A method and system for suppressing false detections of substation equipment

By using a two-stage discriminator and dynamic geometric relationships in substation equipment detection, the problems of high false detection rate and low recall rate in the background are solved, and efficient equipment detection in complex scenarios is achieved.

CN123071308APending Publication Date: 2026-11-10CHINA THREE GORGES UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611336923.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies for substation equipment inspection suffer from high background false detection rates, low recall rates, and the tendency for small targets to be mistakenly deleted. In particular, they are difficult to effectively distinguish between real equipment and high-confidence background false detections in complex scenarios.

Method used

A candidate box is generated using an object detection model and semantically verified using a two-level discriminator. By combining soft preservation coefficients and dynamic geometric relationships, small targets and occluded targets are protected through continuous score adjustment and fusion weights, and candidate box clustering and fusion are dynamically adjusted.

Benefits of technology

It effectively reduces the false positive rate, maintains recall levels, improves the reliability and accuracy of test results, and adapts to complex scene changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071308A_ABST
    Figure CN123071308A_ABST
Patent Text Reader

Abstract

This application provides a method and system for suppressing false detections of substation equipment. The method includes: generating candidate bounding boxes of a substation image using a target detection model and cropping candidate regions to obtain the category and detection confidence of the candidate bounding boxes; inputting the candidate regions into a two-level discriminator and outputting the verification probability that the candidate regions belong to real equipment; combining the detection confidence and verification probability into a continuously corrected score using a soft retention coefficient, without directly deleting candidate bounding boxes; dynamically determining the candidate bounding box clustering threshold and fusion weight, wherein the verification probability participates in the calculation of both the continuously corrected score and the fusion weight; performing retention protection on single candidate bounding boxes with an area ratio lower than the threshold and outputting the fused detection result. The system includes: a candidate bounding box generation module, a candidate region preprocessing module, a hard-to-bear sample two-level discriminator module, a soft score correction module, a dynamic distance-constrained weighted bounding box fusion module, and a result output module, effectively maintaining recall while reducing background false detections.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of computer vision, image target detection, and intelligent inspection of power equipment, and particularly to a method and system for suppressing false detections of substation equipment. Background Technology

[0002] Substations have a variety of primary equipment and complex spatial structures. Target detection models in field images easily misidentify background areas with similar local appearances, such as wires, metal supports, or repetitive textures, as equipment, generating a large number of false background bounding boxes. Existing technologies have at least the following shortcomings when processing these candidate bounding boxes: First, traditional post-processing methods mainly rely on geometric relationships such as intersection-union ratio (IU) and detection confidence for screening, lacking semantic discrimination capabilities and struggling to distinguish between real equipment and high-confidence background false detections. Second, to reduce semantic false detections, existing technologies typically use fixed high thresholds for hard filtering of secondary verification probabilities. This direct deletion of low-probability candidate boxes easily leads to the simultaneous deletion of small targets with incomplete local features and occluded targets, resulting in a significant decrease in recall. Third, existing box fusion rules typically use fixed IU and center distance thresholds, which cannot adaptively adjust with target scale and local density of candidate boxes. In complex scenes, this can easily lead to erroneous merging and the accidental deletion of dense small targets lacking repeating candidate boxes.

[0003] Therefore, there is an urgent need for a method to suppress false detections of substation equipment that can jointly utilize semantic verification probability and dynamic geometric relationships. This method can replace hard deletion with soft score correction to protect small and occluded targets, and adapt to candidate box clustering and fusion in complex scenes, thereby effectively maintaining recall while reducing background false detections. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problems mentioned above and to propose a method for suppressing false detections of substation equipment, comprising the following steps: The object detection model generates candidate bounding boxes for substation images and crops the candidate regions, obtaining the category and detection confidence of the candidate bounding boxes; The candidate region is input into the secondary discriminator, which outputs the verification probability that the candidate region belongs to the real device. The secondary discriminator is trained using hard negative samples automatically constructed from similar real bounding boxes and background false detection boxes. The detection confidence and validation probability are combined into a continuously corrected score using a soft retention coefficient, without directly deleting candidate boxes; The candidate box clustering threshold is dynamically determined based on the target scale, local candidate density, and continuous correction score, and the fusion weight is calculated based on the continuous correction score, intersection-union ratio, center consistency, and verification probability. Single candidate boxes with an area ratio below a threshold are preserved and protected, and the fusion detection results are output.

[0005] In the preferred embodiment, the training data construction steps for the secondary discriminator include: The baseline detection model is used to generate training candidate boxes for the training images, and each training candidate box is matched with the ground truth box of the same category by intersection-union matching. Training candidate regions with a maximum crossover-union ratio (CUI) not less than the matching threshold are labeled as positive samples from the real device, while training candidate regions with a maximum CUI less than the matching threshold are labeled as negative samples from the background that are difficult to detect false positives.

[0006] In the preferred scheme, the calculation steps for the continuous correction score include: assuming the detection confidence level is... The verification probability is The soft retention factor is ,and Continuous score correction satisfy: .

[0007] In the preferred embodiment, the calculation steps for the target scale and local candidate density include: Let the area of ​​the candidate box be The area of ​​the input image is The target scale is at the level of a small target. ,satisfy ; Using the center of the candidate box as the center and three times the length of the candidate box's diagonal as the radius, count the number of candidate boxes within the neighborhood. ,by Represents the density normalization constant, local candidate density satisfy .

[0008] In the preferred scheme, the step of dynamically determining the candidate box clustering threshold includes: The cross-union ratio threshold and the normalized center distance threshold are dynamically calculated by combining the small target size, local candidate density, and detection confidence. If the intersection-union ratio (IU) between the candidate box to be matched and the representative box of the existing cluster is not less than the IU threshold, and the normalized center distance is not greater than the normalized center distance threshold, the candidate box to be matched is assigned to the existing cluster; otherwise, a new cluster is created.

[0009] In the preferred embodiment, the calculation steps for the fusion weights include: Whenever a new candidate box is added to a cluster, the weights are recalculated for all candidate boxes within the cluster. The fusion weight is calculated by continuously adjusted scores, crossover ratio, central consistency and validation probability, and the bounding box coordinates and cluster scores are updated based on the fusion weight.

[0010] In the preferred embodiment, the steps for performing retention protection include: Determine whether the area of ​​the candidate box relative to the image area is less than the threshold for the proportion of small target area. If yes, use the small target retention threshold; otherwise, use the basic retention threshold. After clustering is completed, clusters containing two or more candidate boxes are output directly, while single-box clusters containing only a single candidate box are retained and output when they reach the corresponding retention threshold.

[0011] A substation equipment false detection suppression system is also proposed, including: The candidate box generation module is used to read substation images and output the first... The coordinates, category labels, and detection confidence of each candidate bounding box; The candidate region preprocessing module is used to crop a local image from the original image according to the candidate box, and perform uniform scaling and normalization to obtain the candidate region; The difficult-to-bear sample secondary discrimination module is used to input candidate regions into the secondary discriminator and output the secondary verification probability that the candidate regions belong to the real device targets; The soft score correction module is used to use the secondary validation probability as a continuous adjustment factor, and combine the soft retention coefficient with the detection confidence to form a continuous correction score. This module reduces the weight and rearranges the candidate boxes, but does not directly delete the candidate boxes. The dynamic distance-constrained weighted box fusion module is used to dynamically cluster and calculate fusion weights based on target scale, local candidate density, intersection-union ratio, normalized center distance and validation probability, and to perform small target single-box protection for single candidate boxes; The results output module is used to summarize the fused detection boxes, sort them by fusion score, and give the final detection results according to the preset maximum number of single images to be output.

[0012] In the preferred embodiment, in the secondary discrimination module for hard negative samples, the training data of the secondary discriminator is automatically constructed from the target detection training set. Candidate regions with a maximum crossover union ratio (CUB) of not less than the matching threshold are marked as real device positive samples, and candidate regions with a CUB less than the matching threshold are marked as background false positive hard negative samples.

[0013] In the preferred embodiment, the dynamic distance-constrained weighted box fusion module is specifically used to dynamically calculate the cross-union ratio (CUNR) clustering threshold and the center distance threshold by combining the small target degree, local density and candidate confidence of the candidate boxes to be matched, and to determine whether the candidate boxes belong to the same cluster by using the dual conditions of CUNR and center distance.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention adjusts the clustering threshold by jointly adjusting the target scale, local density, confidence, IoU and center distance, so that the box fusion rule can adapt to ordinary, occluded, complex background and dense small target scenes.

[0015] (2) The present invention identifies semantic false detections such as wires, brackets and metal structures by a hard negative sample discriminator, which makes up for the inability of simple geometric post-processing to distinguish between real equipment and false detections in high confidence background.

[0016] (3) The present invention replaces the fixed high threshold hard deletion with soft score correction, so that false detection candidates are continuously downweighted, while retaining the opportunity for real small targets and occluded targets to enter the subsequent fusion.

[0017] (4) The present invention uses the two-level verification probability to participate in the score re-ranking and dynamic fusion weight calculation simultaneously, forming a semantic verification and geometric fusion closed loop, thereby improving the reliability of the fusion result.

[0018] (5) All thresholds, soft retention coefficients and category protection parameters of the present invention can be configured, and the operating point can be adjusted according to the cost of false detection and the cost of missed detection.

[0019] (6) Experimental data verification: On 752 original verification images, D-WD-WBF reduced FP from 7183 to 6579 compared to WBF, a reduction of 604; Precision increased from 0.6309 to 0.6505; F1 increased from 0.6020 to 0.6099; and Recall changed from 0.5756 to 0.5741. This shows that dynamic geometric fusion can reduce false positives while maintaining recall. After further adding a soft verification mechanism (soft retention coefficient α=0.50), compared with using dynamic distance constraint weighted box fusion alone, the number of false positives decreased from 6579 to 2852, a reduction of 56.65%; Precision increased from 0.6505 to 0.7972; F1 value increased from 0.6099 to 0.6335; and Recall decreased from 0.5741 to 0.5256. The results indicate that the main implementation achieves a better overall balance between false positive suppression and recall maintenance, but this does not mean that recall is completely unaffected. Attached Figure Description

[0020] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0021] Figure 2 This is a flowchart of the D-WD-WBF dynamic clustering and fusion process.

[0022] Figure 3 Flowchart for training a secondary discriminator for difficult negative samples.

[0023] Figure 4 This is a flowchart of soft verification and dynamic fusion reasoning.

[0024] Figure 5 The image shows a comparison of the actual detection results of D-WD-WBF, fixed threshold hard filtering, and soft validation schemes in complex scenarios with high false positives (green box represents GT, blue box represents TP, red box represents FP, and yellow box represents FN). Detailed Implementation

[0025] Example 1 A method for suppressing false detections of substation equipment includes the following steps: The object detection model generates candidate bounding boxes for substation images and crops the candidate regions, obtaining the category and detection confidence of the candidate bounding boxes; The candidate region is input into the secondary discriminator, which outputs the verification probability that the candidate region belongs to the real device. The secondary discriminator is trained using hard negative samples automatically constructed from similar real bounding boxes and background false detection boxes. The detection confidence and validation probability are combined into a continuously corrected score using a soft retention coefficient, without directly deleting candidate boxes; The candidate box clustering threshold and fusion weight are dynamically determined based on the target scale, local candidate density, continuous correction score, cross-union ratio, and normalized center distance, where the validation probability participates in the calculation of both the continuous correction score and the fusion weight. Single candidate boxes with an area ratio below a threshold are preserved and protected, and the fusion detection results are output.

[0026] Preferably, the training data construction steps for the secondary discriminator include: The baseline detection model is used to generate training candidate boxes for the training images, and each training candidate box is matched with the ground truth box of the same category by intersection-union matching. Training candidate regions with a maximum crossover-union ratio (CUI) not less than the matching threshold are labeled as positive samples from the real device, while training candidate regions with a maximum CUI less than the matching threshold are labeled as negative samples from the background that are difficult to detect false positives.

[0027] Preferably, the calculation steps for the continuously corrected score include: setting the detection confidence level as... The verification probability is The soft retention factor is ,and Continuous score correction satisfy: .

[0028] Preferably, the calculation steps for the target scale and local candidate density include: Let the area of ​​the candidate box be The area of ​​the input image is The target scale is at the level of a small target. ,satisfy ; Using the center of the candidate box as the center and three times the length of the candidate box's diagonal as the radius, count the number of candidate boxes within the neighborhood. ,by Represents the density normalization constant, local candidate density satisfy .

[0029] Preferably, the step of dynamically determining the candidate box clustering threshold includes: The cross-union ratio threshold and the normalized center distance threshold are dynamically calculated by combining the small target size, local candidate density, and detection confidence. If the intersection-union ratio (IU) between the candidate box to be matched and the representative box of the existing cluster is not less than the IU threshold, and the normalized center distance is not greater than the normalized center distance threshold, the candidate box to be matched is assigned to the existing cluster; otherwise, a new cluster is created.

[0030] Preferably, the calculation steps for the fusion weights include: Whenever a new candidate box is added to a cluster, the weights are recalculated for all candidate boxes within the cluster. The fusion weight is calculated by continuously adjusted scores, crossover ratio, central consistency and validation probability, and the bounding box coordinates and cluster scores are updated based on the fusion weight.

[0031] Preferably, the steps for performing retention protection include: Determine whether the area of ​​the candidate box relative to the image area is less than the threshold for the proportion of small target area. If yes, use the small target retention threshold; otherwise, use the basic retention threshold. After clustering is completed, clusters containing two or more candidate boxes are output directly, while single-box clusters containing only a single candidate box are retained and output when they reach the corresponding retention threshold.

[0032] A substation equipment false detection suppression system is also proposed, including: The candidate box generation module is used to read substation images and output the first... The coordinates, category labels, and detection confidence of each candidate bounding box; The candidate region preprocessing module is used to crop a local image from the original image according to the candidate box, and perform uniform scaling and normalization to obtain the candidate region; The difficult-to-bear sample secondary discrimination module is used to input candidate regions into the secondary discriminator and output the secondary verification probability that the candidate regions belong to the real device targets; The soft score correction module is used to use the secondary validation probability as a continuous adjustment factor, and combine the soft retention coefficient with the detection confidence to form a continuous correction score. This module reduces the weight and rearranges the candidate boxes, but does not directly delete the candidate boxes. The dynamic distance-constrained weighted box fusion module is used to dynamically cluster and calculate fusion weights based on target scale, local candidate density, intersection-union ratio, normalized center distance and validation probability, and to perform small target single-box protection for single candidate boxes; The results output module is used to summarize the fused detection boxes, sort them by fusion score, and give the final detection results according to the preset maximum number of single images to be output.

[0033] Preferably, in the difficult-to-bear sample secondary discrimination module, the training data of the secondary discriminator is automatically constructed from the target detection training set. Candidate regions with a maximum crossover union ratio (CUB) of not less than the matching threshold are marked as real device positive samples, and candidate regions with a CUB ratio less than the matching threshold are marked as background false detection difficult-to-bear samples.

[0034] Preferably, the dynamic distance-constrained weighted box fusion module is specifically used to dynamically calculate the cross-union ratio (CUN) clustering threshold and the center distance threshold by combining the small target degree, local density and candidate confidence of the candidate boxes to be matched, and to determine whether the candidate boxes belong to the same cluster by using the dual conditions of CUN and center distance.

[0035] Example 2 This invention provides a method for suppressing false detections of substation equipment based on dynamic distance constraint box fusion and hard-score sample soft validation. This method does not limit the specific target detection network; the detection model only needs to output candidate box coordinates, categories, and detection confidence scores. In one embodiment, a trained 15-class single-stage target detection network is used. The system consists of a candidate box generation module, a candidate region preprocessing module, a hard-score sample secondary discrimination module, a soft score correction module, a dynamic distance constraint weighted box fusion module, and a result output module. The overall system architecture of this embodiment is as follows: Figure 1 As shown. The main chain of the system includes candidate box generation, candidate region cropping and secondary discrimination, soft score correction, D-WD-WBF dynamic clustering and fusion, and result output. The secondary discriminator is constructed offline from training set images and labeled data, and outputs the verification probability of candidate regions during the inference stage.

[0036] (1) The candidate box generation module reads the substation image and outputs the k-th candidate box B_k, the category label l_k, and the detection confidence.

[0037] (2) The candidate region cropping module crops a local image from the original image according to the candidate box, and then uniformly scales and normalizes it.

[0038] (3) The secondary discrimination module outputs the secondary verification probability that the candidate region belongs to the real device target.

[0039] (4) The soft score correction module uses the secondary verification probability as a continuous adjustment factor to reduce the weight and rearrange the candidate box scores, instead of directly deleting the candidate boxes at this stage.

[0040] (5) The D-WD-WBF module dynamically clusters and merges clusters based on target scale, local density, IoU, center distance and validation probability.

[0041] (6) The result output module sorts the results by fusion score and gives the final detection results according to the preset maximum number of single images.

[0042] After the input image is inferred by the object detection model, candidate boxes with a detection confidence level not lower than the minimum confidence threshold of the candidate box are retained. In one embodiment, this threshold is set to 0.20, and the candidate box pre-retention overlap threshold is set to 0.99. A maximum of 1000 original candidate boxes are retained per image to avoid premature deletion of candidates needed for subsequent fusion by the duplicate box suppression process within the detection model. Each candidate box is represented by the coordinates of its upper left and lower right corners, and corresponds one-to-one with the category label and detection confidence level.

[0043] During candidate region preprocessing, candidate box coordinates are restricted to within the image boundaries, and invalid boxes with widths or heights not greater than zero are discarded. Valid candidate regions are converted to a uniform color space, scaled to the input size of the secondary discriminator, and normalized. When candidate region extraction fails, the corresponding secondary verification probability is set to zero and the exception is recorded to avoid inference interruption caused by invalid coordinates.

[0044] The training process of the second-level discriminator is as follows: Figure 3 As shown. The baseline detection model first generates candidate boxes on the training set images, then divides positive samples and background false positives into hard negative samples based on the intersection-union ratio (IU) of the candidate boxes and the ground truth boxes of the same category. Subsequently, it uses these samples to train a secondary discriminator and outputs the validation probability. The training data for the secondary discriminator is automatically constructed from the object detection training set; the validation set does not participate in the training of the secondary discriminator. First, the baseline detection model generates candidate boxes on the training images, then matches each candidate box with the ground truth boxes of the same category using the IU: candidate regions with an IU of at least one matching threshold are marked as positive samples from the real device, while candidate regions with an IU of less than this threshold are marked as hard negative samples from background false positives. In one embodiment, the matching threshold is 0.5. This allows the secondary discriminator to directly learn the difference between the local appearance of the real device and the high-frequency background false positives of the detection model. The composition of the candidate region samples for the secondary discriminator is shown in Table 1.

[0045] Table 1. Composition of candidate region samples for the secondary discriminator

[0046] In one embodiment, the secondary discriminator employs a lightweight convolutional binary classification network. Taking MobileNetV3-small as an example, its final classification layer is set to output binary classification of real devices and background false detections. The best-performing model obtained during training appeared in the 6th training epoch, with a validation accuracy of 0.8427; the precision for the real device class was 0.8939 and the recall was 0.8624, while the precision for the background false detection class was 0.7546 and the recall was 0.8052. Since the recall for the real device class is not 1.0, a single high-threshold hard deletion is not used as the main solution during the inference stage.

[0047] 4. Soft validation mechanism for candidate boxes For the k-th candidate box, let the detection confidence output by the detection model be C_{det,k}, and the true device verification probability output by the secondary discriminator be P_{v,k}. Let the soft retention coefficient be α, and 0≤α≤1, then the corrected candidate score C'_k is determined according to equation (1).

[0048] (1) When α=0, the corrected score is entirely determined by the product of the detection confidence and the secondary validation probability; when α=1, no secondary validation probability correction is used; the larger α is, the stronger the protection of true targets and the weaker the suppression of false detections. The main implementation uses α=0.50, ensuring that each candidate box retains at least 50% of its original detection score, while the secondary validation probability continuously reduces the weight and rearranges the candidate boxes. This mechanism does not directly delete candidate boxes in the secondary discrimination stage, mitigating false filtering caused by small targets, occluded targets, or incomplete local features.

[0049] 5. D-WD-WBF Dynamic Distance Constraint Box Fusion The D-WD-WBF dynamic clustering and fusion process is as follows: Figure 2 As shown, candidate boxes are grouped by category and sorted by continuously corrected scores. The dynamic intersection-union threshold and dynamic center distance threshold are calculated sequentially. Candidate boxes that satisfy both constraints are added to the existing cluster; otherwise, a new cluster is created. The representative box and cluster scores are then updated, and small target single-box protection is performed before the results are output.

[0050] A. Target scale factor and local scene density Let the area of ​​the k-th candidate box be A_k, the area of ​​the input image be A_I, and the smallness of the candidate box be s_k. Using the center of the candidate box as the center and three times the length of the candidate box's diagonal as the neighborhood radius, count the number of candidate boxes n_k in the neighborhood, and let N_0 represent the density normalization constant. The smallness of the target and the local density c_k are determined according to equation (2). The closer s_k is to 1, the smaller the candidate target is relative to the image; the larger c_k is, the denser the local candidate boxes are.

[0051] (2) B. Dynamic IoU threshold and dynamic center distance threshold For candidate boxes to be matched, the clustering threshold is dynamically calculated according to Equation (3) and Equation (4) by combining the small target degree, local density and candidate confidence.

[0052] (3) (4) The default values ​​for the basic intersection-union ratio (IoU,0) threshold T_{IoU,0} are 0.55 and the default values ​​for the basic normalized center distance threshold T_{dist,0} are 1.00. The default values ​​for the scale adjustment coefficient λ_s, density adjustment coefficient λ_d, and confidence adjustment coefficient λ_p are 0.25, 0.20, and 0.20, respectively, and the default value for the IoU density adjustment coefficient λ_c is 0.10. , Take values ​​of 0.25 and 0.75 respectively. , The values ​​are 0.35 and 1.80 respectively. A candidate box is only included in the current candidate box cluster when it simultaneously satisfies the overlap constraint and center distance constraint of equation (5).

[0053] (5) C. Dynamic Quality Weights and Dependency Box Updates Let r_k be the intersection-union ratio between the candidate box and the current cluster representative box, d_k be the normalized center distance, and q_k be the center consistency. In the implementation where the verification probability participates in the fusion weight, the fusion weight w_k is determined according to equation (7), so that the soft correction score, geometric overlap relationship, center consistency and semantic authenticity jointly affect the fusion result.

[0054] (6) (7) (8) Where ε is a very small positive number to prevent the weight from being zero, and m is the number of candidate boxes in the current cluster. Whenever a new candidate box is added to the cluster, the weights are immediately recalculated for all candidate boxes in the current cluster, and the representative box coordinates and cluster scores are updated according to equation (8). As an alternative implementation, different weight indices can be used according to the area range of the candidate boxes: increase the confidence weight and decrease the geometric relationship weight for small targets, increase the overlap relationship and center consistency weight for large targets, and use a balanced weight for medium targets. The reasoning process of soft verification and dynamic fusion is as follows Figure 4 As shown, the detection confidence and secondary verification probability first form a continuously corrected score, which is used for candidate box sorting and dynamic fusion; the secondary verification probability can also participate in the fusion weight calculation. Figure 4 In this context, r_k represents the intersection-union ratio of the candidate box and the current cluster representative box, and q_k represents the central consistency obtained from the normalized central distance.

[0055] D. Dynamic confidence and small target frame protection Dynamic distance-constrained weighted bounding box fusion processes each category independently. When the candidate box area is less than the small target area ratio threshold, a small target retention threshold is used; otherwise, a basic retention threshold is used. In one embodiment, both retention thresholds are 0.20, and the small target area ratio threshold is 0.0025. After clustering, clusters containing two or more candidate boxes are directly output; single-box clusters are also retained as long as they reach the retention threshold corresponding to their scale, thus avoiding the accidental deletion of dense small targets due to a lack of duplicate candidate boxes.

[0056] (9) 6. Output, Deployment, and Exception Handling The fused results are sorted from high to low according to the fusion confidence score, and output according to the preset maximum number of single images; in the main embodiment, this number is 300. The output includes at least image identifier, category, candidate box coordinates, and fusion confidence score, and optionally records secondary verification probability, cluster member number, dynamic threshold, and scene density for inspection and verification and parameter backtracking.

[0057] The object detection model and secondary discriminator can be deployed on the same computing device, different computing devices, edge devices, and servers, or processed in batches in a pipeline manner. Candidate regions can be input into the secondary discriminator in batches; when image reading fails, candidate boxes are invalid, or the secondary discriminator is unavailable, the system records the anomaly and degrades to dynamic distance-constrained weighted box fusion using only detection confidence, without affecting the basic detection chain.

[0058] Example 3: Master Scheme Parameters and Complete Reasoning Steps The parameters used in the main embodiment and their functions are shown in Table 2.

[0059] Table 2 Parameters and Functions of the Main Implementation Example

[0060] (1) Read the substation image, scale the image to 640×640 pixels and input it into the trained fifteen-class target detection model to obtain candidate boxes, categories and detection confidence.

[0061] (2) Extract candidate regions one by one, and input them in batches into a lightweight convolutional secondary discriminator to obtain the secondary verification probability of each candidate box.

[0062] (3) Calculate the corrected score according to formula (1), where the soft retention coefficient is 0.50, and the candidate box is kept in one-to-one correspondence with the secondary verification probability.

[0063] (4) Group by category and process in descending order of corrected score, calculate the small target degree, local density, dynamic intersection-union threshold and dynamic center distance threshold for each candidate box.

[0064] (5) Compare with existing cluster representative boxes in turn; add a cluster if both IoU and center distance conditions are met, otherwise create a new cluster.

[0065] (6) The representative box and cluster scores are updated by using the second-level verification probability to participate in the fusion weight, and small target single box protection is implemented.

[0066] (7) Summarize the results of each category cluster, sort by fusion score and output the top 300 detection boxes.

[0067] Example 4: Control Experiment The comparison results of each post-processing and validation scheme on the original validation set are shown in Table 3.

[0068] Table 3. Results of the original validation set control experiment

[0069] The hard filtering scheme with a fixed threshold of 0.60 reduced the number of false positives to 752, but the recall rate dropped to 0.4210, indicating that direct hard deletion was too aggressive. The main scheme controlled the number of false positives to 2852 and restored the recall rate to 0.5256, while achieving the highest F1 score of 0.6335, proving that soft validation does not simply relax the threshold, but rather strikes a balance between semantic weighting of candidate boxes and dynamic geometric fusion.

[0070] Example 5: Subset of Complex Scenes Table 4 shows the changes in false detection suppression and recall of the main solution across various complex scenario subsets.

[0071] Table 4 Experimental Results for Subsets of Complex Scenarios

[0072] The main solution demonstrates significant false positive suppression in complex backgrounds with high false positives, complex backgrounds, and dense small target subsets. Recall decreases by more than 10% in strongly occluded subsets; therefore, larger α values, class-adaptive protection, or occluded target protection thresholds can be used in this scenario, and it should not be claimed that recall remains constant across all scenarios. A comparison of detection results in complex scenarios with high false positives is provided. Figure 5 As shown, the green boxes represent true labeled boxes, the blue boxes represent correctly detected boxes, the red boxes represent falsely detected boxes, and the yellow boxes represent missed detection boxes. Figure 5 The detection results of D-WD-WBF, the combination of hard filtering with a fixed threshold of 0.60 and D-WD-WBF, and the combination of soft validation and D-WD-WBF are shown in turn to visually illustrate the differences between the different schemes in terms of false detection suppression and recall maintenance.

[0073] Example 6: Alternative Implementation (1) The target detection model can be replaced by other single-stage, two-stage or Transformer detectors, as long as they can output bounding boxes, categories and confidence scores.

[0074] (2) The secondary discriminator can be replaced by ResNet, lightweight visual Transformer or other binary classification networks, and can also output the real probability of multiple categories.

[0075] (3) Hard filtering mode can be used in scenarios where the cost of false detection is significantly higher than the cost of false detection; small target protection mode can use a lower verification threshold for candidates with an area ratio of less than 0.003; category adaptive mode can set different thresholds for high-frequency false detection categories.

[0076] (4) The soft retention coefficient can be adjusted in the range of 0 to 1 according to the scenario; for example, when it is 0.30, it is biased towards false detection suppression, and when it is 0.70, it is biased towards recall protection.

[0077] (5) Local density can be obtained by fixed radius counting, K-nearest neighbor distance, candidate box heatmap or scene classification results; the dynamic threshold function can also be a piecewise function or a learning function.

[0078] (6) The secondary verification probability can be used only for soft score correction, only for fusion weight, or both; the main implementation adopts the method of simultaneous participation.

[0079] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for suppressing false detections of substation equipment, characterized in that, Includes the following steps: The object detection model generates candidate bounding boxes for substation images and crops the candidate regions, obtaining the category and detection confidence of the candidate bounding boxes; The candidate region is input into the secondary discriminator, which outputs the verification probability that the candidate region belongs to the real device. The secondary discriminator is trained using hard negative samples automatically constructed from similar real bounding boxes and background false detection boxes. The detection confidence and validation probability are combined into a continuously corrected score using a soft retention coefficient, without directly deleting candidate boxes; The candidate box clustering threshold is dynamically determined based on the target scale, local candidate density, and continuous correction score, and the fusion weight is calculated based on the continuous correction score, intersection-union ratio, center consistency, and verification probability. Single candidate boxes with an area ratio below a threshold are preserved and protected, and the fusion detection results are output.

2. The method for suppressing false detections of substation equipment according to claim 1, characterized in that, The steps for constructing training data for the secondary discriminator include: The baseline detection model is used to generate training candidate boxes for the training images, and each training candidate box is matched with the ground truth labeled box of the same category by intersection-union matching. Training candidate regions with a maximum crossover-union ratio (CUI) not less than the matching threshold are labeled as positive samples from the real device, while training candidate regions with a maximum CUI less than the matching threshold are labeled as negative samples from the background that are difficult to detect false positives.

3. The method for suppressing false detections of substation equipment according to claim 1, characterized in that, The calculation steps for the continuously corrected score include: Let the detection confidence level be... The verification probability is The soft retention factor is ,and Continuous score correction satisfy: .

4. The method for suppressing false detections of substation equipment according to claim 1, characterized in that, The calculation steps for the target scale and local candidate density include: Let the area of ​​the candidate box be The area of ​​the input image is The target scale is at the level of a small target. ,satisfy ; Using the center of the candidate box as the center and three times the length of the candidate box's diagonal as the radius, count the number of candidate boxes within the neighborhood. ,by Represents the density normalization constant, local candidate density satisfy .

5. The method for suppressing false detections of substation equipment according to claim 4, characterized in that, The steps for dynamically determining the candidate box clustering threshold include: The cross-union ratio threshold and the normalized center distance threshold are dynamically calculated by combining the small target size, local candidate density, and detection confidence. If the intersection-union ratio (IU) between the candidate box to be matched and the representative box of the existing cluster is not less than the IU threshold, and the normalized center distance is not greater than the normalized center distance threshold, the candidate box to be matched is assigned to the existing cluster; otherwise, a new cluster is created.

6. The method for suppressing false detections of substation equipment according to claim 1, characterized in that, The steps for calculating the fusion weights include: Whenever a new candidate box is added to a cluster, the weights are recalculated for all candidate boxes within the cluster. The fusion weight is calculated by continuously adjusted scores, crossover ratio, central consistency and validation probability, and the bounding box coordinates and cluster scores are updated based on the fusion weight.

7. The method for suppressing false detections of substation equipment according to claim 1, characterized in that, The steps to perform retention protection include: Determine whether the area of ​​the candidate box relative to the image area is less than the threshold for the proportion of small target area. If yes, use the small target retention threshold; otherwise, use the basic retention threshold. After clustering is completed, clusters containing two or more candidate boxes are output directly, while single-box clusters containing only a single candidate box are retained and output when they reach the corresponding retention threshold.

8. A substation equipment false detection suppression system, characterized in that, include: The candidate box generation module is used to read substation images and output the first... The coordinates, category labels, and detection confidence of each candidate bounding box; The candidate region preprocessing module is used to crop a local image from the original image according to the candidate box, and perform uniform scaling and normalization to obtain the candidate region; The difficult-to-bear sample secondary discrimination module is used to input candidate regions into the secondary discriminator and output the secondary verification probability that the candidate regions belong to the real device targets; The soft score correction module is used to use the secondary validation probability as a continuous adjustment factor, and combine the soft retention coefficient with the detection confidence to form a continuous correction score. This module reduces the weight and rearranges the candidate boxes, but does not directly delete the candidate boxes. The dynamic distance-constrained weighted box fusion module is used to dynamically cluster and calculate fusion weights based on target scale, local candidate density, intersection-union ratio, normalized center distance and validation probability, and to perform small target single-box protection for single candidate boxes; The results output module is used to summarize the fused detection boxes, sort them by fusion score, and give the final detection results according to the preset maximum number of single images to be output.

9. A substation equipment false detection suppression system according to claim 8, characterized in that: In the secondary discrimination module for hard negative samples, the training data of the secondary discriminator is automatically constructed from the target detection training set. Candidate regions with a maximum crossover ratio (CUP) of not less than the matching threshold are marked as real device positive samples, while candidate regions with a CUP of less than the matching threshold are marked as background false positive hard negative samples.

10. A substation equipment false detection suppression system according to claim 8, characterized in that: The dynamic distance-constrained weighted box fusion module is specifically used to dynamically calculate the cross-union ratio (CUI) clustering threshold and the center distance threshold by combining the small target degree, local density and candidate confidence of the candidate boxes to be matched. It uses the dual conditions of CUI and center distance to determine whether the candidate boxes belong to the same cluster.