group. This step is used to extract various information describing the scene characteristics from each pre-divided test scenario group and assemble this information into a scene fingerprint with a unified structure, providing a standardized data format that can be directly compared for subsequent historical scene fingerprint matching.

[0222] Optionally, in addition to the file name pattern of the test scene group (file_pattern), the group number identifier of the test scene group (group_sequence), the scene keywords (scene_keywords), the number of images in the group (image_count), the batch code (batch_code), and the analysis timestamp (timestamp), the scene information also includes whether the test machine (has_test_machine) and the number of reference machines (ref_machine_count) are included.

[0223] Additionally, when scene information lacks valid scene keywords (e.g., all-numeric naming, random hash naming), a downgraded matching path is initiated: global visual features of the images are extracted (e.g., color histogram, edge orientation histogram, or lightweight image embedding vector), and visual similarity across batches of images is calculated for coarse scene matching. The matching results from the downgraded path are marked as low-confidence, and manual verification is recommended before performing incremental comparisons. This downgrade mechanism ensures that cross-batch scene pairing can still be completed even when filename information is insufficient.

[0224] In this embodiment, multiple test scene groups are obtained by identifying scene identifiers (images with the same scene identifier are grouped into the same test scene group). Scene information such as file name pattern, sequence number group identifier, scene keywords, number of images in the group, batch code, and analysis timestamp are extracted from each test scene group. Based on this information, a scene fingerprint corresponding to the test scene group is constructed. Automatic grouping is achieved through the above-mentioned scene identifier identification, and a structured scene fingerprint is automatically constructed based on multi-dimensional scene information. This ensures that each test scene group has a unique identifier containing rich attribute information, providing an accurate and complete structured data foundation for subsequent scene fingerprint matching. This avoids the manual identification of scene identifiers and manual organization of scene information, improving the automation level of the overall process.

[0225] Figure 8 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 5 In the above Figure 2 Based on the illustrated embodiments, as Figure 8 As shown, step S203 is specifically implemented as follows:

[0226] Step S2031: For the scene fingerprint of each test scene group, obtain the similarity of scene keywords, file name pattern, and sequence number between the scene fingerprint and each matching candidate historical scene fingerprint.

[0227] In this embodiment, for each test scenario group's scenario fingerprint, each candidate historical scenario fingerprint matching that test scenario group is sequentially traversed. For the currently traversed candidate historical scenario fingerprint, the intersection and union operations of the scenario keyword set in the current scenario fingerprint and the historical scenario keyword set in the candidate historical scenario fingerprint are performed. The similarity of the scenario keywords is obtained by dividing the intersection size by the union size. The matching degree of the file name strings is divided into three levels: complete match, partial match, and no match. If two file name strings are completely identical, the similarity value of the file name pattern is 1.0; if the main body of two file name strings is the same but there are minor differences (e.g., different version number suffixes or different prefix markers, such as...), the similarity value is 1.0. If indoor_portrait_().jpg and indoor_portrait_hdr_().jpg belong to different sub-variants of the same scene family, the value is 0.5; if the two filename strings are completely different in their main structure, the value is 0.0. The sequence group identifier in the scene fingerprint is compared with the sequence group identifier in the candidate historical scene fingerprint. If they are completely the same, the sequence similarity is 1; if they are different, the value is 0. This step is used to calculate the matching degree between the current test scene group and each candidate historical scene fingerprint in the three dimensions of scene keywords, filename pattern and sequence number, providing the basic values ​​of the three dimensions for the subsequent weighted comprehensive similarity calculation.

[0228] Step S2032: Based on the similarity of scene keywords, file name patterns, and sequence numbers between the scene fingerprint and each matched candidate historical scene fingerprint, and the preset weights, obtain the comprehensive similarity between the scene fingerprint and each matched candidate historical scene fingerprint.

[0229] In this embodiment, for each test scenario group's scenario fingerprint, each matching candidate historical scenario fingerprint in that test scenario group can be sequentially traversed. The similarity of the scenario keywords is multiplied by the scenario keyword weight coefficient to obtain a weighted value. The similarity of the file name pattern is multiplied by the file name pattern weight coefficient to obtain another weighted value. The similarity of the sequence number is multiplied by the sequence number weight coefficient to obtain yet another weighted value. These three weighted values ​​are added together to obtain the comprehensive similarity between the test scenario group and the current candidate historical scenario fingerprint. This step is used to merge the independent similarity values ​​of the three dimensions of scenario keywords, file name pattern, and sequence number into a single comprehensive similarity value according to their respective importance, so that this value can reflect the overall matching degree between the current test scenario group and the candidate historical scenario fingerprint.

[0230] For example, for each scene fingerprint, the following steps are performed to align it with historical scene fingerprints: Candidate fingerprints with the same number of images and a Jaccard similarity to scene keywords exceeding a preset threshold TH_CANDIDATE are selected from historical scene fingerprints. TH_CANDIDATE is a preset candidate selection threshold set for Jaccard similarity to scene keywords when selecting candidate fingerprints from historical scene fingerprints in cross-batch scene matching. A typical value for TH_CANDIDATE is 0.2-0.4. A comprehensive similarity score is then calculated for each candidate fingerprint.

[0231]

[0232] Here, *similarity* represents the overall similarity, and W1, W2, and W3 are preset weights that satisfy W1 + W2 + W3 = 1. Possible weights include W1 = 0.5 (highest weight for keyword matching), W2 = 0.3, and W3 = 0.2. *keyword_jaccard* represents the similarity of scene keywords, *file_pattern_match* represents the similarity of filename patterns, and *group_sequence_match* represents the similarity of sequence numbers. The automatic matching algorithm based on scene fingerprints (keyword Jaccard similarity + filename pattern + sequence number grouping) achieved a scene matching accuracy of 87.5% on the test dataset (10 batches, 120 scene groups in total), with a matching coverage rate of 92% or higher at medium confidence. Compared to manually annotating scene labels, the matching efficiency is improved by approximately 20 times.

[0233] A perfect match means the filename string in the current test scene fingerprint is exactly the same as the filename string in the candidate historical scene fingerprint. In this case, `file_pattern_match` is 1.0. A partial match means the two filename strings have the same main body but have minor differences, such as different version number suffixes or different prefix markers. In this case, `file_pattern_match` is 0.5. A mismatch means the two filename strings are completely different in their main structure. In this case, `file_pattern_match` is 0.0. Same sequence number means the sequence number group identifier in the current test scene fingerprint is exactly the same as the sequence number group identifier in the candidate historical scene fingerprint. This indicates that the test scene group in the two batches corresponds to the same sequence number position in the shooting order. In this case, `group_sequence_match` is 1.0. Different sequence numbers mean the two sequence number group identifiers have different values. In this case, `group_sequence_match` is 0.0.

[0234] Figure 9 This is an example diagram of the scene fingerprint extraction and scene alignment process provided in the embodiments of this application, such as... Figure 9 As shown, starting with a list of new batch files, images are first grouped according to file sequence number. Then, for each group, corresponding scene fingerprints are extracted based on keywords, file patterns, and image data. All historical batch fingerprint indexes are retrieved from the database for candidate screening. Screening criteria include matching the number of images and Jaccard similarity greater than a threshold. Subsequently, a comprehensive similarity is calculated by weighting the three dimensions of keywords, file patterns, and sequence number. The highest similarity value is compared with the matching threshold. If the value meets the threshold, the scene alignment is considered successful, and the matched image group and corresponding matching confidence are output. If the value does not reach the threshold, the current group is marked as unsuccessful in matching, and only a single batch of independent analysis is performed on this group. The entire process realizes the automated matching and division of new batch image groups and historical image groups of the same scene.

[0235] In this embodiment, for each test scenario group's scenario fingerprint, its similarity to each candidate historical scenario fingerprint is obtained in three dimensions: scenario keywords, file name pattern, and sequence number. A comprehensive similarity is calculated based on preset weights. Through the above multi-dimensional weighted similarity calculation method, the determination of comprehensive similarity no longer depends on the matching result of a single dimension, but integrates the similarity information of scenario keywords, file name pattern, and sequence number. This allows the comprehensive similarity to more comprehensively reflect the matching degree between the current test scenario group and the candidate historical scenario fingerprint, providing a more reliable quantitative basis for the selection of the highest comprehensive similarity and the determination of scenario labels, thereby improving the accuracy of scenario matching.

[0236] Figure 10 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 6 In the above Figure 2 Based on the illustrated embodiment, the method may further include:

[0237] Step S1001: Determine the confidence information of each scene fingerprint based on the highest comprehensive similarity of each scene fingerprint and the preset confidence level.

[0238] In this embodiment, for each test scenario group, the highest comprehensive similarity value obtained by scene matching for that test scenario group can be obtained. This value is compared with preset high confidence thresholds and low confidence thresholds. If the highest comprehensive similarity is greater than or equal to the high confidence threshold, the confidence information of the test scenario group is determined to be high confidence. If the highest comprehensive similarity is less than the high confidence threshold but greater than or equal to the low confidence threshold, it is determined to be medium confidence. If the highest comprehensive similarity is less than the low confidence threshold, it is determined to be low confidence. This step is used to convert the comprehensive similarity value corresponding to the optimal matching result of each test scenario group into hierarchical confidence level information to reflect the matching reliability between the test scenario group and the historical scene fingerprint of the optimal match.

[0239] For example, the candidate fingerprint with the highest overall similarity is selected. If the highest similarity is lower than the preset threshold TH_MATCH, the test scenario group in the current batch to be processed is marked as a failed match. The typical value range of TH_MATCH is 0.2-0.4. An overall similarity ≥ TH_HIGH_CONF is considered high confidence (safe for incremental comparison), TH_LOW_CONF to TH_HIGH_CONF is considered medium confidence (manual confirmation is recommended), and below TH_LOW_CONF is considered low confidence. TH_MATCH is the preset matching threshold for determining the final match based on the highest overall similarity in cross-batch scenario matching, and marking it as a failed match if it is lower. TH_HIGH_CONF is the high confidence threshold, and TH_LOW_CONF is the low confidence threshold. TH_HIGH_CONF can be set to 0.7 and TH_LOW_CONF to 0.3.

[0240] The specific implementation method of step S204 is as follows:

[0241] Step S1002: In response to the confidence information being high confidence, according to the preset level quantization mapping rules and based on the visual language model, obtain the structured indicators corresponding to the test scenario groups marked as successful matching and with high confidence information.

[0242] In this embodiment, after obtaining the confidence information of the test scene group, it can be checked whether the confidence information is a high confidence marker. If it is a high confidence marker, the image of the test scene group is input into the visual language model to obtain the natural language analysis results of the test scene group in each image quality dimension. Then, the natural language analysis results and the level quantization mapping rules are input into the language model, and the language model outputs index data that conforms to the structured format as the structured index corresponding to the test scene group. This step is used to automatically trigger the structured index extraction process for the matching high confidence scene group, so that the high confidence scene group can enter the subsequent processing stage without additional operation.

[0243] Step S1003: In response to the confidence information being of medium confidence and the confirmation signal for the test scenario group with medium confidence information, according to the preset level quantization mapping rules and based on the visual language model, obtain the structured indicators corresponding to the test scenario groups marked as successfully matched and with medium confidence information.

[0244] In this embodiment, after obtaining the confidence information of the test scene group, it can be checked whether the confidence information is of medium confidence. If it is of medium confidence, a prompt to wait for confirmation is output. When a confirmation signal for the test scene group is received, the image of the test scene group is input into the visual language model to obtain the natural language analysis results of the test scene group in each image quality dimension. Then, the natural language analysis results and the level quantization mapping rules are input into the language model, and the language model outputs index data that conforms to the structured format as the structured index corresponding to the test scene group. This step is used to start the structured index extraction process only after the medium confidence scene group is confirmed, so as to avoid the matching results with insufficient reliability from directly entering the subsequent processing stage.

[0245] Optionally, the confirmation signal for test scenario groups with medium confidence information can be obtained in the following ways: Manual confirmation: After a test scenario group is determined to be of medium confidence, its information is pushed to the manual confirmation interface. The operator views the matching details of the scenario fingerprint and, after confirming a correct match, clicks the confirmation button or submits a confirmation command. The generated confirmation signal serves as the basis for triggering the extraction of structured indicators for the medium confidence scenario group. Non-manual confirmation: After a test scenario group is determined to be of medium confidence, the system checks whether the test scenario group has been marked as a trusted pair in historical batches or whether it exists in a pre-configured whitelist of no-confirmation options. If the preset conditions are met, a confirmation signal is automatically generated for the test scenario group, serving as the basis for triggering the extraction of structured indicators for the medium confidence scenario group.

[0246] In this embodiment, by comparing the highest comprehensive similarity of each scene fingerprint with a pre-set confidence level, the scene matching results are divided into two levels: high confidence and medium confidence. Differentiated structured indicator acquisition paths are set for different confidence levels: for test scene groups with successful matching and high confidence, the structured indicator acquisition process is directly triggered without additional manual intervention; for test scene groups with successful matching but medium confidence, the structured indicator acquisition is not triggered immediately, but waits for a confirmation signal for the test scene group before acquisition can be performed, ensuring that the medium confidence scene group enters the subsequent analysis after verification and confirmation. The hierarchical processing mechanism enables the high confidence scene group to quickly enter the subsequent processing process, maintaining automation efficiency. At the same time, a confirmation step is set for the medium confidence scene group to avoid misleading deviations in subsequent comparison conclusions due to insufficient matching reliability, thus achieving a balance between automation efficiency and judgment reliability.

[0247] Figure 11 A flowchart illustrating the cross-batch image increment comparison method provided in this application embodiment. Figure 7 In the above Figure 2 Based on the illustrated embodiments, as Figure 11 As shown, step S203 is specifically implemented as follows:

[0248] Step S2033: If the highest comprehensive similarity is greater than or equal to the second preset threshold, then the test scenario group is marked as a successful match.

[0249] In this embodiment, for each test scenario group, the highest comprehensive similarity value calculated among all candidate historical scenario fingerprints for that test scenario group is obtained. This value is compared with a preset second threshold. If the value is greater than or equal to the second threshold, the test scenario group is marked as a successful match. This step is used to determine that test scenario groups with comprehensive similarity meeting the matching requirements are successfully paired with historical scenarios, so that the scenario is marked as a successful match in subsequent processing and structured index extraction and comparison operations are performed.

[0250] Step S2034, or, if the highest overall similarity is less than the second preset threshold, then the test scenario group is marked as a failed match.

[0251] In this embodiment, for each test scenario group, the highest comprehensive similarity value calculated among all candidate historical scenario fingerprints for that test scenario group can be obtained. This value is compared with a preset second threshold. If the value is less than the second preset threshold, the test scenario group is marked as a failed match. This step is used to determine that test scenario groups with comprehensive similarity lower than the matching requirement cannot be paired with historical scenarios, so that the scenario is marked as a failed match in subsequent processing, and subsequent structured index extraction and comparison operations are not performed.

[0252] The first preset threshold is used to filter candidate objects from historical scene fingerprints, and the second preset threshold is used to finally determine whether the test scene group is successfully matched. The two serve different judgment logics and can be set independently in practical applications. They can be set to the same value or different values.

[0253] In this embodiment, by limiting the determination of whether a match is successful to the comparison result of the highest comprehensive similarity and the second preset threshold, a successful match is marked when the result is greater than or equal to the second preset threshold, and a failed match is marked when the result is less than the second preset threshold. This means that the determination of whether a test scenario group is successfully paired with a historical scenario no longer relies on subjective human judgment, but is automatically determined based on the comparison result of the highest comprehensive similarity and the preset threshold. This provides a clear and unified decision-making basis for whether to perform structured index extraction and incremental comparison on the test scenario group in subsequent processes.

[0254] Based on the above embodiments, this application provides a specific implementation example, namely, a comparison system for cross-batch image increments, the system including an input layer, a scene matching layer, an incremental comparison layer, and a trend analysis layer.

[0255] The scene matching layer automatically identifies comparable test scene groups in different batches, providing a pairing basis for incremental comparison: it performs recognition processing on the batch images to be processed, obtaining multiple test scene groups in the batch images and the scene fingerprints corresponding to each test scene group; for the scene fingerprints corresponding to each test scene group, it filters candidate historical scene fingerprints from the historical scene fingerprint database, whose scene keyword similarity is greater than the first preset threshold and whose number of images in the group is the same; then, it calculates the comprehensive similarity based on the similarity between the scene fingerprint and the candidate historical scene fingerprint in three dimensions: scene keywords, file name pattern, and sequence number, as well as preset weights, and selects the highest comprehensive similarity to determine whether the test scene group is marked as a successful match or an unsuccessful match, thus completing the automatic pairing of test scene groups between different batches.

[0256] The incremental comparison layer performs VLM analysis only on new batches, reusing structured metrics from historical batches for incremental comparison, thus reducing token costs from O(N) to O(N). 2To reduce the complexity to O(N): For test scene groups marked as successfully matched, a structured index is extracted based on a visual language model and a pre-defined level quantization mapping rule. This structured index includes scene identifiers, batch codes, and a set of image quality dimension indicators. The image quality dimensions include exposure, white balance, noise, texture, focus, and overall performance. Each dimension's index set includes defect level, level quantization score, defect description summary, and judgment confidence. Then, the level quantization scores of each dimension in the structured index of the test scene group are subtracted dimension by dimension from the level quantization scores of the corresponding dimensions in the historical structured index of the matched historical scene groups in the database. The score difference for each image quality dimension is obtained and mapped to the corresponding change direction label, serving as the incremental comparison result for the test scene group.

[0257] The trend analysis layer detects long-term quality trends and generates structured trend reports based on incremental comparison results from multiple batches. It acquires structured indicators and incremental comparison results for the same test scenario group across multiple batches. Using scene mode and image quality dimensions as indices, it constructs a time-series sequence by arranging batch codes, analysis timestamps, defect levels, level quantification scores, and sample sizes for each batch in chronological order. The least squares method is used to perform linear regression fitting on the level quantification scores of each data point in the time-series sequence to obtain slope values ​​and calculate the coefficient of determination. Based on the slope and coefficient of determination, the long-term trend direction of each image quality dimension is determined. The long-term trend direction of each dimension is combined with the change direction labels in the incremental comparison results of the most recent batch to obtain the comprehensive trend results for the test scenario group across multiple quality dimensions. Trend report generation uses a plain text API, consuming only approximately 3000-6000 tokens per report, resulting in low cost.

[0258] Figure 12 A three-tier architecture example diagram for cross-batch incremental comparison and trend analysis provided in this application embodiment is shown below. Figure 12As shown, the cross-batch incremental comparison and trend analysis architecture is divided into four processing units from top to bottom: input layer, scene matching layer, incremental comparison layer, and trend analysis layer. The input layer imports stored multi-version historical batch indexes and new batch image data of the new firmware to be verified. The scene matching layer completes grouping based on file name and sequence number and extracts scene fingerprints. It completes the same scene alignment matching between the new batch and the historical batch through fingerprint similarity calculation. The incremental comparison layer only performs VLM image parsing on the new batch and extracts standardized multi-dimensional structured indicators. It reuses the historical indicators stored in the database to carry out incremental comparison of new and old data scene by scene, completes the determination of image quality change type and confidence assignment. The trend analysis layer aggregates the multi-batch structured indicators by scene, dimension and time dimension to construct a time series. It completes long-term image quality trend detection with the help of linear regression and significance test. It generates trend interpretation content in natural language form based on LLM. Finally, it outputs a complete cross-batch trend report with image quality change alarm. The whole layered process realizes the fully automated processing of automatic image pairing, incremental quantitative comparison and long-term image quality evolution law analysis in firmware iteration scenarios.

[0259] Figure 13 An example diagram illustrating the extended cross-batch data model provided in this application embodiment, such as... Figure 13 As shown, the model consists of seven types of data objects: AnalysisBatch, AnalysisRecord, StructuredMetrics, BatchComparison, SceneAlignment, DimensionTrend, and Trend report. AnalysisBatch carries batch identifiers, timestamps, and firmware version information, and contains multiple AnalysisRecords. Each AnalysisRecord is bound to the scene, batch information, and analysis results, and is associated with and stores the quantitative scores of each dimension in StructuredMetrics. Multiple AnalysisBatch groups participate in the comparison to generate BatchComparison objects. BatchComparison contains SceneAlignment, which records scene matching information and matching confidence levels. SceneAlignment further encapsulates and stores DimensionTrend data on dimensional change magnitude, slope, and significance. Finally, based on BatchComparison, a Trend report with global confidence, trend summary, and alarm status is generated. The entire data model fully carries the structured data produced by the entire process of scene matching, incremental comparison, and time-series trend calculation, realizing standardized storage and associated traceability of cross-batch image comparison data.

[0260] Figure 14 This is a schematic diagram of the structure of the comparison device for cross-batch image increments provided in an embodiment of this application. Figure 14 As shown, the comparison device for cross-batch image increments includes:

[0261] The recognition module 1401 is used to perform recognition processing on the acquired batch images to be processed, and to acquire multiple test scene groups in the batch images to be processed and the scene fingerprint corresponding to each test scene group. The fingerprint acquisition module 1402 is used to acquire matching candidate historical scene fingerprints from historical scene fingerprints in historical batches for the scene fingerprint corresponding to each test scene group; wherein the similarity between the historical scene keywords in the candidate historical scene fingerprint and the scene keywords in the scene fingerprint is greater than a first preset threshold, and the number of images within a group in the historical scene fingerprint is the same as the number of images within a group in the scene fingerprint. The determination module 1403 is used to acquire a comprehensive similarity for the scene fingerprint corresponding to each test scene group, based on the scene fingerprint and the matching candidate historical scene fingerprints, and to acquire the highest comprehensive similarity from the comprehensive similarity, and to determine whether the label of the test scene group is a successful match based on the highest comprehensive similarity. The indicator acquisition module 1404 is used to acquire the structured indicators corresponding to each test scene group in the batch of images to be processed, according to a preset level quantization mapping rule and based on a visual language model. The structured indicators include: the scene identifier of the test scene group, the batch code, and a set of indicators for the image quality dimension within the test scene group. The comparison module 1405 is used to compare the structured indicators corresponding to the test scene groups marked as successfully matched with the structured indicators of historical scene groups that have successfully matched the test scene groups, thereby obtaining the incremental comparison results of the batch of images to be processed.

[0262] The cross-batch image increment comparison device provided in this embodiment can perform... Figure 2 , Figure 4 , Figure 6 , Figure 7 , Figure 8 , Figure 10 and Figure 11 The technical solution of the cross-batch image increment comparison method embodiment shown is implemented in accordance with the principle and technical effect of the method. Figure 2 , Figure 4 , Figure 6 , Figure 7 , Figure 8 , Figure 10 and Figure 11 The embodiment of the cross-batch image increment comparison method shown is similar and will not be described in detail here.

[0263] Figure 15 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application. Figure 15As shown, the electronic device 150 includes at least one processor 1501 and a memory 1502. The electronic device 150 also includes a communication component 1503. The processor 1501, the memory 1502, and the communication component 1503 are connected via a bus 1504.

[0264] In the specific implementation process, at least one processor 1501 executes computer execution instructions stored in memory 1502, so that at least one processor 1501 is used to implement a cross-batch image increment comparison method of the above embodiment.

[0265] The specific implementation process of processor 1501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0266] In the above embodiments, it should be understood that the processor 1501 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0267] The memory 1502 may include high-speed RAM memory, and may also include non-volatile memory NVM, such as at least one disk storage.

[0268] Bus 1504 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 1504 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 1504 in the accompanying drawings of this application is not limited to only one bus or one type of bus.

[0269] The above description of the functions implemented by electronic devices and main control devices has introduced the solutions provided by the embodiments of the present invention. It is understood that, in order to implement the above functions, the electronic device or main control device includes hardware structures and / or software modules corresponding to the execution of each function. By combining the units and algorithm steps of the various examples described in the embodiments of the present invention, the embodiments of the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the technical solutions of the embodiments of the present invention.

[0270] This application also provides a computer-readable storage medium storing computer-executable instructions. When executed by a processor, these instructions implement a cross-batch image incremental comparison method as described above. In the specific implementation of the aforementioned cross-batch image incremental comparison method, each module can be implemented as a processor.

[0271] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0272] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, the processor and the readable storage medium can exist as discrete components in an electronic device or a host device.

[0273] This application also provides a computer program product, including a computer program, which, when executed by a processor, is used to implement a cross-batch image increment comparison method according to the above embodiments.

[0274] The computer program is stored in a readable storage medium, and at least one processor can read the computer program from the readable storage medium and execute the computer program to perform the scheme provided in any of the above embodiments.

[0275] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disk, or optical disk.

[0276] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for comparing incremental images across batches, characterized in that, include: The acquired batch images to be processed are identified and processed to obtain multiple test scene groups in the batch images to be processed and scene fingerprints corresponding to each test scene group; For each test scenario group, a matching candidate historical scenario fingerprint is obtained from the historical scenario fingerprints in the historical batch; wherein, the similarity between the historical scenario keywords in the candidate historical scenario fingerprint and the scenario keywords in the scenario fingerprint is greater than a first preset threshold, and the number of images within the group in the historical scenario fingerprint is the same as the number of images within the group in the scenario fingerprint. For each test scenario group, based on the scenario fingerprint and the matched candidate historical scenario fingerprints, a comprehensive similarity is obtained, and the highest comprehensive similarity is obtained from the comprehensive similarity. Based on the highest comprehensive similarity, it is determined whether the test scenario group is successfully matched. For each test scene group in the batch of images to be processed, according to the preset level quantization mapping rules and based on the visual language model, the structured indicators corresponding to the test scene group are obtained; wherein, the structured indicators include: the scene identifier, batch code, and the set of indicators for the image quality dimension in the test scene group. For test scene groups marked as successfully matched, the structured indicators corresponding to the test scene groups are compared with the structured indicators of historical scene groups that were successfully matched by the test scene groups to obtain the incremental comparison results of the batch images to be processed.

2. The method according to claim 1, characterized in that, The image quality dimensions include: exposure, white balance, noise, texture, focus and / or a combination thereof, and the set of indicators for the image quality dimensions includes: defect level, level quantification score, defect description summary and judgment confidence. The step of comparing the structured metrics corresponding to the test scenario group with the structured metrics of historical scenario groups that were successfully matched with the test scenario group to obtain the incremental comparison results of the batch images to be processed includes: Based on the image quality dimension index set in the structured indicators corresponding to the test scenario group, the image quality dimension index set in the structured indicators of the historical scenario group that successfully matched the test scenario group is compared dimension by dimension to obtain the change in each image quality dimension; wherein, the structured indicators of the historical scenario group are obtained by reading data from a preset storage terminal.

3. The method according to claim 2, characterized in that, Also includes: Obtain incremental comparison results of multiple batches of images, and structured metrics for each batch of images; A time series is constructed based on the structured indicators under each batch of images; wherein, the time series includes: scene mode, image quality dimension and data point list, and the data point list includes: batch code, analysis timestamp, defect level, level quantification score and number of samples; Based on the time series, linear regression and significance tests are used to obtain the trend direction, linear fit degree and significance probability of each image quality dimension; A comprehensive trend result is obtained based on the incremental comparison results of the multiple batches of images, the trend direction of each image quality dimension, the linear fitting degree, and the significance probability.

4. The method according to claim 3, characterized in that, The step of obtaining the trend direction, linear fit degree, and significance probability of each image quality dimension based on the time series using linear regression and significance testing includes: Based on the time span between the analysis timestamp corresponding to each historical data point in the time series and the analysis timestamp of the batch images to be processed, the time decay weight of each historical data point is determined; wherein, the time decay weight decreases monotonically as the time span increases. Historical data points whose time decay weight is less than a preset expiration threshold are marked as expired data. Using the time decay weight as a weighting coefficient, a weighted linear regression is performed on the rank quantification scores of each historical data point in the time series that is not marked as expired data to obtain the weighted slope and the linear fit degree. Based on the weighted slope and the preset slope threshold, the trend direction of each image quality dimension is obtained; When the number of historical data points in the time series that are not marked as expired data is greater than or equal to a preset number, a significance test is performed based on the weighted slope to obtain the significance probability.

5. The method according to claim 4, characterized in that, The comprehensive trend result includes a comprehensive trend score and a trend insight report. The comprehensive trend result is obtained based on the incremental comparison results of the multiple batches of images, the trend direction of each image quality dimension, the linear fit degree, and the significance probability. Based on the trend direction of each image quality dimension, determine the trend score corresponding to each image quality dimension; The overall trend score is obtained by multiplying the trend score corresponding to each image quality dimension by its respective preset dimension weight and then summing the results. Based on the comparison between the significance probability and the preset significance threshold, the significance level corresponding to each image quality dimension is determined. The trend direction of each image quality dimension is classified to obtain a list of improvement dimensions, a list of regression dimensions, and a list of stable dimensions; Based on the incremental comparison results of the multiple batches of images, as well as the trend direction, linear fitting degree and significance probability of each image quality dimension, regression alarm flags and regression severity levels are generated. The trend score corresponding to each image quality dimension, the overall trend score, the significance level corresponding to each image quality dimension, the list of improvement dimensions, the list of regression dimensions, the list of stable dimensions, the regression alarm flag, and the regression severity level are all input into the large language model to obtain the trend insight report.

6. The method according to claim 4, characterized in that, Also includes: When the structured metrics corresponding to the test scenario group are obtained, the corresponding historical scenario group is determined based on the correspondence established by the successful matching of the test scenario group with the historical scenario group. Before adding the structured indicators of the test scenario group as new data points to the time series corresponding to the historical scenario group, the new data points are predicted using a weighted linear regression equation fitted based on historical data points in the historical scenario group that are not marked as expired data. If the deviation between the actual quantized score of the new data point and the predicted value exceeds a preset detection threshold, the new data point is marked as a mutation point; wherein, the preset detection threshold is the product of a preset multiple and the residual standard deviation of all historical data points in the time series corresponding to the historical scene group that have not been marked as expired data.

7. The method according to claim 1, characterized in that, The step of performing recognition processing on the acquired batch images to be processed, and obtaining multiple test scene groups in the batch images to be processed and the scene fingerprint corresponding to each test scene group, includes: The scene identifiers in the batch of images to be processed are identified to obtain multiple test scene groups; wherein the scene identifiers of the images in each test scene group are the same; Extract scene information from each test scene group, and construct a scene fingerprint corresponding to the test scene group based on the scene information; The scenario information includes: the file name pattern of the test scenario group, the serial number group identifier of the test scenario group, scenario keywords, the number of images in the group, batch code, and analysis timestamp.

8. The method according to claim 7, characterized in that, The step of obtaining a comprehensive similarity score for each test scenario group based on its corresponding scenario fingerprint and the matched candidate historical scenario fingerprints includes: For each test scenario group, the similarity of scenario keywords, file name patterns, and sequence numbers between the scenario fingerprint and each matching candidate historical scenario fingerprint is obtained. Based on the similarity of scene keywords, file name patterns, and sequence numbers between the scene fingerprint and each matching candidate historical scene fingerprint, and a preset weight, the comprehensive similarity between the scene fingerprint and each matching candidate historical scene fingerprint is obtained.

9. The method according to claim 1, characterized in that, Also includes: The confidence information of each scene fingerprint is determined based on the highest comprehensive similarity of each scene fingerprint and the preset confidence level. The step of obtaining the structured indicators corresponding to the test scenario group according to the preset level quantization mapping rules and based on the visual language model includes: In response to the high confidence level information, according to the preset level quantization mapping rules and based on the visual language model, the structured indicators corresponding to the test scenario groups marked as successful matching and with high confidence level information are obtained respectively. In response to the confidence information being of medium confidence and the confirmation signal for the test scenario group with medium confidence information, the structured indicators corresponding to the test scenario groups marked as successfully matched and with medium confidence information are obtained according to the preset level quantization mapping rules and based on the visual language model.

10. The method according to claim 1, characterized in that, The step of determining whether the marker of the test scene group is a successful match based on the highest overall similarity includes: If the highest overall similarity is greater than or equal to the second preset threshold, then the test scenario group is marked as a successful match. or, If the highest overall similarity is less than the second preset threshold, then the test scenario group is marked as a failed match.

11. A comparison device for cross-batch image increments, characterized in that, include: The recognition module is used to perform recognition processing on the acquired batch images to be processed, and to obtain multiple test scene groups in the batch images to be processed and the scene fingerprint corresponding to each test scene group; The fingerprint acquisition module is used to acquire matching candidate historical scene fingerprints from historical scene fingerprints in historical batches for each scene fingerprint corresponding to the test scene group; wherein, the similarity between the historical scene keywords in the candidate historical scene fingerprint and the scene keywords in the scene fingerprint is greater than a first preset threshold, and the number of images within the group in the historical scene fingerprint is the same as the number of images within the group in the scene fingerprint. The determination module is used to obtain a comprehensive similarity for each scene fingerprint corresponding to each test scene group, based on the scene fingerprint and the matched candidate historical scene fingerprints, and to obtain the highest comprehensive similarity from the comprehensive similarity, and to determine whether the tag of the test scene group is a successful match based on the highest comprehensive similarity; The indicator acquisition module is used to acquire the structured indicators corresponding to each test scene group in the batch of images to be processed, according to a preset level quantization mapping rule and based on a visual language model; wherein, the structured indicators include: the scene identifier, batch code, and a set of indicators for the image quality dimension in the test scene group. The comparison module is used to compare the structured indicators corresponding to the test scene group marked as a successful match with the structured indicators of the historical scene group that was successfully matched by the test scene group, and obtain the incremental comparison results of the batch images to be processed.

12. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; When the processor executes the computer execution instructions stored in the memory, it is used to implement the cross-batch image increment comparison method as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the cross-batch image increment comparison method as described in any one of claims 1 to 10.

14. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, is used to implement the cross-batch image increment comparison method as described in any one of claims 1 to 10.

Images