Grid beam coverage area anomaly detection method

By combining multi-scale convolutional neural networks and attention mechanisms with temporal consistency analysis, the problems of low efficiency and false positives in grid beam detection are solved, achieving high-precision anomaly detection in complex backgrounds, which is suitable for large-area grid beam inspection.

CN120932139APending Publication Date: 2025-11-11POLY CHANGDA ENGINEERING CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202511054512.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-30
Publication Date
2025-11-11

AI Technical Summary

Technical Problem

In existing technologies, grid beam detection methods are inefficient, highly subjective, and have difficulty identifying grid structures with significant changes in size and orientation. They are also easily confused with the background environment and lack the ability to perceive multiple scales and analyze temporal information.

Method used

By employing a multi-scale convolutional neural network combined with attention mechanisms and temporal consistency analysis, image sequences are acquired via UAVs, and features of the grid beam region are extracted. Anomaly detection is performed using a deep model with embedded attention mechanisms, and similarity calculation and edge computing deployment are combined to achieve accurate identification.

Benefits of technology

It improves the accuracy and real-time performance of anomaly detection in the grid beam coverage area, is suitable for large-area inspection, and can accurately identify grid beam structural anomalies in complex backgrounds, reducing the false judgment rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120932139A_ABST
    Figure CN120932139A_ABST
Patent Text Reader

Abstract

The invention relates to the field of grid beam detection, and discloses a grid beam coverage area anomaly detection method, which comprises the steps of S1, acquiring an image sequence or a video frame comprising a to-be-detected grid beam coverage area; s2, performing the following processing on each frame of image in the image sequence; s3, based on time sequence consistency analysis between adjacent frames, judging the structure or texture change condition of the grid beam candidate region; s4, performing similarity calculation on the candidate region and a preset abnormal image sample to generate an abnormal score of each image block; s5, if the abnormal score exceeds a preset threshold value, judging that the area is an abnormal grid beam; according to the method, multiple key technologies such as multi-scale feature extraction, an attention mechanism, time sequence consistency analysis and depth anomaly recognition are combined, so that the abnormal region of the grid beam structure can be accurately recognized under a complex background, and the detection precision, the real-time performance and the automation degree are remarkably improved; the method is especially suitable for large-area inspection and intelligent maintenance system integration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of grid beam detection, and more particularly to a method for detecting anomalies in the coverage area of ​​a grid beam. Background Technology

[0002] Grid beams, also known as grid beams or grid structures, are widely used in engineering fields such as slope protection, soil and water conservation, and high slope treatment. Their main function is to construct an interlaced grid-like support structure on the ground surface to limit soil slippage and support vegetation growth, thereby achieving slope stability control. During long-term service, grid beam structures may experience anomalies due to natural weathering, soil slippage, vegetation degradation, construction defects, or human damage. These anomalies mainly manifest as exposure, damage, collapse, displacement, localized cracking, or detachment of cover in the grid beam area. If these anomalies are not identified in time, they may trigger serious geological disasters and endanger the safety of the project.

[0003] Currently, engineering maintenance units typically assess anomalies in grid beam structures through manual inspections or static image detection. Manual inspections are inefficient, subjective, and unsuitable for large-scale, remote, or hazardous operations. Automatic identification methods based on traditional image segmentation or classification often suffer from the following technical shortcomings: detection algorithms rely on static single-frame images, lacking temporal information analysis capabilities and struggling to identify latent anomalies caused by gradual changes; feature extraction networks lack the ability to perceive targets at different scales, resulting in low accuracy in detecting grid structures with significant changes in size and orientation, and easily overlooking the saliency analysis of key areas in the image, making grid beams easily confused with the background environment (such as vegetation and shadows). Therefore, there is an urgent need for a method that can automatically, efficiently, and accurately identify anomalies in grid beam-covered areas, especially one with multi-scale perception capabilities and dynamic anomaly detection to improve identification efficiency and practical applicability. Summary of the Invention

[0004] The purpose of this application is to provide an anomaly detection method for grid beam coverage areas to address the following technical shortcomings of existing automatic recognition methods for image segmentation or classification: Feature extraction networks lack the ability to perceive targets at different scales, resulting in low detection accuracy for grid structures with significant changes in size and orientation; the method neglects the saliency analysis of key regions in the image, making grid beams easily confused with the background environment (such as vegetation and shadows). The specific technical solution is as follows: A method for detecting anomalies in a grid beam-covered area includes: S1. Acquire image sequences or video frames including the area covered by the grid beam to be detected; S2. Perform the following processing on each frame of the image sequence: S2A utilizes multi-scale convolutional neural networks to extract image features; S2B extracts features of the mesh beam region using a deep model with an embedded attention mechanism; S2C, based on a time-series consistency detection mechanism, filters out regions of structural change; S3. Based on temporal consistency analysis between adjacent frames, determine the structural or texture changes of candidate regions for mesh beams. S4. Calculate the similarity between the candidate region and the pre-set abnormal image samples to generate an abnormal score for each image block. S5. If the abnormal score exceeds the preset threshold, the area is determined to be an abnormal grid beam. S6. Output the image coordinates of the abnormal mesh beam area, or calculate its corresponding geographical location information based on the image metadata; S7. Transmit anomaly detection results to the edge or cloud terminal in real time.

[0005] As an improvement to the above technical solution, in step S1, an image sequence of the area to be detected is acquired by a high-definition camera mounted on a drone, with an image frame rate of 2-10 fps and a resolution of not less than 1080p.

[0006] As an improvement to the above technical solution, each time the drone takes an image, it first divides the grid beams into two triangular regions along the diagonal direction, with each group consisting of 3×3 cells. The drone then hovers and takes pictures of the two triangular regions respectively.

[0007] As one of the improvements to the above technical solutions, the attention mechanism is embedded in the multi-scale feature extraction network to perform channel dimension and spatial dimension weighted enhancement on image feature maps of different scales, so as to improve the recognition saliency of the grid beam region features in the multi-scale fusion process.

[0008] As one of the improvements to the above technical solutions, temporal consistency analysis is based on the attention-enhanced feature map. It calculates the structural similarity index or the distance change value of the feature embedding space for image blocks at the same position in the image sequence to determine whether there is an abnormal risk caused by dynamic changes in the region.

[0009] As one of the improvements to the above technical solution, the anomaly score is obtained by calculating the similarity between the attention-enhanced feature map and the embedded features of the preset anomaly sample in a unified metric space. The similarity calculation adopts a neural network model optimized by metric learning, and outputs the score results for screening abnormal grid beam regions.

[0010] As an improvement to the above technical solution, when outputting image coordinates or geographic location information, contour extraction and boundary fitting are performed on the abnormal regions detected in the image to generate refined abnormal region boundary information.

[0011] As one of the improvements to the above technical solution, the threshold for anomaly scoring is an adaptive dynamic threshold, which is automatically adjusted according to the overall brightness, texture complexity, or anomaly distribution density of the current image sequence.

[0012] As an improvement to the above technical solution, the method is deployed on devices that support edge computing, and the deep neural network model used is a lightweight structure.

[0013] The beneficial effects of this application are as follows: By combining multiple key technologies such as multi-scale feature extraction, attention mechanism, temporal consistency analysis and deep anomaly recognition, this method can accurately identify abnormal areas of grid beam structures in complex backgrounds, significantly improving detection accuracy, real-time performance and automation, and is especially suitable for integration into large-area inspection and intelligent maintenance systems.

[0014] In grid beam anomaly detection, images may contain the following: complex backgrounds (such as tree shadows, roads, and grass), grid beams with inconsistent positions and sizes, and anomaly areas that are often not obvious (such as slight exposure). If the model does not focus on the "grid beam region," it is likely to make a misjudgment. By introducing an attention mechanism, the model can automatically improve its ability to represent the grid beam region; focus on key features such as structural integrity and coverage in the feature map; effectively remove background information interference; and improve the model's accuracy in complex scenes.

[0015] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Of course, implementing any product or method of this application does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Existing automatic recognition methods based on image segmentation or classification typically suffer from the following technical shortcomings: Feature extraction networks lack the ability to perceive targets at different scales, resulting in low detection accuracy for grid structures with significant variations in size and orientation; the neglect of saliency analysis of key regions in the image makes grid beams easily confused with the background environment (such as vegetation and shadows). Please refer to [link to relevant documentation]. Figure 1 The present invention provides some embodiments to solve the above problems. A method for detecting anomalies in a grid beam coverage area includes the following steps: S1. Acquire image sequences or video frames including the area covered by the grid beam to be detected. The image sequence of the area to be detected is acquired by the high-definition camera mounted on the drone. The image frame rate is 2-10 fps and the resolution is not less than 1080p. Preferably, when the drone takes an image each time, it first divides the grid beam into two triangular areas along the diagonal direction with 3×3 cells as a group. The drone then hovers and takes pictures of the two triangular areas respectively. S2. Perform the following processing on each frame of the image sequence: S2A uses a multi-scale convolutional neural network to extract image features. Each frame of image is input into the multi-scale convolutional neural network to extract feature maps at different levels. After fusion, a unified feature map is output. S2B extracts features of the grid beam region through a deep model with an embedded attention mechanism, and inputs the fused feature map into a neural network with an attention module to highlight the saliency of the grid beam region in the spatial and channel dimensions. S2C, based on the temporal consistency detection mechanism, filters out structural change regions, uses target detection methods to locate candidate regions of "triangular" grid beams from the enhanced feature map, and generates image blocks; S3. Based on temporal consistency analysis between adjacent frames, determine the structural or texture changes of candidate regions of the grid beams. Calculate the structural similarity (SSIM) of candidate image blocks at the same location in the image sequence. If the SSIM changes significantly, mark it as a potential anomalous region. S4. Calculate the similarity between the candidate region and the pre-set abnormal image samples to generate an abnormal score for each image block. S5. If the anomaly score exceeds the preset threshold, the area is determined to be an abnormal mesh beam. Specifically, the current image block is compared with the constructed anomaly template library (such as exposed, collapsed, or damaged) by feature embedding, and the similarity is calculated as the anomaly score. For example, if the score is higher than the empirical threshold (such as 0.7), it is marked as an abnormal area, and the image coordinates (such as the upper left corner x, y, width w, and height h) are recorded. S6. Output the image coordinates of the abnormal mesh beam area, or calculate its corresponding geographical location information based on the image metadata; S7. Transmit the anomaly detection results to the edge terminal or cloud terminal in real time. Optionally, upload the detection results to the cloud or operation terminal in real time via a 4G / 5G module.

[0020] Using a nine-square grid system combined with diagonal imaging for drone inspection of grid beam structures ensures complete beam coverage and eliminates blind spots. Each nine-square grid area contains three horizontal beams and three vertical beams (a total of six beams). By taking two images (one from each diagonal side), all beams are ensured to be fully displayed in at least one image, avoiding the partial beam obstruction, blurred edges, or structural deformation caused by improper angles in traditional vertical imaging. Structural anomalies such as cracks and corrosion usually extend along the beams; if the beams are not fully visible, they are easily missed or misjudged. Diagonal shooting is more conducive to visual analysis along the beam direction, which is beneficial for identifying linear cracks, checking the spread of rust areas, and judging whether there is obvious bending or misalignment. In addition, the image always contains continuous and regularly arranged horizontal and vertical structures, which is convenient for deep learning models to extract features.

[0021] Regarding temporal consistency, when observing continuous images taken over a period of time (such as video or multi-frame images from a drone), if a certain region (such as a grid beam) maintains stable appearance characteristics across multiple frames, this region is said to have "temporal consistency". If the region suddenly undergoes significant changes in a few frames (such as color abrupt changes or structural deformation), the system considers it to have temporal inconsistency, which may indicate an anomaly.

[0022] This method combines several key technologies, such as multi-scale feature extraction, attention mechanism, temporal consistency analysis and deep anomaly recognition, to achieve accurate identification of abnormal areas in grid beam structures in complex backgrounds. It significantly improves detection accuracy, real-time performance and automation, and is especially suitable for integration into large-area inspection and intelligent maintenance systems.

[0023] In grid beam anomaly detection, images may contain the following: complex backgrounds (such as tree shadows, roads, and grass), grid beams with inconsistent positions and sizes, and anomaly areas that are often not obvious (such as slight exposure). If the model does not focus on the "grid beam region," it is likely to make a misjudgment. By introducing an attention mechanism, the model can automatically improve its ability to represent the grid beam region; focus on key features such as structural integrity and coverage in the feature map; effectively remove background information interference; and improve the model's accuracy in complex scenes.

[0024] In step S2, multi-scale feature extraction is achieved by introducing a feature pyramid network or a convolutional neural network with residual structure. Specifically, the multi-scale feature extraction module performs upsampling and fusion in the output of each layer through lateral connections, preserving shallow details and deep semantics, and improving the detection capability of mesh beam regions of different sizes and orientations.

[0025] In step S2, the attention mechanism is embedded in the multi-scale feature extraction network to perform channel and spatial dimension weighted enhancement on image feature maps at different scales, thereby improving the recognition saliency of the grid beam region features in the multi-scale fusion process. The attention module is added after the feature extraction backbone, and the corresponding module performs channel importance weighting enhancement. Alternatively, a CBAM structure combined with spatial attention can be used to improve the model's ability to focus on the "triangular shape" feature region of the grid beam and reduce background interference. After introducing channel attention and spatial attention mechanisms, the model can adaptively focus on the more discriminative grid beam region in the image, suppress background clutter (such as vegetation, shadows, etc.), improve the quality of feature representation, and enhance the generalization ability and accuracy of the detection model.

[0026] Temporal consistency analysis, based on attention-enhanced feature maps, calculates the structural similarity index or distance change value in the feature embedding space for image patches at the same location in an image sequence to determine whether there is anomaly risk caused by dynamic changes in the region. Image patches at the same location extracted from the image sequence are kept consistent in position through inter-frame registration. If the change in this region in consecutive frames exceeds a preset threshold, it is considered a structurally unstable region and enters the anomaly judgment process. Temporal consistency analysis can effectively identify structural anomalies that evolve slowly over time, such as vegetation degradation and foundation subsidence, making up for the weakness of traditional methods in detecting latent slow-changing anomalies and significantly enhancing the dynamic response capability and early warning capability of detection.

[0027] In some embodiments, a scoring method can be used for screening. Specifically, the anomaly score is obtained by calculating the similarity between the attention-enhanced feature map and the embedded features of a preset anomaly sample in a unified metric space. The similarity calculation uses a neural network model optimized by metric learning, and outputs a score result for screening anomaly grid beam regions. Specifically, the anomaly score is generated by an embedding network trained by metric learning. The network structure class and training samples include normal and anomaly grid beam image patch pairs. The similarity between the detected image patch and the anomaly template is calculated through this network. For example, the scoring range can be set to [0,1], with the closer to 1 representing the more anomaly. The threshold for the anomaly score is an adaptive dynamic threshold, which is automatically adjusted according to the overall brightness, texture complexity, or anomaly distribution density of the current image sequence. The threshold is not statically fixed, but dynamically adjusted in combination with factors such as the overall brightness histogram distribution of the current frame of the image sequence, the number of candidate regions, and texture complexity. By analyzing the global features of the current scene, the similarity judgment threshold is automatically updated to improve the algorithm's adaptability. Image quality is improved through image preprocessing (brightness equalization, noise reduction, etc.), which helps the model maintain stable performance in complex lighting environments. It is especially suitable for drone inspection image input under natural light conditions, effectively improving the overall robustness and adaptability of the system.

[0028] In step S5, when outputting image coordinates or geographic location information, contour extraction and boundary fitting are performed on the detected abnormal regions in the image to generate refined abnormal region boundary information. Before feature extraction, the original image undergoes the following enhancement processing: 1. Enhance contrast; 2. Blur to suppress image noise; 3. Grayscale normalization is performed to reduce lighting interference; This preprocessing step is particularly suitable for sloping areas with significant shading variations, and helps improve recognition.

[0029] The method is deployed on devices that support edge computing. The deep neural network model used is a lightweight structure. Specifically, the method is deployed on devices equipped with edge AI chips (such as NVIDIA Jetson Nano, RK3588, etc.). The model used is a compressed structure, which is well known to those skilled in the art. The quantized model size is <10MB, supports real-time detection of 8-15fps, and is adapted to the computing power limitations of battery-powered drones and mobile terminals.

[0030] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0031] It should be noted that the terms "first," "second," etc., used in this document are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features.

[0032] The terms “center,” “upper,” “lower,” “front,” “back,” “left,” “right,” “vertical,” “horizontal,” “top,” “bottom,” “inner,” and “outer” used herein to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings and are used only for the convenience of describing this application and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0033] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A method for detecting anomalies in a grid beam-covered area, characterized in that, Includes the following steps: S1. Acquire image sequences or video frames including the area covered by the grid beam to be detected; S2. Perform the following processing on each frame of the image sequence: S2A: Extracting image features using a data model; S2B extracts features of the mesh beam region using a deep model with an embedded attention mechanism; S2C, based on a time-series consistency detection mechanism, filters out regions of structural change; S3. Based on temporal consistency analysis between adjacent frames, determine the structural or texture changes of candidate regions for mesh beams; S4. Calculate the similarity between the candidate region and the pre-set abnormal image samples to generate an abnormal score for each image block. S5. If the abnormal score exceeds the preset threshold, the area is determined to be an abnormal grid beam. S6. Output the image coordinates of the abnormal mesh beam area, or calculate its corresponding geographical location information based on the image metadata; S7. Transmit anomaly detection results to the edge or cloud terminal in real time.

2. The method for detecting anomalies in a grid beam coverage area according to claim 1, characterized in that: In step S1, an image sequence of the area to be detected is acquired by a high-definition camera mounted on the drone. The image frame rate is 2-10 fps and the resolution is not less than 1080p.

3. The method for detecting anomalies in a grid beam coverage area according to claim 2, characterized in that: Each time the drone takes an image, it first divides the grid beams into two triangular regions along the diagonal, with each group consisting of 3×3 cells. The drone then hovers and takes images of each of the two triangular regions.

4. The method for detecting anomalies in a grid beam coverage area according to claim 1, characterized in that: An attention mechanism is embedded in a multi-scale feature extraction network to perform channel and spatial dimension weighted enhancement on image feature maps at different scales, thereby improving the recognition saliency of grid beam region features in the multi-scale fusion process.

5. A method for detecting anomalies in a grid beam coverage area according to claim 4, characterized in that: Temporal consistency analysis, based on attention-enhanced feature maps, calculates the structural similarity index or the distance change value in the feature embedding space for image patches at the same location in the image sequence to determine whether there is anomaly risk caused by dynamic changes in the region.

6. A method for detecting anomalies in a grid beam coverage area according to claim 5, characterized in that: The anomaly score is obtained by calculating the similarity between the attention-enhanced feature map and the embedded features of the preset anomaly samples in a unified metric space. The similarity calculation uses a neural network model optimized by metric learning, and the output is a score result used to screen out anomaly grid beam regions.

7. A method for detecting anomalies in a grid beam coverage area according to claim 1, characterized in that: When outputting image coordinates or geographic location information, contour extraction and boundary fitting are performed on the abnormal regions detected in the image to generate refined abnormal region boundary information.

8. A method for detecting anomalies in a grid beam coverage area according to claim 1, characterized in that: The threshold for anomaly scoring is an adaptive dynamic threshold that is automatically adjusted based on the overall brightness, texture complexity, or anomaly distribution density of the current image sequence.

9. A method for detecting anomalies in a grid beam coverage area according to claim 1, characterized in that: The method is deployed on devices that support edge computing, and the deep neural network model used is a lightweight structure.

Citation Information

Cited By

  • Video abnormal behavior real-time detection method and system based on deep learning

    CN122244956A

  • Beidou skyeye global sensing safety early warning base large model system

    CN122313672A