On-site small target defect instant screening method and device for power transmission line inspection and storage medium

Abstract

The present application relates to the technical field of power transmission line inspection and edge intelligent identification, and discloses a kind of on-site small target defect instant screening method, device and storage medium for power transmission line inspection.The method carries out effectiveness verification and standardization pretreatment after portable terminal receives unmanned aerial vehicle inspection image, carries out preliminary inspection using lightweight defect detection model, and obtains candidate defect target;For the candidate target that meets the small target trigger condition, generate extended local area, and carry out cutting, amplification, normalization and secondary rejudgment;Then fuse the preliminary inspection result and the rejudgment result, determine the serious defect target in combination with the severity score, generate the on-site screening report, and selectively return the image and structured result corresponding to the serious defect to the background.The method can improve the on-site screening efficiency of small target defects in power transmission line inspection and reduce the data return pressure.

Description

Technical Field

[0001] This invention relates to the fields of power transmission line inspection, edge computing, and computer vision, specifically to a method, device, and storage medium for real-time screening of small target defects in power transmission line inspection. Background Technology

[0002] Unmanned aerial vehicle (UAV) inspections have become an important means of refined operation and maintenance of power transmission lines. A single field operation typically generates thousands of high-resolution inspection images. Current engineering practices mostly adopt a processing mode of "field acquisition—full data transmission—centralized backend identification—result transmission." While this mode facilitates unified backend management, it still has the following shortcomings in the field inspection environment: 1. Long transmission time and low on-site work efficiency. The large number of inspection images and high resolution of each image mean that full transmission is time-consuming in weak network or intermittent network environments. This often results in a long wait before the analysis phase can begin after the on-site work is completed, seriously affecting the overall efficiency of the inspection operation.

[0003] Second, the ability to detect defects in small targets is insufficient, leading to a high risk of missed detections. Defects such as missing pins and minor damage to hardware occupy a very small percentage of pixels in the entire inspection image (usually less than 0.05% of the total image area), with strong background interference and large scale variations. General defect detection models have limited stability in recognizing such targets, easily resulting in missed detections or low confidence levels. Furthermore, on-site personnel cannot immediately know the location and type of defects after taking the picture, making it difficult to decide in a timely manner whether to retake the picture, re-fly, or conduct on-site verification.

[0004] Third, delayed backend results hinder real-time decision-making support on-site. The identification task relies heavily on the backend server, and there is a significant delay window between image transmission and obtaining analysis results. On-site personnel cannot obtain alarm information for serious defects in real time during inspections, and therefore cannot make critical decisions on-site such as "whether immediate review is needed" or "whether re-photographing is needed."

[0005] Fourth, full data transmission continuously consumes transmission bandwidth, storage space, and backend computing power. The indiscriminate transmission of a large number of images with no defects or only minor defects not only wastes communication resources but also increases the pressure on backend storage and analysis, which is not conducive to the efficient organization of large-scale inspection operations.

[0006] Therefore, there is an urgent need for an instant screening method that can run directly on resource-constrained portable terminals in the field and can perform targeted re-judgment of small target defects, so as to quickly provide preliminary results at the inspection site and reduce unnecessary data backhaul. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method, device, and storage medium for real-time screening of small target defects in power transmission line inspections. This real-time screening method can be run directly on resource-constrained portable terminals in the field and can perform specific re-judgments of small target defects, so as to quickly provide preliminary results at the inspection site while reducing unnecessary data transmission.

[0008] This invention is achieved through the following technical solution: A method for real-time screening of small-target defects in power transmission line inspection is provided, which is applied to portable terminals with edge computing capabilities. The method includes the following steps: Receive images from drone inspections and perform validity verification and standardized preprocessing on the images; The lightweight defect detection model deployed on the portable terminal is invoked to perform the first detection on the preprocessed inspection image, and the bounding box, candidate defect category and first confidence score of each candidate defect target are obtained. For candidate defect targets that meet the small target triggering conditions, an extended local region is generated by expanding the bounding box of the candidate defect target outward by a preset ratio. After cropping, enlarging and normalizing the size of the extended local region, it is input into the small target enhancement recognition model for secondary re-judgment to obtain the second recognition result and the second confidence level. The small target enhancement recognition model includes a local feature enhancement subnetwork and a confidence level calibration subnetwork. The local feature enhancement subnetwork includes a lightweight spatial attention submodule. The first detection result and the second re-judgment result are fused according to the preset fusion rules to obtain the final defect category and the final confidence level; Based on the final defect category, final confidence level, and relative size information of the defect target, the severity score S = 0.4 × f is calculated according to the preset severity scoring rules. type +0.4×f conf +0.2×f size , where f type To score the defect type, f conf For the final confidence score, f size Scoring for defect size; Defect targets with S≥0.70 are identified as critical defect targets, an on-site screening report is generated, and the original image, partially cropped image and structured results corresponding to the critical defect targets are selectively sent back to the back-end system. The triggering conditions for small targets include at least one of the following: The bounding box area of ​​the candidate defect target is less than 0.05% of the total image area; The candidate defect target belongs to a preset set of sub-target defect categories, and the set of sub-target defect categories includes at least missing pins; The first confidence level of the candidate defect target is in the range of 0.50 to 0.80.

[0009] Furthermore, the preset proportion for expanding outwards in all directions is 20% for each side of the bounding box, and the target size for size normalization is 128×128 pixels.

[0010] Furthermore, the integration rules specifically include: When the defect category detected in the first detection matches the defect category identified in the second identification result, the defect category is retained, and the larger of the first confidence score and the second confidence score is taken as the final confidence score. When the two defect categories are inconsistent and the second confidence level is not lower than 0.80, the second identification result is used as the final defect category, and the second confidence level is used as the final confidence level. When the two defect categories are inconsistent and the second confidence level is below 0.80, the first detection result is retained as the final defect category, and a manual review mark is added to the target.

[0011] Furthermore, in the severity rating, for defects such as missing pins, ftype is set to 1.0; fconf is directly taken as the final confidence level; and fsize is calculated based on the normalized ratio of the defect area to the area of ​​the component in which it is located.

[0012] Furthermore, it also includes a tiered processing strategy for defective targets: If S ≥ 0.70, it is judged as a serious defect, and a key alarm is generated and sent back immediately; If 0.50≤S<0.70, it is judged as a defect of concern, cached locally and returned in batches when the network is available or made available for manual review; If S < 0.50, it is judged as a general defect, and only written to the local inspection record or a simplified report is generated.

[0013] Furthermore, the lightweight defect detection model is an improved YOLOv5s model quantized with INT8, with a model file size of no more than 28MB and an input size of 640×640 pixels. Inference is performed on a portable terminal using the OpenVINO inference engine.

[0014] The on-site screening report should include at least the work batch information, image number, tower number, defect category, defect location, severity level, final confidence level, and recommended treatment items; for targets that do not reach the severity threshold, the results information is only cached locally on the portable terminal without triggering immediate data transmission.

[0015] The present invention also provides an apparatus for real-time screening of small target defects in power transmission line inspection, deployed on a portable terminal with edge computing capabilities, comprising: The image receiving and preprocessing unit is used to receive UAV inspection images and perform validity verification and standardization preprocessing. The initial inspection unit is used to call the lightweight defect detection model to perform the first detection on the preprocessed inspection image, and output the bounding box of the candidate defect target, the candidate defect category and the first confidence score; The small target re-judgment unit is used to generate an extended local region by expanding outward from the bounding box based on the candidate defect target that meets the small target triggering conditions. After being cropped, enlarged, and normalized in size, the extended local region is input into the small target enhancement recognition model for secondary re-judgment to obtain a second recognition result and a second confidence level. The small target enhancement recognition model includes a local feature enhancement sub-network with a lightweight spatial attention sub-module and a confidence calibration sub-network with a fully connected layer. The result fusion unit is used to fuse the first detection result and the second re-judgment result according to the preset fusion rules to obtain the final defect category and the final confidence level. The severity assessment unit is used to calculate the severity score based on the final defect category, final confidence level, and relative defect size, and to determine those with S≥0.70 as severe defects. The report generation and feedback unit is used to generate on-site screening reports and feed back the original images, partially cropped images, and structured results corresponding to the severely defective targets to the backend system.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method.

[0017] The beneficial effects of this invention are: This invention brings defect identification capabilities to the inspection site, allowing for initial inspection and key re-judgment to be completed after the inspection images are imported. This reduces waiting time caused by full image transmission and background queuing analysis, and improves the timeliness of obtaining preliminary results on site.

[0018] By implementing local cropping and magnification, along with secondary enhancement recognition steps, for small target defects, the system improves the identification capability of key targets such as missing pins and minor damage, reducing the risk of missed detections, without recalculating the entire image. In particular, by setting a hesitancy zone of 0.50 to 0.80 confidence level as one of the triggering conditions, the system precisely allocates computational resources to candidate targets where the initial detection model lacks confidence but still has a high probability of being true positives, achieving optimal allocation of computational resources.

[0019] Based on the severity determination rules, the identification results are filtered and classified. On-site operators can view the list, thumbnails, enlarged partial images and highlighted boxes of serious defects on the terminal interface in real time, and decide whether to review, retake or report them, thus filling the gap in on-site real-time decision support.

[0020] This invention only transmits back images and structured information corresponding to severely defective targets, thereby reducing network bandwidth usage and backend storage pressure, and forming a closed-loop processing link of on-site initial inspection, local re-judgment, severe screening, and selective transmission.

[0021] Furthermore, this invention employs an improved YOLOv5s target detection model quantized with INT8 as the primary detection model, and performs inference on a portable terminal using the OpenVINO inference engine. Combined with a unified input of 640×640 pixels and a lightweight deployment method with a model file size not exceeding 28MB, the terminal can complete preprocessing of inspection images, primary detection, small target re-judgment, severity determination, and screening report generation on-site, thereby reducing reliance on centralized computing power in the background. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0023] Figure 2 This is a schematic diagram illustrating the data interaction between the field terminal, the drone, and the back-end system of this invention.

[0024] Figure 3 This is a schematic diagram of the secondary re-judgment process for small target candidate regions in this invention.

[0025] Figure 4 This is a schematic diagram of the severity determination and selective feedback process in this invention. Detailed Implementation

[0026] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution. Example

[0027] like Figure 1 As shown, the on-site small target defect real-time screening method of the present invention mainly includes the following processing stages: inspection image reception and batch association, image validity verification and standardized preprocessing, initial inspection of the terminal side lightweight model, small target candidate region screening, local cropping, magnification and enhancement re-judgment, fusion of initial inspection results and re-judgment results, severity determination, on-site screening report generation and selective back transmission of severely defective targets. Figure 1 This invention demonstrates the overall closed-loop processing flow consisting of on-site import, real-time identification, key re-judgment, hierarchical screening, and on-demand data transmission.

[0028] In this embodiment, the drone collects inspection images of towers, insulator strings, hardware connection points, conductors, and accessories during power transmission line inspections. After the inspection, personnel receive the inspection images via an industrial tablet, ruggedized portable device, or other portable terminal with edge computing capabilities. Image reception methods include direct USB data cable connection, direct wireless LAN transmission, memory card import, or import from a mobile storage device. Simultaneously with image import, the terminal acquires or establishes work batch information corresponding to the inspection images and associates and stores the image number with information such as tower number, shooting time, and flight path segment number.

[0029] like Figure 2 As shown, a drone is used to collect images of power transmission line inspections and transmits them to a portable terminal with edge computing capabilities. The portable terminal performs image preprocessing, initial defect detection, small target re-judgment, severity screening, and on-site report generation locally. The backend system receives images of severe defects, partially cropped images, and / or structured defect information filtered by the terminal according to rules, for subsequent archiving, review, work order linkage, or integration with the maintenance platform. This invention does not transmit all inspection images back in full; instead, it selectively uploads images after initial screening at the terminal, thereby reducing communication burden and improving on-site processing timeliness.

[0030] After receiving the inspection image, the terminal first performs validity verification and standardization preprocessing. Validity verification includes, but is not limited to, file format verification, integrity verification, image decoding verification, and duplicate file identification. Images that fail verification (such as corrupted files or duplicate images) are removed from the identification queue, and the reason for the anomaly is recorded in the log. Standardization preprocessing includes: orientation correction, resolution normalization (uniform scaling to the model input size), pixel value normalization (e.g., dividing by 255), and color channel unification (e.g., RGB order confirmation).

[0031] After preprocessing, the system calls the lightweight defect detection model deployed on the terminal to perform the first detection. The first detection outputs the bounding box position, candidate category, and first confidence score of each candidate defect target. For candidate defect targets that do not trigger a second review, the system directly retains the results of the first detection as input for subsequent severity determination; for candidate defect targets that meet the triggering conditions, the system enters the local review process. Example

[0032] like Figure 3As shown, after initial detection of candidate defect targets on the terminal side, the system first filters targets to be re-judged based on preset small target triggering conditions. For candidate targets that meet the re-judgment conditions, the system generates an extended local region based on its bounding box and sequentially performs local cropping, enlargement, size normalization, and detail enhancement processing. Then, it calls the small target enhancement recognition model to output a second recognition result and a second confidence score. Finally, the system performs consistency correction, confidence score fusion, or result replacement on the initial detection result and the re-judgment result according to preset fusion rules. This invention provides a local re-judgment mechanism for small defect targets to improve the reliability of small target recognition while controlling terminal computing power consumption.

[0033] In this embodiment, the lightweight defect detection model adopts an improved YOLOv5s target detection model quantized with INT8, with a model file size of approximately 28MB; the inspection images are uniformly scaled to 640×640 pixels before entering the first inspection. The terminal calls the OpenVINO inference engine to perform model inference to meet the computing power and memory constraints of the field equipment.

[0034] After the initial detection, the system does not recalculate all candidate targets uniformly. Instead, it filters targets for re-evaluation based on preset small target triggering conditions. In this embodiment, the small target triggering conditions include: the area of ​​the candidate target is less than 0.05% of the total image area; the candidate target belongs to a preset small target defect category set (at least including missing pins); and the first confidence level of the candidate target is between 0.50 and 0.80. The preset small target defect category set includes at least missing pins.

[0035] For each target to be re-evaluated, the system expands outwards by 20% in each direction based on the corresponding bounding box, generating an extended local region. This extended local region is then cropped, enlarged, and normalized to meet the input requirements of the small target enhancement recognition model. In this embodiment, the local region is normalized to 128×128 pixels.

[0036] The small target enhancement recognition model includes a local feature enhancement subnetwork and a confidence calibration subnetwork. The local feature enhancement subnetwork includes a lightweight spatial attention submodule to enhance edge information and high-frequency texture information in the local image; the confidence calibration subnetwork includes fully connected layers to perform classification and confidence calibration on the enhanced local features to output a second recognition result and a second confidence level.

[0037] In the first detection, a candidate target is identified as a missing pin target with an initial confidence level of 0.65 and its bounding box area is lower than the area threshold. After local feature enhancement and confidence level calibration, the second confidence level of the candidate target is increased to 0.82.

[0038] During the results fusion phase, the system processes the data according to the following rules: When the defect category of the first detection result matches that of the second identification result, the system retains the defect category and uses the larger confidence level between the two as the final confidence level. When the defect categories of the two results are inconsistent and the second confidence level is not lower than 0.80, the system adopts the second identification result as the final defect category; When the two results are inconsistent and the second confidence level is below 0.80, the system retains the first test result and adds a manual verification mark.

[0039] The above processing can improve the reliability of identifying defects in small targets while controlling the overall inference overhead.

[0040] Table 1 below summarizes examples of key parameters in this embodiment: Table 1 Key Parameters First detection model Improved YOLOv5s Deployed on the terminal side after INT8 quantization First test of input dimensions 640×640 pixels Used to standardize model input specifications Model file size Approximately 28MB Meeting the deployment requirements of portable terminals Small target area trigger threshold 0.05% of the total area of ​​the map Local re-judgment is triggered when the value is below this threshold. Local expansion ratio 20% on each side of the bounding box Preserve the contextual information surrounding the defective target Re-judgment of input size 128×128 pixels Input for small target enhancement recognition model Example of raw confidence level 0.65 Confidence of candidate targets obtained from the first detection Example of confidence level for re-judgment 0.82 Second confidence level after local re-judgment Replacement threshold 0.80 Result replacement conditions when categories are inconsistent Example

[0041] like Figure 4 As shown, after obtaining the final defect category and corresponding confidence level, the system further combines factors such as defect type risk, target relative size, and location risk information to perform severity scoring or level determination on the defect target. For targets that reach the severity threshold, the terminal prioritizes generating key alarm information and immediately sends back the corresponding original image, partially cropped image, and / or structured results to the backend. For targets that do not meet the conditions for immediate feedback, local caching, delayed upload, or only writing to the local inspection record are performed. This invention achieves a selective feedback mechanism by prioritizing the feedback of key targets and processing ordinary targets as needed through severity grading.

[0042] After obtaining the final defect category and final confidence level, the system enters the severity determination stage. This stage does not rely solely on a single confidence threshold for screening; instead, it calculates a severity score for each defect target by combining the defect type, final confidence level, and defect size, thereby identifying critical defect targets that require immediate feedback.

[0043] In this embodiment, the severity score is calculated according to the following formula: S=0.4×f type + 0.4×f conf + 0.2×f size ; Among them, f type For defect type scoring, for missing pin defects, f type Set the score to 1.0; other defect types can be pre-set with a score based on the risk level (e.g., set hardware breakage to 0.8).

[0044] f confFor the final confidence score, directly take the final confidence score (value range 0~1).

[0045] f size The defect size score is calculated by normalizing the ratio of the defect area to the area of ​​the component it belongs to. A smaller ratio results in a lower score (indicating that smaller defects require more attention; however, a higher score here indicates a more severe defect). The specific mapping can be set by those skilled in the art; for example, when the defect area percentage is less than 0.1%, f... size =0.5, when the proportion is greater than 1%, f size =1.0).

[0046] The system classifies defect targets into three levels based on the S-value and adopts different processing strategies, as shown in Table 2 below: S ≥ 0.70 Critical defects Real-time generation of key alerts and transmission of original images, local images, and structured results. 0.50 ≤ S<0.70 Focus on defects Local caching allows for batch uploads or manual review when network availability. S<0.50 General defects Write only to local inspection logs or generate simplified reports If the final confidence level of a missing pin target after reassessment is 0.82 and the size score is 0.80, then its severity score is: S=0.4×1.0+0.4×0.82+0.2×0.80=0.4+0.328+0.16=0.888; The target was identified as a critical defect (S≥0.70), triggering an immediate feedback.

[0047] The terminal generates a field screening report based on the list of serious defects. The field screening report may include work batch information, image number, tower number, defect category, defect location, severity level, final confidence level, and recommended actions. For targets that reach the severity threshold, the system displays them on the terminal interface using lists, thumbnails, zoomed-in views, and highlighted boxes, allowing operators to quickly decide whether to conduct a review or retake.

[0048] To reduce the pressure on on-site data transmission, the system does not need to upload all inspection images to the backend. Instead, it only sends back the original images, partially cropped images, and structured results corresponding to severely defective targets. The backend system can then further perform archiving, manual review, work order linkage, or data integration with the operation and maintenance platform. This forms a closed-loop processing chain of on-site initial inspection, partial reassessment, severe screening, and selective data return. Example

[0049] To more clearly demonstrate the complete working process of the method of the present invention, a specific inspection image is used as an example to give a full-process numerical example from image reception to selective backhaul.

[0050] Image reception and preprocessing: During a routine inspection, a drone captured an image of insulator strings and fittings at a resolution of 4000×3000 pixels. The terminal received the image and verified its validity (JPEG format, normal integrity, no duplicates). Standardization preprocessing scaled the image to 640×640 pixels and normalized the pixel values.

[0051] First detection: Using an improved YOLOv5s model with INT8 quantization (input 640×640, model 28MB, OpenVINO inference). Output: Two candidate defect targets. Candidate Target A: Category "Pin Missing", bounding box (120, 150, 30, 25), area 750 square pixels, total image area 409,600 square pixels, area percentage 0.183% (higher than 0.05%), first confidence level 0.72.

[0052] Candidate Target B: Category "Pin Missing", bounding box (400, 500, 18, 15), area 270 square pixels, area percentage 0.066% (higher than 0.05%), first confidence level 0.62.

[0053] Small target trigger condition judgment: Candidate Target A: Area percentage 0.183% > 0.05%, does not meet the area requirement; Category belongs to the preset small target category (pin missing); First confidence level 0.72 is within the range of 0.50~0.80. Triggering condition (confidence condition) is met, proceed to re-evaluation.

[0054] Candidate Target B: Area percentage 0.066% > 0.05%, does not meet the area condition; Category belongs to the preset small target category; First confidence level 0.62 is within the range of 0.50~0.80. The trigger condition (confidence condition) is met, proceed to re-evaluation.

[0055] Partial cropping and re-judgment: Candidate target A: bounding box (120, 150, 30, 25), expanded by 20% to obtain an expanded region: (108, 138, 54, 45). This region is cropped and enlarged to 128×128 pixels, then input into a small target enhancement recognition model (including spatial attention + fully connected calibration). The output is a second recognition result: category "missing pin," with a second confidence score of 0.88.

[0056] Candidate target B: bounding box (400, 500, 18, 15), expanded by 20% to obtain the expanded region: (388, 488, 42, 33). Crop and enlarge to 128×128, input into the re-judgment model. Output the second recognition result: category "missing pin", second confidence level 0.78.

[0057] Results fusion: Candidate Target A: The initial detection category "pin missing" is consistent with the re-judgment category, and the final category is "pin missing". The final confidence level is max(0.72, 0.88) = 0.88.

[0058] Candidate Target B: The initial detection category "pin missing" is consistent with the re-judgment category, and the final category is "pin missing". The final confidence level is max(0.62, 0.78) = 0.78.

[0059] Severity rating: Candidate target A: f type =1.0 (pin missing), f conf =0.88, where f size The defect area is 750 square pixels. The typical area of ​​the component (pin) in the image is 10,000 square pixels (estimated), accounting for 7.5%. Normalized f size =0.90 (assuming the mapping function is linear). Then S = 0.4 × 1.0 + 0.4 × 0.88 + 0.2 × 0.90 = 0.4 + 0.352 + 0.18 = 0.932. S ≥ 0.70, which is considered a serious defect.

[0060] Candidate target B: f type =1.0, f conf =0.78, f size The defect area is 270 square pixels, accounting for 2.7% of the total area, with a normalized f-value. size =0.60. Then S = 0.4 × 1.0 + 0.4 × 0.78 + 0.2 × 0.60 = 0.4 + 0.312 + 0.12 = 0.832. S ≥ 0.70, which is judged as a serious defect.

[0061] Report generation and selective return: The terminal generates an on-site screening report, displaying two severely defective targets. The terminal sends back the original images (4000×3000 original images), partially cropped images (expanded region images), and structured results (category, coordinates, confidence level, score) corresponding to candidate targets A and B to the backend. The backend receives and archives the data, triggering a work order. On-site personnel see the highlighted marks on the terminal interface and decide to re-photograph the locations of these two pins for confirmation.

[0062] The above end-to-end example demonstrates the complete processing flow of the method of the present invention in a real-world inspection scenario. Example

[0063] The method of this invention is deployed on a portable terminal device with edge inference capabilities. This portable terminal device can be an industrial tablet, a ruggedized portable device, or other mobile computing device with a processor, memory, and display module. The memory stores inspection image data, a lightweight defect detection model file, a small target enhancement recognition model file, and program instructions for executing the method of this invention. The processor calls the program instructions to complete image preprocessing, initial defect detection, small target re-judgment, result fusion, severity determination, report generation, and selective data transmission.

[0064] Specifically, the portable terminal integrates the following functional modules: Image receiving and preprocessing unit: Receives inspection images imported by the UAV and calls the image processing library to complete file format verification, integrity verification, orientation correction, size normalization, pixel value normalization, and image index establishment.

[0065] Initial inspection unit: Call the lightweight defect detection model (INT8 quantization improved YOLOv5s, input 640×640, model ≤28MB, inference via OpenVINO) to perform the first inspection, and output the defect category, bounding box and first confidence score of the candidate defect target.

[0066] Small target re-judgment unit: performs local pruning, outward expansion by 20%, and normalization to 128×128 on candidate targets that meet the triggering conditions, and feeds them into the small target enhancement recognition model (including spatial attention submodule and fully connected calibration subnetwork), and outputs the second recognition result and the second confidence score.

[0067] The result fusion unit obtains the final defect category and final confidence level according to the fusion rules described in Example 2.

[0068] Severity determination unit: Based on the final defect category, final confidence level, and defect size information, according to the formula S=0.4f type +0.4f conf +0.2f size Calculate the severity score and classify it according to Table 2.

[0069] Report generation and feedback unit: Generates on-site screening reports based on the list of severely defective targets, and only feeds back the original images, partially cropped images, and structured results corresponding to the severely defective targets to the backend system.

[0070] Image data and intermediate results are transmitted between units through internal data streams within the terminal. The entire processing is completed locally on the portable terminal, without relying on the real-time computing capabilities of the backend server.

[0071] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A method for real-time screening of small-target defects in power transmission line inspection, characterized in that: Applied to portable terminals with edge computing capabilities, the method includes the following steps: Receive images from drone inspections and perform validity verification and standardized preprocessing on the images; The lightweight defect detection model deployed on the portable terminal is invoked to perform the first detection on the preprocessed inspection image, and the bounding box, candidate defect category and first confidence score of each candidate defect target are obtained. For candidate defect targets that meet the small target triggering conditions, an extended local region is generated by expanding the bounding box of the candidate defect target outward by a preset ratio. After cropping, enlarging and normalizing the size of the extended local region, it is input into the small target enhancement recognition model for secondary re-judgment to obtain the second recognition result and the second confidence level. The small target enhancement recognition model includes a local feature enhancement subnetwork and a confidence level calibration subnetwork. The local feature enhancement subnetwork includes a lightweight spatial attention submodule. The first detection result and the second re-judgment result are fused according to the preset fusion rules to obtain the final defect category and the final confidence level; Based on the final defect category, final confidence level, and relative size information of the defect target, the severity score S = 0.4 × f is calculated according to the preset severity scoring rules. type +0.4×f conf +0.2×f size , where f type To score the defect type, f conf For the final confidence score, f size Scoring for defect size; Defect targets with S≥0.70 are identified as critical defect targets, an on-site screening report is generated, and the original image, partially cropped image and structured results corresponding to the critical defect targets are selectively sent back to the back-end system. The triggering conditions for small targets include at least one of the following: The bounding box area of ​​the candidate defect target is less than 0.05% of the total image area; The candidate defect target belongs to a preset set of sub-target defect categories, and the set of sub-target defect categories includes at least missing pins; The first confidence level of the candidate defect target is in the range of 0.50 to 0.

80.

2. The method for real-time screening of small target defects in transmission line inspection according to claim 1, characterized in that: The preset proportion for expanding outwards in all directions is 20% for each side of the bounding box, and the target size for size normalization is 128×128 pixels.

3. The method for real-time screening of small target defects in transmission line inspection according to claim 1, characterized in that: The specific rules for fusion include: When the defect category detected in the first detection matches the defect category identified in the second identification result, the defect category is retained, and the larger of the first confidence score and the second confidence score is taken as the final confidence score. When the two defect categories are inconsistent and the second confidence level is not lower than 0.80, the second identification result is used as the final defect category, and the second confidence level is used as the final confidence level. When the two defect categories are inconsistent and the second confidence level is below 0.80, the first detection result is retained as the final defect category, and a manual review mark is added to the target.

4. The method for real-time screening of small target defects in transmission line inspection according to claim 1, characterized in that: In the severity rating, for defects such as missing pins, f type Take 1.0; f conf Take the final confidence level directly; f size The calculation is based on the normalization of the ratio of the defect area to the area of ​​the component in which it is located.

5. The method for real-time screening of small target defects in transmission line inspection according to claim 1, characterized in that: It also includes a tiered approach to handling defective targets: If S ≥ 0.70, it is judged as a serious defect, and a key alarm is generated and sent back immediately; If 0.50 ≤ S < 0.70, it is judged as a defect of concern, cached locally and returned in batches when the network is available or made available for manual review; If S < 0.50, it is judged as a general defect, and only written to the local inspection record or a simplified report is generated.

6. The method for real-time screening of small target defects in transmission line inspection according to claim 1, characterized in that: The lightweight defect detection model is an improved YOLOv5s model quantized with INT8. The model file size is no more than 28MB, the input size is 640×640 pixels, and inference is performed on a portable terminal through the OpenVINO inference engine.

7. The method for real-time screening of small target defects in transmission line inspection according to claim 1, characterized in that: The on-site screening report should include at least the work batch information, image number, tower number, defect category, defect location, severity level, final confidence level, and recommended treatment items; for targets that do not reach the severity threshold, the results information is only cached locally on the portable terminal without triggering immediate data transmission.

8. A field small-target defect instant screening device for transmission line inspection using the method described in any one of claims 1 to 7, characterized in that: Deployed on portable terminals with edge computing capabilities, including: The image receiving and preprocessing unit is used to receive UAV inspection images and perform validity verification and standardization preprocessing. The initial inspection unit is used to call the lightweight defect detection model to perform the first detection on the preprocessed inspection image, and output the bounding box of the candidate defect target, the candidate defect category and the first confidence score; The small target re-judgment unit is used to generate an extended local region by expanding outward from the bounding box based on the candidate defect target that meets the small target triggering conditions. After being cropped, enlarged, and normalized in size, the extended local region is input into the small target enhancement recognition model for secondary re-judgment to obtain a second recognition result and a second confidence level. The small target enhancement recognition model includes a local feature enhancement sub-network with a lightweight spatial attention sub-module and a confidence calibration sub-network with a fully connected layer. The result fusion unit is used to fuse the first detection result and the second re-judgment result according to the preset fusion rules to obtain the final defect category and the final confidence level. The severity assessment unit is used to calculate the severity score based on the final defect category, final confidence level, and relative defect size, and to determine those with S≥0.70 as severe defects. The report generation and feedback unit is used to generate on-site screening reports and feed back the original images, partially cropped images, and structured results corresponding to the severely defective targets to the backend system.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When a computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 7.

Images