Transparent film abnormality rephotographing and termination control method and device based on information gain

Abstract

The embodiment of the application provides a transparent film abnormality rephotographing and termination control method and device based on information gain, which comprises the following steps: collecting an initial image set and an initial analysis result, and constructing a current decision state vector containing classification, horizon, artifact and evidence gap information; generating a candidate rephotographing action set based on the state vector, and calculating the expected information gain of each action for the state vector, which comprehensively considers the marginal improvement of classification confidence, horizon inference, artifact exclusion and interpretation integrity, and deducts the acquisition cost and time delay; under resource constraints, selecting the action with the maximum information gain to execute and updating the state vector, and ending the acquisition process when the updated state vector meets the preset termination acquisition condition, which can improve the expected income of rephotographing acquisition actions and increase the visual detection efficiency.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for abnormal reshooting and termination control of transparent films based on information gain. Background Technology

[0002] In industrial visual inspection of transparent films and complex reflective materials, traditional methods typically acquire all focal layers, angles, and exposure combinations in one go according to preset parameters. While this approach is simple to operate, it has two prominent problems: first, the acquisition redundancy is large, leading to a longer inspection cycle, a surge in data volume, and a heavier computational burden; second, although there are a large number of images, the information most valuable for anomaly detection may not be acquired first and is difficult to identify in the early stages.

[0003] Existing re-shooting control technologies mostly rely on a single threshold or a fixed template, such as adding a focal layer when sharpness is insufficient or increasing exposure groups when exposure is abnormal. However, for complex anomalies such as transparent films, foreign objects under films, surface contamination, and weakly reflective particles, a single threshold often cannot effectively determine whether "new acquisition actions truly improve judgment capabilities." In practical engineering, the system needs to simultaneously consider multiple acquisition targets. However, the lack of a unified information gain metric during the acquisition process makes it difficult for the system to make stable decisions on whether to re-shoot, which parameters to re-shoot, and when to terminate.

[0004] Therefore, there is an urgent need for a transparent film abnormality re-shooting and termination control method based on information gain, which can improve the expected benefits of re-shooting acquisition and increase visual detection efficiency. Summary of the Invention

[0005] To address the problems in the prior art, this application provides a method and apparatus for abnormal re-shooting and termination control of transparent films based on information gain, which can improve the expected benefits of re-shooting acquisition actions and increase visual detection efficiency.

[0006] To solve at least one of the above problems, this application provides the following technical solution: In a first aspect, this application provides a method for abnormal re-shooting and termination control of transparent films based on information gain, comprising: Acquire initial acquisition results for transparent membrane anomaly detection, the initial acquisition results including an initial multi-parameter image set and initial image analysis results; A judgment state vector is constructed based on the initial acquisition results. The judgment state vector is used to characterize the comprehensive judgment state of the transparent membrane abnormality. A corresponding set of candidate reshoot actions is determined based on the judgment state vector. Marginal benefit calculation is performed on each candidate reshoot action in the set of candidate reshoot actions to determine the corresponding expected information gain. Under the condition of satisfying the preset resource constraints, the target reshoot is performed based on the candidate reshoot action with the largest expected information gain to determine the corresponding new image data. The newly added image data and the initial acquisition results are fused together, and the determination state vector is updated according to the fused acquisition results. The corresponding comprehensive uncertainty is calculated based on the updated determination state vector, and it is determined whether the termination reshoot condition is met. If it is met, the acquisition process ends, and the updated determination state vector is output to determine the abnormal state of the transparent film.

[0007] Further, the step of constructing a determination state vector based on the initial acquisition results includes: The anomaly classification confidence, layer probability confidence, and artifact discrimination confidence are extracted from the initial image analysis results, and the corresponding anomaly classification results, layer results, and artifact discrimination results are determined respectively. The evidence coverage integrity of the current image analysis results is evaluated based on the initial multi-parameter image set to identify the corresponding key evidence gaps. Based on the data collection actions that have been performed, determine the corresponding parameter coverage, cumulative data collection cost, and cumulative time overhead. The anomaly classification results, the layer results, the artifact determination results, the key evidence gaps, the parameter coverage, the cumulative acquisition cost, and the cumulative time overhead are stored in a preset hierarchical data structure to determine the corresponding determination state vector. The determination state vector is used to characterize the comprehensive determination state of transparent membrane anomalies.

[0008] Further, determining the corresponding candidate retake action set based on the determination state vector includes: Based on the key evidence gaps and parameter coverage in the judgment state vector, the corresponding collection parameter dimension is determined. The collection parameter dimension represents the parameter dimension that has insufficient sampling confidence or is not currently covered. Based on the acquisition parameter dimensions, corresponding executable actions are enumerated and matched from the preset candidate reshoot action type library to determine the corresponding candidate reshoot action set. The action types in the candidate reshoot action type library include additional focal layer sampling, expanding the focal layer range, reducing the focal layer step size, adding observation angle, switching bands, adding exposure groups, switching illumination modes, performing polarization acquisition, performing high-magnification resampling on local regions of interest, and reselecting reference pose.

[0009] Further, the step of calculating the marginal benefit of each candidate retake action in the candidate retake action set to determine the corresponding expected information gain includes: For each candidate reshoot action in the candidate reshoot action set, the marginal benefit component of the reshoot action to the determination state vector is calculated according to the preset data processing method. The marginal benefit component includes marginal classification gain, marginal layer gain, marginal artifact removal gain, marginal interpretation gain, reshoot action acquisition cost, and time delay. The weighted sum of each marginal benefit item is used to determine the corresponding expected information gain.

[0010] Furthermore, the data processing method for calculating the marginal revenue components includes: Rule calculation involves matching the candidate reshoot action and the judgment state vector according to a preset rule base to determine the corresponding marginal benefit item. The lightweight visual model is used to perform regression prediction on the potential benefits after the candidate reshoot action is executed, based on the feature vector of the acquired image, to determine the corresponding marginal benefit items. Historical experience database matching calculation: retrieve the historical sample with the highest similarity to the current judgment state vector, and calculate its actual marginal benefit as the marginal benefit component of the current candidate retake action; Semantic intelligent agent computing, based on large language models or knowledge graphs, infers the causal relationship between the current anomaly type and candidate reshoot actions, and determines the corresponding marginal benefit items.

[0011] Furthermore, the step of calculating the corresponding comprehensive uncertainty based on the updated decision state vector includes: Based on the updated judgment state vector, extract the corresponding anomaly classification uncertainty, layer uncertainty, artifact attribution uncertainty, and evidence insufficiency. The uncertainty of anomaly classification, the uncertainty of stratum, the uncertainty of artifact attribution, and the uncertainty of insufficient evidence interpretation are weighted and summed to determine the corresponding comprehensive uncertainty.

[0012] Furthermore, the conditions for terminating reshooting include at least one of the following: The overall uncertainty is lower than a preset uncertainty threshold; The expected information gain of all candidate reshoot actions in the candidate reshoot action set is lower than a preset gain threshold. The cumulative collection cost or cumulative time expenditure has reached the preset resource constraints.

[0013] Secondly, this application provides a transparent film abnormal reshoot and termination control device based on information gain, comprising: The initial acquisition result determination module is used to obtain the initial acquisition results of the transparent membrane anomaly detection. The initial acquisition results include an initial multi-parameter image set and initial image analysis results. The reshooting action determination module is used to construct a judgment state vector based on the initial acquisition results. The judgment state vector is used to characterize the comprehensive judgment state of the transparent membrane abnormality. Based on the judgment state vector, a corresponding set of candidate reshooting actions is determined. Marginal benefit calculation is performed on each candidate reshooting action in the set of candidate reshooting actions to determine the corresponding expected information gain. Under the condition of satisfying the preset resource constraints, the target reshooting is performed based on the candidate reshooting action with the largest expected information gain to determine the corresponding new image data. The reshoot termination control module is used to fuse the newly added image data and the initial acquisition result, update the determination state vector according to the fused acquisition result, calculate the corresponding comprehensive uncertainty according to the updated determination state vector, and determine whether the reshoot termination condition is met. If it is met, the acquisition process ends, and the updated determination state vector is output to determine the abnormal state of the transparent film.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the information gain-based transparent film abnormal reshoot and termination control method.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the information gain-based transparent film abnormal reshoot and termination control method.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the information gain-based transparent film abnormal reshoot and termination control method.

[0017] As can be seen from the above technical solution, this application provides a method and apparatus for abnormal re-shooting and termination control of transparent membranes based on information gain. It collects an initial image set and initial analysis results to construct a current judgment state vector containing classification, layer, artifact, and evidence gap information. Based on this state vector, a set of candidate re-shooting actions is generated, and the expected information gain of each action on the state vector is calculated. This gain integrates the marginal improvement of classification confidence, layer inference, artifact elimination, and interpretation completeness, while deducting acquisition cost and time delay. Under resource constraints, the action with the largest information gain is selected for execution and the state vector is updated. When the updated state vector meets the preset termination acquisition conditions, the acquisition process ends. This improves the expected benefits of re-shooting acquisition actions and increases visual detection efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the information gain-based transparent film abnormal reshoot and termination control method in the embodiments of this application. Figure 2 This is a structural diagram of the information gain-based transparent film abnormal reshoot and termination control device in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0022] Existing visual inspection and re-enhancing technologies for transparent films often rely on single thresholds or fixed templates, making it difficult to simultaneously address the multi-dimensional acquisition targets required in practical engineering, resulting in low re-enhancing control efficiency. This application provides a method and apparatus for abnormal re-enhancing and termination control of transparent films based on information gain. The method involves acquiring an initial image set and initial analysis results to construct a current judgment state vector containing classification, layer, artifact, and evidence gap information. Based on this state vector, a set of candidate re-enhancing actions is generated, and the expected information gain of each action on the state vector is calculated. This gain integrates the marginal improvement in classification confidence, layer inference, artifact elimination, and interpretation completeness, while deducting acquisition costs and time delays. Under resource constraints, the action with the highest information gain is selected for execution, updating the state vector. The acquisition process ends when the updated state vector meets preset termination conditions. This improves the expected benefits of re-enhancing actions and increases visual inspection efficiency.

[0023] To improve the expected benefits of re-capture actions and increase visual detection efficiency, this application provides an embodiment of a transparent film abnormality re-capture and termination control method based on information gain. See [link to embodiment]. Figure 1 The information gain-based transparent film abnormal reshoot and termination control method specifically includes the following: Step S101: Obtain the initial acquisition results of transparent membrane anomaly detection, the initial acquisition results including an initial multi-parameter image set and initial image analysis results; Optionally, in this embodiment, the focus is on anomaly detection scenarios involving transparent films, functional thin films, composite optical layers and their surfaces, interiors, and subsurfaces. The emphasis is on how to determine, based on multiple anomaly detection images already captured, "whether further imaging is needed," "what the most valuable image to capture next," and "when to stop acquisition and proceed to subsequent analysis" during ultra-depth-of-field, multi-focal-layer, multi-angle, and multi-band imaging processes. This aims to balance acquisition pace, evidence quality, and computational overhead within a unified framework.

[0024] The logic behind this embodiment is that useful evidence of transparent membrane anomalies is not evenly distributed, and the benefits of subsequent data collection actions do not necessarily increase linearly with the number of samples. By incorporating both benefits and costs into a unified function, the re-collection behavior can be transformed from an empirical behavior into a measurable, interpretable, and auditable optimization decision-making behavior.

[0025] Optionally, this step is the starting point of the transparent membrane abnormality re-shooting and termination control method, which obtains raw data and preliminary analysis conclusions about the current abnormal transparent membrane target to be detected from the physical acquisition device.

[0026] Specifically, the pre-configured imaging acquisition module, including a multi-focal-layer scanning unit, a multi-angle illumination unit, and a multi-band sensing unit, is invoked. Based on initial acquisition parameters (such as the default focal layer range, the basic observation angle, and the standard visible light band), a limited but key-dimensional acquisition operation is performed on the transparent film sample to obtain an initial set of multi-parameter images. This acquisition is not exhaustive but rather a rapid acquisition of a set of "probing" images that reflect the basic features of anomalies at a low cost.

[0027] After data acquisition, the initial analysis process is triggered. This process performs rapid preprocessing on the initial multi-parameter image set, including image alignment, noise filtering, and preliminary location of regions of interest (ROIs). Subsequently, a pre-defined decision model or rule engine is used to output the initial image analysis results.

[0028] The initial image analysis results contain at least three levels of information: First, a preliminary classification probability for the anomaly category of the current ROI (e.g., suspected sub-membrane foreign object, surface contamination, or material defect); second, a preliminary inference for the spatial layer where the anomaly is located (on the membrane, inside the membrane, or under the membrane); and third, an initial judgment on whether there is a risk of artifacts in the existing evidence.

[0029] This step provides a quantifiable starting point for subsequent information gain calculations and reshoot decisions, namely, "what is currently known and what is uncertain".

[0030] Step S102: Construct a judgment state vector based on the initial acquisition results. The judgment state vector is used to characterize the comprehensive judgment state of the transparent membrane abnormality. Determine the corresponding candidate reshoot action set based on the judgment state vector. Calculate the marginal benefit of each candidate reshoot action in the candidate reshoot action set to determine the corresponding expected information gain. Under the condition of satisfying the preset resource constraints, perform target reshoot based on the candidate reshoot action with the largest expected information gain to determine the corresponding new image data. Optionally, in this embodiment, this step is a transparent film abnormal reshoot control step based on information gain. By structurally modeling the current judgment state, possible reshoot actions are automatically generated, and the marginal benefits that each action may bring are quantitatively predicted to obtain information gain. Finally, under resource constraints, the optimal reshoot action is selected for execution to obtain new image data.

[0031] The term "information gain" does not simply refer to an increase in the number of image pixels, nor does it merely refer to improved sharpness. Rather, it refers to the increase in marginally effective evidence brought about by a new acquisition action to the final detection decision. Any action that can significantly improve anomaly classification, layer inference, artifact removal, human interpretation, or process attribution can be considered to have positive information gain.

[0032] The term "supplementary sampling" refers to additional data collection actions performed relative to the currently completed data set to address insufficient evidence, excessive uncertainty, or missing key dimensions. Supplementary sampling can involve adding new parameter dimensions or increasing the sampling density on existing dimensions.

[0033] Specifically, after acquiring the initial images of the transparent film anomaly target (e.g., images under several focal layers, angles, or exposure conditions), the system first needs to transform these raw image data into a structured, computable state representation, i.e., a judgment state vector. This vector is not a simple stacking of image features, but rather an abstract summary of the overall anomaly judgment. The determination state vector should include information in at least the following dimensions: the current preliminary determination result of the anomaly category and its confidence level (e.g., 70% confidence level for "sub-membrane foreign body" and 20% confidence level for "on-membrane contamination"); the current probability distribution of the inferred layer (on-membrane, intra-membrane, or sub-membrane); the current estimation of whether artifacts (such as specular reflection, out-of-focus, halo, etc.) exist in the image and the degree of interference; the spatial coverage of the acquired parameters (e.g., which focal layers have been captured and which angles are missing); and the cumulative acquisition time and computational cost already consumed.

[0034] Specifically, given the current state vector, the system can automatically generate a series of candidate actions for "what to do next". These actions are not fixed combinations of all parameters, but are dynamically generated based on the evidence gaps in the current state.

[0035] For example, if the state vector indicates "the focal layer covers the upper surface and intermediate layers, but the bottom layer is missing," candidate actions could include "adding bottom focal layer sampling" or "expanding the focal layer search range." If it indicates "low classification confidence and suspected artifact interference," candidate actions could include "switching to polarization acquisition" or "adding a short exposure group." Common candidate actions also include: adding observation angles, switching near-infrared bands, adding high-magnification local resampling, and reducing the focal layer step size.

[0036] Specifically, for each candidate re-enhancing action, it is necessary to predict: if the action is performed and a new image is obtained, how much "marginal effective evidence enhancement" it will bring to the final anomaly determination relative to the current state. This enhancement is quantified as "expected information gain." This is also the core judgment criterion of this embodiment.

[0037] Expected information gain is not a single indicator, but a weighted combination of multiple dimensions.

[0038] Specifically, this includes: an expected improvement in the confidence level of anomaly classification (whether it can be more definitively identified as a foreign object, scratch, or bubble); an expected improvement in the stability of layer inference (whether it can be more definitively identified as an anomaly on or beneath the membrane); an expected improvement in the ability to eliminate artifacts (whether it can distinguish whether the currently observed anomaly is a real defect or a reflected shadow); and an expected improvement in the completeness of result interpretation (whether a more complete and manually verifiable chain of evidence can be formed). Simultaneously, this gain will deduct the acquisition costs (such as additional photo captures) and time delays (such as the time spent switching light sources or turntables) incurred by this action.

[0039] By calculating the expected information gain, the previously incomparable "adding a focal layer" and "adding a polarization acquisition" can be compared on a unified "net information gain" scale. For example, if the current classification is clear but the layer position is uncertain, the ΔL (layer position improvement) of polarization acquisition may be much greater than the ΔL of adding a focal layer, thus obtaining a higher overall gain.

[0040] Specifically, after calculating the expected information gain for each candidate action, the first step is to check whether resource constraints are met. These constraints include: whether the remaining acquisition time budget is sufficient, whether the number of hardware actions has reached its limit, and whether the current cumulative cost is approaching a threshold.

[0041] An action is only included in the comparison if it falls within the resource allowance. The system then selects the action with the highest expected information gain as the target replacement action. If multiple actions have very similar gains, further selection can be made based on secondary factors such as switching costs and equipment wear.

[0042] In this step, the inclusion of resource constraints ensures that the decision-making process is not purely about maximizing information, but rather about seeking the optimal balance under realistic constraints such as the cycle time of the industrial production line, cost, and equipment lifespan.

[0043] Specifically, after obtaining the target re-shooting action, the acquisition device is controlled to execute the selected target re-shooting action, such as moving the camera to a specified focal length, switching to polarized illumination mode, or triggering high-magnification local shooting. After execution, one or more new image data are obtained. The new data will be integrated into the existing image set in subsequent processes and used to update the decision state vector.

[0044] Understandably, the purpose of this embodiment is to transform reshoot decisions from "experience-based, fixed-template" to "state-driven, value-driven." This reduces low-value or repetitive reshoots, ensuring that collected resources are always focused on the direction that best helps eliminate uncertainty in the current decision. At the same time, since each reshoot has a clear expected information gain, the system behavior becomes interpretable and auditable, facilitating optimization by process engineers and production line traceability.

[0045] Step S103: Fuse the newly added image data and the initial acquisition results, update the determination state vector according to the fused acquisition results, calculate the corresponding comprehensive uncertainty according to the updated determination state vector, and determine whether the termination reshoot condition is met. If it is met, end the acquisition process, and output the updated determination state vector to determine the abnormal state of the transparent film.

[0046] Optionally, in this embodiment, this step is a termination control step for abnormal re-shooting of the transparent film. After each re-shooting action is completed, the newly acquired image information is organically integrated with the historical acquisition results, and the overall uncertainty level of the current system in determining the anomaly is reassessed based on the integrated data, thereby deciding whether to continue acquisition or terminate the process, and finally outputting the determination result.

[0047] "Termination control" refers to the system actively ending data acquisition after meeting preset reliability requirements, rather than passively capturing the entire maximum parameter space. The significance of termination control is to prevent continuous sampling of low-yield regions, ensuring the system completes its task within industrial cycle time.

[0048] Specifically, after the system completes a re-shooting action (such as adding a focal layer, switching bands, or increasing angle sampling), it acquires a set of new image data. This data needs to be strictly registered with the initial acquisition results in terms of spatial location, focal layer alignment, exposure parameters, and viewpoint relationships. Based on the parameter dimensions corresponding to the re-shooting action, the system aligns the new image into the existing multi-parameter image tensor structure according to the ROI region. If the re-shooting is of a new focal layer, it is inserted into the corresponding position in the focal layer sequence; if the re-shooting is of a new angle, it is added to the data slot in the angle dimension; if the re-shooting is of a polarized image, it is fused with the original grayscale or color image as a new modal channel.

[0049] The fusion process also needs to address scale differences between different acquisition parameters. For example, the initial acquisition might use low-magnification full-field imaging, while re-enhancing might involve high-magnification re-imaging of a local ROI. In this case, the system needs to downsample the high-magnification image or interpolate the low-magnification image to ensure that the two are fused at the same spatial resolution. The fused acquisition result is a multi-parameter image data cube with richer dimensions and higher sampling density, providing a data foundation for more reliable state inference later.

[0050] After data fusion is complete, the current anomaly detection state vector is recalculated. During the update process, the system re-invokes the classifier, layer inference model, and artifact detection model to infer the fused multi-parameter image data cube. New image information often alters the probability distribution output by the original model. For example, if the confidence level for identifying a region as a sub-membrane foreign object in the initial acquisition is only 0.6, but a re-captured near-infrared image shows that the region has typical spectral characteristics of sub-membrane foreign objects, the updated confidence level may increase to 0.9. Similarly, if layer inference is ambiguous in the initial acquisition, a re-captured mid-angle image can distinguish between reflections on the membrane and scattering within the membrane, thus changing the layer probability distribution from dispersed to concentrated.

[0051] Simultaneously, the system will update the structured description of "critical evidence gaps." If, after reshooting, it is found that evidence is still insufficient in a certain dimension (e.g., lack of polarization information to rule out specular reflection artifacts), the gap list will retain that dimension; otherwise, it will be removed. Cumulative costs and time will also be added to the overhead of this reshoot.

[0052] Specifically, based on the updated decision state vector, the system calculates the corresponding comprehensive uncertainty.

[0053] The overall uncertainty is a weighted fusion of uncertainties in multiple dimensions, including: classification uncertainty (how hesitant the model is in determining the anomaly category), layer uncertainty (the degree of ambiguity in attributing the anomaly to the membrane / inside / below the membrane), artifact attribution uncertainty (the degree of doubt as to whether the feature is a real anomaly or an imaging artifact), and inadequacy of evidence interpretation (whether the existing evidence is sufficient to provide a complete explanation for manual review or subsequent process steps).

[0054] The overall uncertainty can be calculated using mathematical forms such as entropy, the Gini coefficient, or the probability margin of the model output. The lower the value, the more confident the system is in the current decision; the higher the value, the greater the discrepancy or information gap, requiring further data collection.

[0055] Based on the updated comprehensive uncertainty and the preset termination rule set, determine whether the conditions for stopping data collection are met.

[0056] The termination conditions of the termination rule set include: the overall uncertainty is lower than a preset threshold (indicating sufficient confidence); the expected information gain of all candidate reshoot actions is lower than the gain threshold (indicating that continuing to collect data will not yield marginal improvement); and the cumulative cost or time has reached the budget limit (indicating that resource constraints do not allow for continued use). The system determines that termination is permissible if any one of these conditions is met.

[0057] Once the determination is terminated, the system outputs the updated determination status vector as the final output. This output vector includes not only the final anomaly category label and layer location, but also metadata such as confidence level, uncertainty decomposition, and a list of key evidence, for subsequent defect labeling, process tracing, or manual review. At this point, the detection status of the transparent membrane anomaly is officially determined.

[0058] This example demonstrates how, in multi-parameter acquisition of the same anomalous object, this embodiment uses a uniform information gain function to measure the improvement of the next acquisition action on classification, layer inference, artifact elimination, and result interpretation, thereby adaptively determining the timing of reshooting and termination.

[0059] As described above, the information gain-based transparent membrane abnormality re-shooting and termination control method provided in this application can acquire an initial image set and initial analysis results, construct a current judgment state vector containing classification, layer, artifact, and evidence gap information; generate a candidate re-shooting action set based on the state vector, and calculate the expected information gain of each action for the state vector. This gain integrates the marginal improvement of classification confidence, layer inference, artifact elimination, and interpretation completeness, while deducting acquisition cost and time delay; under resource constraints, select the action with the largest information gain to execute and update the state vector; when the updated state vector meets the preset termination acquisition conditions, the acquisition process ends, thereby improving the expected benefits of re-shooting acquisition actions and increasing visual detection efficiency.

[0060] In one embodiment of the information gain-based transparent film abnormal reshooting and termination control method of this application, it may further include the following: Step S201: Extract the anomaly classification confidence, layer probability confidence, and artifact discrimination confidence from the initial image analysis results, and determine the corresponding anomaly classification results, layer results, and artifact discrimination results respectively; Step S202: Evaluate the evidence coverage integrity of the current image analysis results based on the initial multi-parameter image set, and identify the corresponding key evidence gaps; Step S203: Based on the data collection actions that have been performed, determine the corresponding parameter coverage, cumulative data collection cost, and cumulative time overhead; Step S204: Store the anomaly classification result, the layer result, the artifact determination result, the key evidence gap, the parameter coverage, the cumulative acquisition cost, and the cumulative time cost in a preset hierarchical data structure, and determine the corresponding determination state vector. The determination state vector is used to characterize the comprehensive determination state of the transparent membrane anomaly.

[0061] Optionally, in this embodiment, a pre-trained analysis model is called on the obtained initial multi-parameter image set (e.g., images at different focal layers and angles). The analysis model can be a comprehensive analysis model or a single-state analysis model, which does not affect the application of this embodiment.

[0062] Analysis model outputs: anomaly classification confidence (e.g., "sub-membrane foreign body" probability 75%, "membrane surface dirt" probability 15%); layer probability confidence (e.g., probability distribution of anomalies located on, inside, or under the membrane); artifact discrimination confidence (e.g., probability that the current observation result belongs to a real anomaly or a specular reflection artifact).

[0063] Based on the aforementioned confidence level, the system determines the corresponding anomaly classification results, layer results, and artifact determination results by using preset thresholds or maximum probability criteria.

[0064] Optionally, in this embodiment, multiple evidentiary dimensions of the current image analysis result are checked one by one: whether the focal layer coverage is complete (whether there is a lack of intra- or sub-film layers), whether the observation angle is sufficient (whether there are blind spots), and whether the band information is sufficient (whether near-infrared auxiliary judgment is needed). By comparing with the evidence completeness standard, the system identifies the currently missing key evidentiary dimensions and marks them as "key evidence gaps". The purpose of this step is to clarify "what is still missing" in the current judgment, providing a direct basis for generating subsequent reshoot actions.

[0065] Optionally, in this embodiment, the executed acquisition actions are quantified and statistically analyzed. The cumulative acquisition cost includes the number of image frames captured and the number of hardware actions (such as the number of focus switches and light source switches); the cumulative time cost includes the total time consumed from the start of acquisition to the current moment. This provides a cost benchmark for subsequent information gain calculations, enabling decisions to weigh the "benefits" against the "costs."

[0066] Optionally, in this embodiment, the information from all the above dimensions is organized and stored using a preset hierarchical data structure. The resulting judgment state vector represents the comprehensive judgment state of the current transparent membrane anomaly.

[0067] The hierarchical data structure includes: the top layer is the overall judgment status, the middle layer is the results of each dimension (classification, level, artifact), and the bottom layer is the detailed indicators (confidence distribution, gap list, cost details).

[0068] Through step S204, this embodiment transforms the scattered initial image analysis results into a hierarchical decision state vector, which lays the foundation for the subsequent realization of "decision reshooting based on information gain".

[0069] In one embodiment of the information gain-based transparent film abnormal reshooting and termination control method of this application, it may further include the following: Step S301: Based on the key evidence gaps and parameter coverage in the determination state vector, determine the corresponding collection parameter dimension, wherein the collection parameter dimension represents the parameter dimension that has insufficient sampling confidence or is not currently covered. Step S302: Enumerate and match the corresponding executable actions from the preset candidate reshoot action type library according to the acquisition parameter dimension, and determine the corresponding candidate reshoot action set. The action types in the candidate reshoot action type library include additional focal layer sampling, expanding the focal layer range, reducing the focal layer step size, adding observation angle, switching bands, adding exposure groups, switching illumination modes, performing polarization acquisition, performing high-magnification resampling on local regions of interest, and reselecting reference pose.

[0070] Optionally, in this embodiment, based on the evidence gaps and covered parameters recorded in the current determination state vector, it automatically identifies which collected parameter dimensions are still "insufficient in confidence" or "completely missing".

[0071] The determination state vector contains the parameter records that the system has completed for acquiring abnormal targets, such as which focal layers, angles, wavelengths, or exposure conditions have been captured. At the same time, the vector also contains confidence assessments for various dimensions, such as "the current confidence in determining the subfilm layer is only 40%" or "polarization information that artifact exclusion depends on has not yet been acquired."

[0072] By analyzing this information, we can identify parameter dimensions with insufficient sampling confidence—that is, although some data has been collected, it is not enough to make a stable judgment; and parameter dimensions that have not yet been covered—that is, new dimensions that have not been collected at all but may play a key role in anomaly judgment.

[0073] For example, if the state vector shows that all current focal layer images are concentrated in the supra-film region, while the sub-film region has no data at all, then "sub-film focal layer" is an uncovered parameter dimension. Similarly, if multiple focal layers have been acquired but the classification confidence still fluctuates significantly, then "band information" may be a dimension with insufficient confidence, requiring supplementation with near-infrared or other band data.

[0074] Optionally, in this embodiment, after identifying which parameter dimensions are insufficient, these dimensions are further mapped to specific executable reshoot actions.

[0075] A library of candidate reshoot action types is maintained in advance, which includes a series of standard action types designed for transparent membrane anomaly detection scenarios.

[0076] The action types cover the most commonly used parameter adjustment methods, including: adding focal layer sampling (adding new layers at both ends of the existing focal layer range), expanding the focal layer range (extending to a deeper or shallower direction), reducing the focal layer step size (densifying sampling near key layers), adding observation angles (e.g., adding 30° or 60° oblique shots), switching bands (e.g., switching from visible light to near-infrared), adding exposure groups (e.g., adding long exposures in overly dark areas), switching illumination modes (e.g., switching from bright field to dark field), performing polarization acquisition (eliminating specular reflection interference), performing high-magnification resampling on local regions of interest, and reselecting the reference attitude (adjusting the relative orientation of the camera or sample).

[0077] It is understood that the candidate reshoot actions are derived from the preset action types in the action type library, and the addition of candidate reshoot actions does not affect the implementation of this solution embodiment.

[0078] For each missing or insufficiently confident parameter dimension identified in step S301, the system will perform a matching enumeration in the action type library. For example, if "missing subfilm layer" is identified as an uncovered dimension, the matched actions may include "adding focal layer sampling (towards the subfilm direction)" and "expanding the focal layer range (including deeper layers)". If "severe artifact interference" is identified as an insufficiently confident dimension, the matched actions may include "performing polarization acquisition" and "switching illumination mode".

[0079] In step S302, this embodiment transforms the abstract parameter dimension gap into a set of explicit and executable acquisition instructions, forming a set of candidate reshoot actions for subsequent information gain evaluation module to perform quantitative comparison.

[0080] In one embodiment of the information gain-based transparent film abnormal reshooting and termination control method of this application, it may further include the following: Step S401: For each candidate reshoot action in the candidate reshoot action set, calculate the marginal benefit component of the reshoot action to the determination state vector according to the preset data processing method. The marginal benefit component includes marginal classification gain, marginal layer gain, marginal artifact removal gain, marginal interpretation gain, reshoot action acquisition cost, and time delay. Step S402: Weight the marginal revenue items and sum them to determine the corresponding expected information gain.

[0081] Optionally, in this embodiment, the marginal changes brought about by each candidate action in six different dimensions are quantified for the candidate reshoot action set obtained in step S3.

[0082] Specifically, for a candidate action (such as "adding a focal layer" or "switching to polarization acquisition"), we first evaluate how much the confidence level of anomaly classification can be improved relative to the current state after the action is executed, i.e., the marginal classification gain. For example, if the current judgment of "sub-membrane foreign body" is only 60% certain, and the action is expected to improve the confidence level to 85%, then the classification gain is relatively high.

[0083] Secondly, the system evaluates the marginal layer gain, i.e., the extent to which this action helps determine whether the anomaly is located on, inside, or under the membrane. For transparent membrane inspection, layer determination is often the key to distinguishing between true defects and surface interference.

[0084] Third, the marginal artifact elimination gain measures whether the action can effectively distinguish between real anomalies and artifacts such as specular reflections, out-of-focus areas, and halos. Polarized acquisition or variable-angle shooting usually contributes significantly to this dimension.

[0085] Fourth, the marginal interpretation gain reflects whether the action can make the final test results form a more complete and traceable chain of evidence, which is convenient for manual review or process attribution.

[0086] In addition, two cost items are calculated: the cost of capturing reshooting actions (such as additional shooting times and hardware actions) and time delay (such as the time required to switch light sources and move the platform).

[0087] These six sub-items are calculated independently, allowing the "good" and "bad" of each action to be broken down into interpretable dimensions, avoiding the one-sidedness of a single indicator. The independent calculation process of each sub-item can call rule models, lightweight visual models, historical experience bases, or semantic agents to estimate the marginal benefits that each candidate action may bring.

[0088] After obtaining six marginal benefit components, the system performs a weighted summation of them to obtain the overall expected information gain of the candidate action. The weight of each component can be dynamically adjusted according to the needs of the scenario.

[0089] For example, when production line cycles are tight, the weight of time delay is increased, so that even if time-consuming actions have high gains in other dimensions, the overall score will be suppressed; while in R&D verification scenarios, the weight of explanatory gain and layer gain is higher in order to obtain richer evidence data.

[0090] After weighted summation, each candidate retake action is assigned a comparable value—the expected information gain. This value comprehensively reflects how "worth doing" the action: the higher the positive gain, the more significant the increase in effective evidence after deducting costs; if the gain is zero or negative, the action is not worth performing.

[0091] Optionally, for example, in a preferred embodiment, let the state vector of the current candidate ROI be X. t The set of the next candidate replacement shots is A={a1,a2,...,a...} n}, then action a i The overall information gain function can be expressed as:

[0092] Wherein, ΔC(a) i ) indicates the expected increase in confidence level for the anomaly category determination due to this action, ΔL(a) i ) represents the expected improvement in the stability of inferred on-membrane / intra-membrane / submembrane sites, ΔR(a i ) represents the expected improvement in the ability to eliminate artifacts and confirm real anomalies, ΔE(a) i Cost(a) represents the expected improvement in the interpretability of the final result. i ) represents the acquisition cost, Delay(ai This represents the cost of clock delay, with w1 to w6 being weighting coefficients.

[0093] The prediction calculation of the above marginal revenue items can be performed by calling rule models, lightweight visual models, historical experience bases, or semantic agents.

[0094] Through step S402, this embodiment successfully transforms the originally ambiguous decision problem of "which action is better" into a numerical calculation problem. The decision-making process is transparent and interpretable, and the contribution of each dimension is traceable.

[0095] In one embodiment of the information gain-based transparent film abnormal reshooting and termination control method of this application, it may further include the following: Step S501: Rule calculation, performing rule matching on the candidate reshoot action and the judgment state vector according to the preset rule library to determine the corresponding marginal benefit item; Step S502: Lightweight visual model calculation: Based on the feature vector of the acquired image, perform regression prediction on the potential benefits after the candidate reshoot action is executed, and determine the corresponding marginal benefit items. Step S503: Historical experience database matching calculation, retrieve the historical sample with the highest similarity to the current judgment state vector, and calculate its actual marginal benefit as the marginal benefit component of the current candidate retake action; Step S504: Semantic agent computation, based on a large language model or knowledge graph, reasoning about the causal relationship between the current anomaly type and candidate reshoot actions, and determining the corresponding marginal benefit items.

[0096] Optionally, in this embodiment, the following four methods can be used to predict the marginal revenue values ​​of the six components of the comprehensive information gain function: First, the system operates based on a pre-defined rule base. This rule base, compiled by domain experts, is structured as "If the current state satisfies a certain condition, then a certain component of the reward for a candidate action is a certain value." The system receives the current decision state vector and candidate reshoot actions, and then matches each rule with the others. For example, if the state vector shows "classification confidence is below 0.6 and artifact probability is above 0.7," the rule could output "the marginal benefit of artifact exclusion for polarization acquisition actions is high." The rule base can quickly and interpretably provide basic reward estimates, making it suitable for scenarios with clear experience and causal relationships, and it can run stably without model training.

[0097] Second, a small neural network or regression model is used. Feature vectors (such as depth features or statistical features) of the acquired images are extracted and input into the model. The model directly regresses and predicts the improvement in indicators such as classification confidence and layer stability after each candidate reshoot action. This model is usually trained offline, using the actual gains before and after historical acquisition actions as supervision signals. The lightweight model can capture the non-linear relationship between image content and reshoot gains, such as predicting that "the current image already contains weak edge information, and the layer gain can reach 0.3 after adding a focal layer," making it more flexible and adaptable to diverse materials and defect morphologies than rule-based methods.

[0098] Third, maintain a historical experience database. Each record in the database contains the decision state vector at that time, the re-action performed, and the actual marginal benefit gained. At the current moment, the system uses the current state vector as a query, retrieves the most similar historical samples from the database, and calculates the average actual benefit of the same candidate action in these samples as the current marginal benefit component. The experience database can use real-world experience for analogical decision-making, which is especially suitable for periodic production scenarios—if an action was effective when similar anomalies occurred in the past, it is likely to be equally effective now, eliminating the need for repeated modeling.

[0099] Fourth, integrate large language models or domain knowledge graphs. Information such as the current anomaly type, collected parameters, and candidate reshoot actions are organized into natural language prompts or structured queries. The agent, based on its inherent knowledge or causal edges in the knowledge graph, infers the potential benefits of the action. For example, the agent might output, "For foreign objects under transparent membranes, near-infrared bands can enhance deep scattering signals, resulting in higher benefits from layer location determination." Large language models can handle zero-sample or rare anomaly scenarios, utilizing pre-trained general knowledge for reasonable extrapolation, filling blind spots not covered by rules, models, or historical samples.

[0100] The above four steps can be used in parallel or one of them can be used, without affecting the implementation of this solution embodiment.

[0101] Through step S504, this embodiment successfully enables the system to obtain a reliable, multi-dimensional marginal benefit estimate under any circumstances, thereby supporting the accurate selection of the optimal reshoot action.

[0102] In one embodiment of the information gain-based transparent film abnormal reshooting and termination control method of this application, it may further include the following: Step S601: Based on the updated judgment state vector, extract the corresponding anomaly classification uncertainty, layer uncertainty, artifact attribution uncertainty, and evidence interpretation insufficiency. Step S602: The anomaly classification uncertainty, the layer uncertainty, the artifact attribution uncertainty, and the evidence interpretation insufficiency uncertainty are weighted and summed to determine the corresponding comprehensive uncertainty.

[0103] Optionally, in this embodiment, the determination process of comprehensive uncertainty is described.

[0104] After each reshoot and update of the judgment state vector, the system first needs to decouple four key uncertainty components from this state vector: anomaly classification uncertainty, layer uncertainty, artifact attribution uncertainty, and insufficient evidence interpretation uncertainty. This decoupling process can be achieved through mapping rules or a lightweight prediction model.

[0105] Anomaly classification uncertainty reflects the degree of hesitation the system has in determining which type of defect (such as particles, scratches, dirt, foreign objects, etc.) the current anomaly belongs to. It originates from the output probability distribution of the classification model: if the probability of one category is much higher than other categories, the uncertainty is low; if the probabilities of multiple categories are similar, the uncertainty is high.

[0106] Layer location uncertainty reflects the reliability of the system's judgment regarding whether an anomaly is located on, inside, or below the membrane. In transparent membrane testing, layer location determination directly determines subsequent processing techniques (such as cleaning, replacement, or ignoring), and its uncertainty stems from peak ambiguity or cross-layer response overlap in focal layer sequence analysis.

[0107] Artifact attribution uncertainty reflects the extent to which currently observed anomalous features in the system may be optical artifacts rather than real defects. For example, artifacts such as specular reflection, halos, and focal plane intersections may be highly similar to real anomalies in an image, and this uncertainty is used to quantify this risk of confusion.

[0108] Insufficiency of evidence is a higher-level measure, indicating whether the existing set of evidence is sufficient to provide a complete, self-consistent, and traceable explanation for manual review or process steps. Even if the classification and stratification are clear, the insufficiency score will still be high if evidence for key dimensions is missing.

[0109] After obtaining the four uncertainty components, the system performs a weighted summation to obtain the comprehensive uncertainty. In one embodiment, the comprehensive uncertainty U can be expressed as: U=α·U cls +β·U layer +γ·U artifact +δ·U explain Among them, U cls U represents the classification uncertainty. layer U represents the layer uncertainty. artifactU represents the uncertainty in artifact attribution. explain The α, β, γ, and δ values ​​represent the degree of inadequacy of the evidence interpretation and are adjustable weights.

[0110] The setting of the comprehensive uncertainty weight coefficient reflects the priority of concern in different application scenarios. For example, in detection tasks with high safety levels, the weight γ of artifact attribution uncertainty will be increased to prevent artifacts from being misclassified as real anomalies; in scenarios with strict requirements for process attribution, the weight δ of insufficient evidence interpretation will be significantly increased to ensure that the output results are supported by a complete chain of evidence; and in scenarios where time-priority is prioritized, the weights α and β of classification uncertainty and hierarchical uncertainty can dominate, enabling the system to quickly converge to a clear determination and then terminate.

[0111] Through step S602, this embodiment successfully compresses the high-dimensional, heterogeneous determination state into a quantifiable comprehensive uncertainty index by weighted summation of comprehensive uncertainties. This index ultimately serves as one of the core criteria for terminating data collection, guiding the system to proactively stop data collection when the uncertainty is sufficiently low, avoiding unnecessary redundancy, and ensuring the reliability and interpretability of the output results.

[0112] In one embodiment of the information gain-based transparent film abnormal reshooting and termination control method of this application, it may further include the following: Step S701: The overall uncertainty is lower than the preset uncertainty threshold; Step S702: The expected information gain of all candidate reshoot actions in the candidate reshoot action set is lower than a preset gain threshold; Step S703: The cumulative acquisition cost or cumulative time expenditure reaches the preset resource constraint condition.

[0113] Optionally, in this embodiment, the system recalculates the current termination index T_stop after each round of supplementary shooting, and terminates the acquisition when at least one of the following conditions is met: 1. The overall uncertainty is lower than the threshold U_th; 2. The information gain of all candidate actions is lower than the threshold G_th; 3. The cost budget, time budget, or hardware count budget reaches the upper limit; or the key judgment result has been stably and continuously maintained for a preset number of rounds.

[0114] Specifically, the first condition is that the overall uncertainty is below a preset uncertainty threshold. The core issue is determining whether the system's overall confidence in the anomaly detection results has reached an acceptable level. The preset uncertainty threshold is set in advance based on the production line's requirements for detection reliability. For example, a lower threshold can be set for production lines with high requirements, while a more relaxed threshold can be set for production lines sensitive to cycle time. When the overall uncertainty is below this threshold, it means that the existing image evidence is sufficient for the system to form stable and reliable conclusions regarding key aspects such as the anomaly type, its location, and whether it is an artifact. Continued acquisition offers very limited marginal improvement.

[0115] The second condition is that the expected information gain of all candidate reshoot actions is lower than a preset gain threshold. The decision to continue acquisition is based on the value of the actions. Each action in the candidate reshoot action set corresponds to an expected information gain, which integrates its improvement to classification, layering, artifact elimination, and interpretive completeness, while deducting acquisition costs and time. The preset gain threshold is a minimum acceptable benefit threshold. When the expected gain of all candidate actions is lower than this threshold, it means that regardless of what reshoot action the system performs next, the marginally effective evidence obtained will be very small, and may even be counterproductive. Continuing acquisition at this point will not only fail to improve the judgment result but will also consume resources. Triggering termination prevents the system from repeatedly sampling in "low-reward areas," ensuring that acquisition resources are always allocated to high-value actions.

[0116] The third condition is a hard termination boundary set from an engineering resource perspective. During the data acquisition process, the system continuously accumulates the equipment wear and tear costs, storage costs, and total time consumed by the executed actions. Preset resource constraints include the maximum allowed number of acquisitions, the maximum allowed time budget, or the maximum storage capacity. When the accumulated cost reaches any of these constraint limits, the system must be forcibly terminated even if it has not yet reached the desired level of confidence. Its effect is to ensure that the system completes its tasks within the cycle time constraints and hardware lifespan constraints of the industrial production line, preventing production line blockages or equipment overload caused by the pursuit of perfect evidence. This condition provides the system with a controllable "worst-case boundary," ensuring that the algorithm can still output the current optimal judgment result under limited resources.

[0117] Through step S703, this embodiment successfully controls the anomaly detection and reshooting process by setting termination conditions, thereby reducing redundant acquisition and improving cycle efficiency.

[0118] To improve the expected benefits of re-capture actions and increase visual detection efficiency, this application provides an embodiment of an information gain-based transparent film abnormal capture and termination control device for implementing all or part of the aforementioned information gain-based transparent film abnormal capture and termination control method. See [link to embodiment]. Figure 2 The information gain-based transparent film abnormal reshoot and termination control device specifically includes the following components: The initial acquisition result determination module 10 is used to acquire the initial acquisition results of the transparent membrane anomaly detection. The initial acquisition results include an initial multi-parameter image set and initial image analysis results. The reshooting action determination module 20 is used to construct a judgment state vector based on the initial acquisition results. The judgment state vector is used to characterize the comprehensive judgment state of the transparent membrane abnormality. Based on the judgment state vector, a corresponding set of candidate reshooting actions is determined. Marginal benefit calculation is performed on each candidate reshooting action in the set of candidate reshooting actions to determine the corresponding expected information gain. Under the condition of satisfying the preset resource constraints, the target reshooting is performed based on the candidate reshooting action with the largest expected information gain to determine the corresponding new image data. The reshoot termination control module 30 is used to fuse the newly added image data and the initial acquisition result, update the determination state vector according to the fused acquisition result, calculate the corresponding comprehensive uncertainty according to the updated determination state vector, and determine whether the reshoot termination condition is met. If it is met, the acquisition process ends, and the updated determination state vector is output to determine the abnormal state of the transparent film.

[0119] As described above, the information gain-based transparent film abnormality re-shooting and termination control device provided in this application can acquire an initial image set and initial analysis results, construct a current judgment state vector containing classification, layer, artifact, and evidence gap information; generate a candidate re-shooting action set based on the state vector, and calculate the expected information gain of each action for the state vector. This gain integrates the marginal improvement of classification confidence, layer inference, artifact elimination, and interpretation completeness, while deducting acquisition cost and time delay; under resource constraints, select the action with the largest information gain to execute and update the state vector; when the updated state vector meets the preset termination acquisition conditions, the acquisition process ends, thereby improving the expected benefits of re-shooting acquisition actions and increasing visual detection efficiency.

[0120] To further illustrate this solution, this application also provides a specific application example of the information gain-based transparent film abnormal reshooting and termination control method implemented using the aforementioned information gain-based transparent film abnormal reshooting and termination control device, which specifically includes the following: Example 1: For suspected sub-membrane foreign objects, the system has obtained sufficient focal layer evidence in the initial acquisition, but the confidence level for determining the stratigraphic position is still insufficient. Candidate actions include adding near-infrared band, adding mid-angle shots, and expanding the focal layer range. It is predicted that the near-infrared band will have the greatest improvement on stratigraphic stability and the switching cost is moderate, so near-infrared supplementary shooting is performed first; if the stratigraphic probability distribution has converged significantly after supplementary shooting, the process is terminated.

[0121] Example 2: For highly reflective targets, the system found the classification results to be relatively stable, but artifact removal was insufficient. Candidate actions included polarization acquisition, short-exposure group reshooting, and local ROI refinement. Since polarization acquisition can significantly reduce the influence of specular reflection, it has the highest artifact removal gain, and the system prioritizes polarization reshooting.

[0122] Example 3: For point particles, if the system continues to increase the number of focal layers without significantly improving the judgment quality when there is insufficient multi-angle data, the comprehensive gain function will suppress the invalid focal layer reshoot and prioritize the angle expansion action, thereby prioritizing the observation of highlight center drift and halo diffusion in a limited number of beats.

[0123] Example 4: In scenarios with tight production line cycles, the weight of Delay(a_i) can be significantly increased, causing the system to tend to terminate as early as possible after reaching the minimum confidence threshold; in R&D verification scenarios, the weights of interpretive completeness and hierarchical stability can be increased to obtain richer evidence data.

[0124] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described information gain-based transparent film abnormal re-shooting and termination control method.

[0125] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for abnormal re-shooting and termination control of transparent films based on information gain.

[0126] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for abnormal re-shooting and termination control of transparent films based on information gain.

[0127] In this embodiment of the invention, an initial image set and initial analysis results are collected to construct a current judgment state vector containing information on classification, layer, artifacts, and evidence gaps. Based on this state vector, a set of candidate re-shooting actions is generated, and the expected information gain of each action on the state vector is calculated. This gain integrates the marginal improvement of classification confidence, layer inference, artifact elimination, and interpretation completeness, while deducting the acquisition cost and time delay. Under resource constraints, the action with the largest information gain is selected for execution and the state vector is updated. The acquisition process ends when the updated state vector meets the preset termination conditions. This improves the expected benefits of re-shooting actions and increases visual detection efficiency.

[0128] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0130] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0132] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for abnormal reshooting and termination control of transparent films based on information gain, characterized in that, The method includes: Acquire initial acquisition results for transparent membrane anomaly detection, the initial acquisition results including an initial multi-parameter image set and initial image analysis results; A judgment state vector is constructed based on the initial acquisition results. The judgment state vector is used to characterize the comprehensive judgment state of the transparent membrane abnormality. A corresponding set of candidate reshoot actions is determined based on the judgment state vector. Marginal benefit calculation is performed on each candidate reshoot action in the set of candidate reshoot actions to determine the corresponding expected information gain. Under the condition of satisfying the preset resource constraints, the target reshoot is performed based on the candidate reshoot action with the largest expected information gain to determine the corresponding new image data. The newly added image data and the initial acquisition results are fused together, and the determination state vector is updated according to the fused acquisition results. The corresponding comprehensive uncertainty is calculated based on the updated determination state vector, and it is determined whether the termination reshoot condition is met. If it is met, the acquisition process ends, and the updated determination state vector is output to determine the abnormal state of the transparent film.

2. The method for abnormal reshooting and termination control of transparent films based on information gain according to claim 1, characterized in that, The step of constructing a determination state vector based on the initial acquisition results includes: The anomaly classification confidence, layer probability confidence, and artifact discrimination confidence are extracted from the initial image analysis results, and the corresponding anomaly classification results, layer results, and artifact discrimination results are determined respectively. The evidence coverage integrity of the current image analysis results is evaluated based on the initial multi-parameter image set to identify the corresponding key evidence gaps. Based on the data collection actions that have been performed, determine the corresponding parameter coverage, cumulative data collection cost, and cumulative time overhead. The anomaly classification results, the layer results, the artifact determination results, the key evidence gaps, the parameter coverage, the cumulative acquisition cost, and the cumulative time overhead are stored in a preset hierarchical data structure to determine the corresponding determination state vector. The determination state vector is used to characterize the comprehensive determination state of transparent membrane anomalies.

3. The method for abnormal reshooting and termination control of transparent films based on information gain according to claim 2, characterized in that, The step of determining the corresponding candidate retake action set based on the determination state vector includes: Based on the key evidence gaps and parameter coverage in the judgment state vector, the corresponding collection parameter dimension is determined. The collection parameter dimension represents the parameter dimension that has insufficient sampling confidence or is not currently covered. Based on the acquisition parameter dimensions, corresponding executable actions are enumerated and matched from the preset candidate reshoot action type library to determine the corresponding candidate reshoot action set. The action types in the candidate reshoot action type library include additional focal layer sampling, expanding the focal layer range, reducing the focal layer step size, adding observation angle, switching bands, adding exposure groups, switching illumination modes, performing polarization acquisition, performing high-magnification resampling on local regions of interest, and reselecting reference pose.

4. The method for abnormal reshooting and termination control of transparent films based on information gain according to claim 1, characterized in that, The step of calculating the marginal benefit of each candidate retake action in the candidate retake action set to determine the corresponding expected information gain includes: For each candidate reshoot action in the candidate reshoot action set, the marginal benefit component of the reshoot action to the determination state vector is calculated according to the preset data processing method. The marginal benefit component includes marginal classification gain, marginal layer gain, marginal artifact removal gain, marginal interpretation gain, reshoot action acquisition cost, and time delay. The weighted sum of each marginal benefit item is used to determine the corresponding expected information gain.

5. The method for abnormal reshooting and termination control of transparent films based on information gain according to claim 4, characterized in that, The data processing methods for calculating the marginal revenue components include: Rule calculation involves matching the candidate reshoot action and the judgment state vector according to a preset rule base to determine the corresponding marginal benefit item. The lightweight visual model is used to perform regression prediction on the potential benefits after the candidate reshoot action is executed, based on the feature vector of the acquired image, to determine the corresponding marginal benefit items. Historical experience database matching calculation: retrieve the historical sample with the highest similarity to the current judgment state vector, and calculate its actual marginal benefit as the marginal benefit component of the current candidate retake action; Semantic intelligent agent computing, based on large language models or knowledge graphs, infers the causal relationship between the current anomaly type and candidate reshoot actions, and determines the corresponding marginal benefit items.

6. The method for abnormal reshooting and termination control of transparent films based on information gain according to claim 1, characterized in that, The calculation of the corresponding comprehensive uncertainty based on the updated decision state vector includes: Based on the updated judgment state vector, extract the corresponding anomaly classification uncertainty, layer uncertainty, artifact attribution uncertainty, and evidence insufficiency. The overall uncertainty is determined by weighted summation of the anomaly classification uncertainty, the layer uncertainty, the artifact attribution uncertainty, and the inadequacy of evidence interpretation.

7. The method for abnormal reshooting and termination control of transparent films based on information gain according to claim 1, characterized in that, The conditions for terminating reshooting include at least one of the following: The overall uncertainty is lower than a preset uncertainty threshold; The expected information gain of all candidate reshoot actions in the candidate reshoot action set is lower than a preset gain threshold. The cumulative collection cost or cumulative time expenditure has reached the preset resource constraints.

8. A transparent film abnormality reshoot and termination control device based on information gain, characterized in that, The device includes: The initial acquisition result determination module is used to obtain the initial acquisition results of the transparent membrane anomaly detection. The initial acquisition results include an initial multi-parameter image set and initial image analysis results. The reshooting action determination module is used to construct a judgment state vector based on the initial acquisition results. The judgment state vector is used to characterize the comprehensive judgment state of the transparent membrane abnormality. Based on the judgment state vector, a corresponding set of candidate reshooting actions is determined. Marginal benefit calculation is performed on each candidate reshooting action in the set of candidate reshooting actions to determine the corresponding expected information gain. Under the condition of satisfying the preset resource constraints, the target reshooting is performed based on the candidate reshooting action with the largest expected information gain to determine the corresponding new image data. The reshoot termination control module is used to fuse the newly added image data and the initial acquisition result, update the determination state vector according to the fused acquisition result, calculate the corresponding comprehensive uncertainty according to the updated determination state vector, and determine whether the reshoot termination condition is met. If it is met, the acquisition process ends, and the updated determination state vector is output to determine the abnormal state of the transparent film.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the information gain-based transparent film abnormal reshoot and termination control method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the information gain-based transparent film abnormal reshoot and termination control method as described in any one of claims 1 to 7.

Images