Power transmission channel risk target detection method and related device

Abstract

The application discloses a power transmission channel risk target detection method and related equipment, and relates to the technical field of power transmission channel operation and maintenance. The power transmission channel risk target detection method comprises the following steps: acquiring a to-be-detected image of a power transmission channel, and performing target detection on the to-be-detected image; acquiring a historical scene semantic analysis result corresponding to the to-be-detected image, performing time sequence registration and abnormal frame filtering on the historical scene semantic analysis result, and obtaining a time sequence stable scene semantic analysis result; performing semantic analysis on the to-be-detected image based on a preset scene semantic analysis model, combining the scene semantic analysis result, extracting decision response features of a region where a candidate target is located, and the decision response features representing internal decision information on which the scene semantic analysis model determines the region as the candidate target; and performing physical authenticity verification on the candidate target according to the decision response features, and obtaining a risk target detection result. The application solves the technical problem of low power transmission channel operation and maintenance efficiency caused by frequent false alarms.

Description

Technical Field

[0001] This application relates to the field of power transmission channel operation and maintenance technology, and in particular to methods and related equipment for detecting risk targets in power transmission channels. Background Technology

[0002] Transmission corridors are the core backbone of the power grid's safe operation. They are widely distributed and traverse complex environments, constantly facing threats from external forces such as construction machinery intrusion, wildfires, and suspended debris. Timely and accurate detection of these potential hazards is a crucial prerequisite for ensuring the safe and stable operation of the power grid. With the large-scale deployment of visual monitoring devices in transmission corridors, computer vision-based automated detection methods have gradually replaced manual inspections, becoming the mainstream means of identifying potential hazards in transmission corridors.

[0003] In conventional techniques, hazard detection in power transmission channel images typically employs deep learning models, such as target detection models based on convolutional neural networks or transformer architectures, to locate and classify potential risk targets from the input image. To reduce false alarms, some conventional techniques further incorporate semantic segmentation models, filtering detection results by calculating the pixel area ratio of specific scene categories (such as sky, water surface, and towers) within the detection box. However, these conventional techniques suffer from poor physical authenticity discrimination. Filtering based on pixel area ratio only reflects the semantic composition ratio within the detection box and cannot distinguish whether the target is generated by a real physical entity or by spatial artifacts such as water reflections, cloud reflections, or internal tower textures. When the semantic category of the artifact region highly overlaps with that of the real target at the pixel level, numerous false alarms that violate physical principles are generated, crowding out the window for addressing real risks and reducing operational efficiency. Summary of the Invention

[0004] The main purpose of this application is to provide a method and related equipment for detecting risk targets in power transmission channels, aiming to solve the technical problem of low operation and maintenance efficiency of power transmission channels caused by frequent false alarms.

[0005] To achieve the above objectives, this application proposes a method for detecting risk targets in power transmission channels, the method comprising: Acquire an image of the power transmission channel to be detected, perform target detection on the image to be detected, and obtain at least one candidate target; The historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected are obtained, and the historical scene semantic analysis results are temporally registered and abnormal frames are filtered to obtain temporally stable scene semantic analysis results. Based on a preset scene semantic analysis model, the image to be detected is semantically parsed, and the decision response features of the region where the candidate target is located are extracted by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. Based on the decision response characteristics, the physical authenticity of the candidate targets is verified to filter out non-real risk targets and obtain the risk target detection results.

[0006] In one embodiment, the step of extracting the decision response features of the region where the candidate target is located includes: The image to be detected is semantically parsed according to a preset scene semantic analysis model, and the semantic parsing results are organized to generate a normalized attention heatmap. Based on the bounding box coordinates of the candidate target in the image to be detected, a local attention sub-map that is registered with the bounding box coordinate space is extracted from the normalized attention heatmap as the decision response feature.

[0007] In one embodiment, the normalized attention heatmap includes semantic category response weights assigned by the scene semantic analysis model to each spatial location in the image to be detected, and the step of verifying the physical authenticity of the candidate target based on the decision response features includes: Based on the scene semantic analysis model, the scene semantic category of the region where the candidate target is located in the image to be detected is determined; Based on the semantic category response weights of each spatial location in the decision response features, the concentration of the semantic category response weights is calculated, and the concentration is used as the singleness index of the decision semantics on which the scene semantic analysis model determines the region as the candidate target. Calculate the degree of deviation between the singleness index of the decision semantics and the preset benchmark singleness index corresponding to the scene semantic category; When the deviation exceeds a preset deviation threshold, the candidate target is determined to have failed the physical authenticity verification.

[0008] In one embodiment, the step of verifying the physical authenticity of the candidate target based on the decision response characteristics includes: Extract the first semantic feature distribution inside the bounding box of the candidate target; According to the preset extraction rules, the distribution of second semantic features of the outer edge region adjacent to the bounding box is extracted; Calculate the semantic boundary difference degree between the first semantic feature distribution and the second semantic feature distribution. The semantic boundary difference degree is used to measure the transition continuity between the candidate target and the surrounding scene at the semantic level. When the semantic boundary difference is lower than a preset difference threshold, the candidate target is determined to have failed the physical authenticity verification.

[0009] In one embodiment, the step of calculating the semantic boundary difference between the first semantic feature distribution and the second semantic feature distribution includes: The first semantic feature distribution and the second semantic feature distribution are normalized respectively to obtain the first semantic probability distribution and the second semantic probability distribution; Calculate the cross-entropy between the first semantic probability distribution and the second semantic probability distribution, and calculate the initial semantic boundary difference degree based on the cross-entropy; Based on the initial semantic boundary difference, the boundary region of the bounding box is subjected to topological structure reasoning according to the preset graph convolutional network, and the final semantic boundary difference is output. The preset graph convolutional network is trained on a training set containing physical entity samples and non-physical entity samples, and the boundary semantic features of the non-physical entity samples have smooth transition characteristics.

[0010] In one embodiment, the step of obtaining the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and performing temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results includes: Obtain the temporal historical frame corresponding to the image to be detected, and obtain the historical scene semantic analysis result corresponding to the temporal historical frame; The image to be detected and the temporal history frame are input into a preset offset prediction network to obtain the deformation offset vector between the image to be detected and the temporal history frame. Based on the deformation offset vector, the historical scene semantic analysis result is geometrically transformed, the historical scene semantic analysis result is registered to the coordinate space of the image to be detected, and the registered historical scene semantic analysis result is written into a preset temporal buffer. Calculate the first global scene entropy of the current frame in the time-series buffer, and obtain the second global scene entropy of adjacent historical frames in the time-series buffer; When the fluctuation amplitude of the first global scene entropy relative to the second global scene entropy exceeds a preset fluctuation threshold, an abnormal state signal is generated, the scene semantic analysis result of the most recent valid frame in the time-series buffer is latched, and the writing operation of the analysis result of the current frame to the time-series buffer is suspended. Furthermore, to achieve the above objectives, this application also proposes a power transmission channel risk target detection device, which includes: The detection module is used to acquire the image to be detected of the power transmission channel, perform target detection on the image to be detected, and obtain at least one candidate target; The enhancement module is used to obtain the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and to perform temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results. The extraction module is used to perform semantic parsing on the image to be detected based on a preset scene semantic analysis model, and extract the decision response features of the region where the candidate target is located by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. The verification module is used to verify the physical authenticity of the candidate targets based on the decision response characteristics, so as to filter out non-real risk targets and obtain risk target detection results.

[0011] In addition, to achieve the above objectives, this application also proposes a power transmission channel risk target detection device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the power transmission channel risk target detection method described above.

[0012] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the power transmission channel risk target detection method described above.

[0013] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the power transmission channel risk target detection method described above.

[0014] One or more technical solutions proposed in this application have at least the following technical effects: Compared to related technologies that use pixel area ratio statistics within the detection box to filter false detections, this application acquires a target image of a power transmission channel, performs target detection on the target image to obtain at least one candidate target; acquires historical scene semantic analysis results of the corresponding temporal historical frames of the target image, and performs temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results; based on a preset scene semantic analysis model, performs semantic parsing on the target image, and combines the temporally stable scene semantic analysis results to extract decision response features of the region where the candidate target is located. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target; based on the decision response features, performs physical authenticity verification on the candidate target to filter out non-real risk targets and obtain risk target detection results. Understandably, when the scene semantic analysis model makes a target determination on the image of the power transmission channel, this application extracts the internal decision information on which the scene semantic analysis model determines the image region as a candidate target, thereby realizing the reverse tracing of the model's decision logic. Therefore, based on this internal decision information, it is possible to identify whether the semantic source of the candidate target matches its physical category, and then determine whether the candidate target has physical authenticity. Finally, it completes the filtering of non-real risk targets, obtains reliable risk target detection results, and solves the technical problem of low operation and maintenance efficiency of power transmission channels caused by frequent false alarms. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an embodiment of the power transmission channel risk target detection method of this application; Figure 2 This is a schematic diagram of the overall structure of the transmission channel risk target detection method provided in this application; Figure 3 A schematic diagram of the target detection model network structure provided in the embodiment of the power transmission channel risk target detection method of this application; Figure 4 The temporal semantic scene map provided for the embodiment of the power transmission channel risk target detection method of this application; Figure 5 The semantic logic filtering diagram provided for the embodiments of the power transmission channel risk target detection method of this application; Figure 6 This application provides a method result diagram for an embodiment of the method for detecting risk targets in power transmission channels. Figure 7 This is a schematic diagram of the module structure of the power transmission channel risk target detection device according to an embodiment of this application; Figure 8 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the power transmission channel risk target detection method in the embodiments of this application.

[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0021] The main solution in this application's embodiments is: Acquire an image of the power transmission channel to be detected, perform target detection on the image to be detected, and obtain at least one candidate target; The historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected are obtained, and the historical scene semantic analysis results are temporally registered and abnormal frames are filtered to obtain temporally stable scene semantic analysis results. Based on a preset scene semantic analysis model, the image to be detected is semantically parsed, and the decision response features of the region where the candidate target is located are extracted by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. Based on the decision response characteristics, the physical authenticity of the candidate targets is verified to filter out non-real risk targets and obtain the risk target detection results.

[0022] In this embodiment, the application uses a power transmission channel risk target detection device as the execution subject, which will be referred to as "device" for ease of description.

[0023] Because existing technologies lack feature-level interaction between object detection models and semantic segmentation models, and rely solely on rule filtering based on the percentage of pixel area within the detection box, they are unable to distinguish between real physical entities and spatial artifacts such as water reflections and cloud reflections, resulting in a large number of false alarms.

[0024] This application provides a solution that enables the target detection process to inherently verify the physical authenticity of the detection results. Unlike conventional techniques that rely on external rules for post-processing filtering after the detection results are output, this application obtains the internal decision-making basis upon which the scene semantic analysis model bases its target judgments, and uses this internal decision-making basis as direct evidence to verify the physical authenticity of candidate targets. When the model identifies a region as a candidate target, this application does not directly accept the judgment result, but instead examines the decision response features upon which the model bases its judgment, and uses these features to verify the physical authenticity of the candidate target. If there is a contradiction between the model's internal basis for the judgment and the corresponding scene semantics, it indicates that the candidate target may be triggered by background texture or optical artifacts rather than by a real physical entity, thus filtering out the candidate target as a non-real risk target. This verification process operates at the model's decision level, not the pixel level, fundamentally distinguishing between real physical entities and spatial artifacts, achieving a leap from spatial overlap judgment to feature-based causal reasoning, and effectively reducing the impact of frequent false alarms on the operation and maintenance efficiency of power transmission channels.

[0025] Based on this, embodiments of this application provide a method for detecting risk targets in power transmission channels, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the power transmission channel risk target detection method of this application.

[0026] In this embodiment, the power transmission channel risk target detection method includes steps S10~S40: Step S10: Obtain the image to be detected of the power transmission channel, perform target detection on the image to be detected, and obtain at least one candidate target; It should be noted that the images to be detected include, but are not limited to, visible light images, infrared thermal images, or aerial images taken by inspection drones collected by power transmission channel visualization monitoring devices. These images may contain scene elements such as power transmission towers, conductors, vegetation, water surfaces, and the sky, as well as potential risk targets such as construction machinery, smoke, and foreign objects. Target detection refers to locating and identifying potentially risky areas in the images to be detected using a preset target detection model, and outputting the corresponding bounding box and category label for that area. Candidate targets are those initially determined by the target detection model to be risky; their detection results may include false targets composed of water reflections, cloud reflections, or internal textures of the towers.

[0027] Understandably, this step completes the preliminary target detection of the image to be detected in the power transmission channel, providing a set of candidate targets for subsequent physical authenticity verification. The target detection in this embodiment is not limited to a specific model architecture; it can employ anchor-bound or anchor-free detection paradigms. The output will undergo causal logic verification based on scene semantics in subsequent steps.

[0028] Step S20: Obtain the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and perform temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results; It should be noted that a time-series historical frame refers to a continuous historical image frame that was acquired before the image to be detected and belongs to the same monitoring point of the power transmission channel as the image to be detected. The time-series historical frame and the image to be detected constitute a continuous image sequence in the time dimension. The historical scene semantic analysis result refers to the stored scene semantic analysis output obtained by the preset scene semantic analysis model after performing semantic parsing on the time-series historical frame, including but not limited to the semantic segmentation mask or semantic feature map corresponding to the time-series historical frame. Since the power transmission channel monitoring equipment is exposed to the outdoor environment all year round, there are common minor displacements of the equipment caused by micro-movements of the pan-tilt unit, wind vibration, or temperature changes. Directly using the semantic analysis result of historical frames will introduce registration errors due to the spatial position deviation between frames. At the same time, the monitoring screen may be distorted as a whole due to direct sunlight, lens obstruction, or violent shaking of the pan-tilt unit, causing the semantic analysis result of a single frame to be completely invalid. To this end, this step acquires the historical semantic analysis results of temporal historical frames, performs spatial registration and anomaly frame filtering, accurately maps the semantic information of historical frames to the coordinate space of the current image to be detected, and removes anomaly frames caused by overall image distortion, ultimately obtaining temporally stable scene semantic analysis results with temporal consistency and spatial accuracy. This process provides a robust and high-confidence temporally enhanced semantic reference for subsequent extraction of decision response features, ensuring that the acquisition of internal decision information is rooted in the semantic parsing of the current frame and integrates stable semantic evidence from historical frames, thereby improving the reliability of model decision tracing in complex outdoor scenes.

[0029] Step S30: Based on the preset scene semantic analysis model, perform semantic parsing on the image to be detected, and combine the time-stable scene semantic analysis results to extract the decision response features of the region where the candidate target is located. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. It should be noted that the preset scene semantic analysis model includes, but is not limited to, a semantic segmentation model based on a transformer architecture. This model is pre-trained to have the ability to perform pixel-level semantic parsing of elements in the power transmission channel scene. Decision response features refer to the quantitative representation of the internal decision information relied upon by the scene semantic analysis model when classifying the region where the candidate target is located into a specific category. These include, but are not limited to, the normalized attention heatmap, intermediate layer feature activation map, or gradient response map generated by the model during semantic parsing. These representations reflect the degree of attention and semantic tendency of the model to each spatial location within the region when making classification judgments.

[0030] Understandably, this step, by extracting decision information from within the scene semantic analysis model, achieves a reverse tracing from the result level of "what is the target" to the level of the basis for the model's judgment. The decision response features extracted in this implementation do not rely on pixel category statistics within the detection box, but directly obtain the internal logical basis for the model's target determination, providing an analytical foundation for subsequent judgments on whether there are logical contradictions in this basis.

[0031] Step S40: Based on the decision response characteristics, perform physical authenticity verification on the candidate targets to filter out non-real risk targets and obtain risk target detection results.

[0032] It should be noted that physical authenticity verification refers to the process of logically judging whether a candidate target corresponds to a real physical entity based on the model decision-making criteria reflected in the decision response characteristics. This verification differs from hard rule filtering based on pixel area ratio. Instead, it determines whether the candidate target is generated by a real physical entity or a spatial artifact by analyzing the consistency between the semantic tendency relied upon by the scene semantic analysis model when classifying the region as a candidate target and the actual semantic attributes of the region. The risk target detection result refers to the remaining targets, along with their location and category information, that are judged to have a real physical correspondence after the physical authenticity verification, filtering out false targets that violate physical logic.

[0033] Understandably, this step, by applying physical authenticity verification to candidate targets, achieves the distinction between genuine and false detection results at the feature causal reasoning level. This implementation establishes a consistency verification mechanism between the internal basis of the model's judgment and the final output scene semantics. When there is a logical contradiction between the two, the candidate target is deemed to lack physical authenticity and is filtered out, thereby effectively suppressing false alarms caused by spatial artifacts such as water reflections and cloud reflections, and ultimately outputting risk target detection results with physical credibility.

[0034] This embodiment provides a method for detecting risk targets in power transmission channels. When a scene semantic analysis model makes a target determination on a power transmission channel image, this application extracts the internal decision information on which the scene semantic analysis model determines the image region as a candidate target. This enables reverse tracing of the model's decision logic. Therefore, based on this internal decision information, it is possible to identify whether the semantic source of the candidate target matches its physical category, thereby determining whether the candidate target has physical authenticity. Finally, non-real risk targets are filtered out, and reliable risk target detection results are obtained. This solves the technical problem of low operation and maintenance efficiency of power transmission channels caused by frequent false alarms.

[0035] In one feasible implementation, the step of extracting the decision response features of the region where the candidate target is located includes: The image to be detected is semantically parsed according to a preset scene semantic analysis model, and the semantic parsing results are organized to generate a normalized attention heatmap. Based on the bounding box coordinates of the candidate target in the image to be detected, a local attention sub-map that is registered with the bounding box coordinate space is extracted from the normalized attention heatmap as the decision response feature.

[0036] It should be noted that semantic parsing refers to the pixel-level or region-level classification of scene elements in the image to be detected by the preset scene semantic analysis model. These scene elements include, but are not limited to, physically meaningful region types within the power transmission channel environment, such as the sky, water surface, towers, power lines, vegetation, roads, and buildings. Organizing the semantic parsing results involves normalizing and aggregating the semantic category response values ​​assigned to each spatial location by the scene semantic analysis model during the parsing process, forming a unified, indexable two-dimensional response intensity distribution map. The normalized attention heatmap is a two-dimensional feature map after normalization. This heatmap includes the semantic category response weights assigned by the scene semantic analysis model to each spatial location in the image to be detected. These weights characterize the level of attention the scene semantic analysis model pays to each spatial location when making semantic judgments. Bounding box coordinates refer to the coordinate parameters of the rectangular positioning box corresponding to the candidate target in the image to be detected, including but not limited to the center point coordinates, width, and height values ​​of the rectangular positioning box, or the coordinates of the upper left and lower right vertices of the rectangular positioning box. Spatial registration refers to mapping the bounding box coordinates within a spatial range defined in the image to be detected to the corresponding position in the normalized attention heatmap, ensuring a strict spatial correspondence between the cropped region and the region where the candidate target is located. A local attention subgraph is an attention response subgraph extracted from the normalized attention heatmap according to the spatially registered coordinate range, covering only the region where the candidate target is located. This local attention subgraph retains the local decision information relied upon by the scene semantic analysis model when identifying the region as the candidate target.

[0037] It is understood that this implementation extracts a normalized attention heatmap from the semantic parsing results of the scene semantic analysis model, and maps the bounding box coordinates of candidate targets to this heatmap to extract a local attention subgraph as a decision response feature. This achieves precise localization and extraction of the model's decision-making basis from the global semantic parsing results to the region where the candidate target is located. The local attention subgraph obtained in this step directly reflects the attention distribution used by the scene semantic analysis model when classifying the region where the candidate target is located into a specific category, providing a quantifiable analytical object for subsequent verification of whether there are semantic logical contradictions in this decision-making basis.

[0038] In one feasible implementation, the normalized attention heatmap includes semantic category response weights assigned by the scene semantic analysis model to each spatial location in the image to be detected, and the step of verifying the physical authenticity of the candidate target based on the decision response features includes: Based on the scene semantic analysis model, the scene semantic category of the region where the candidate target is located in the image to be detected is determined; Based on the semantic category response weights of each spatial location in the decision response features, the concentration of the semantic category response weights is calculated, and the concentration is used as the singleness index of the decision semantics on which the scene semantic analysis model determines the region as the candidate target. Calculate the degree of deviation between the singleness index of the decision semantics and the preset benchmark singleness index corresponding to the scene semantic category; When the deviation exceeds a preset deviation threshold, the candidate target is determined to have failed the physical authenticity verification.

[0039] It should be noted that the scene semantic category refers to the semantic label assigned to the region where the candidate target is located by the scene semantic analysis model after pixel-level semantic parsing. These semantic labels include, but are not limited to, category identifiers with clear physical meaning in the power transmission channel environment, such as the sky, water surface, towers, wires, vegetation, roads, and buildings. The semantic category response weight refers to the response intensity value of each spatial location in the normalized attention heatmap to a specific semantic category. This response intensity value reflects the confidence level of the scene semantic analysis model in determining that the spatial location belongs to a specific semantic category. Concentration refers to the distribution clustering measure obtained by statistically aggregating the semantic category response weights of each spatial location in the decision response features. This concentration reflects whether the scene semantic analysis model's decision-making basis for identifying the region as the candidate target is highly concentrated on a single semantic category. The singleness index of decision semantics refers to a quantitative indicator determined based on the concentration, characterizing the specificity of the semantic basis relied upon by the scene semantic analysis model when making target determinations. A higher singleness index indicates that the scene semantic analysis model is more inclined to rely solely on features of a single semantic category for its determinations. The preset benchmark singularity index refers to a pre-defined singularity reference value for the semantic category of the scene. This index characterizes the semantic distribution characteristics that the decision-making basis of the scene semantic analysis model should exhibit under normal judgment scenarios of real physical targets. The deviation degree refers to the difference between the decision semantic singularity index and the preset benchmark singularity index. This deviation degree measures the degree of inconsistency between the scene semantic analysis model's judgment logic for the current candidate target and the expected judgment logic for the real physical target. The preset deviation threshold is a pre-set critical value for the deviation degree. When the deviation degree exceeds the preset deviation threshold, it indicates a significant contradiction between the scene semantic analysis model's judgment logic for the candidate target and the expected judgment logic for the real physical target.

[0040] Understandably, this implementation method quantifies the internal decision-making basis relied upon by the scene semantic analysis model when judging candidate targets by calculating the concentration of semantic category response weights and constructing a singleness index for decision semantics. In normal scenarios involving the judgment of real physical targets, the decision-making basis of the scene semantic analysis model typically exhibits semantic distribution characteristics matching the target category. However, when a candidate target is composed of spatial artifacts such as water reflections or cloud reflections, although the model may misclassify it as a specific target category, the semantic basis upon which its internal decision relies often concentrates on background semantic categories that do not match the target category. This leads to a significant deviation between the singleness index of decision semantics and the preset benchmark singleness index. This step, by comparing the aforementioned degree of deviation with a preset deviation threshold, achieves a quantitative verification of the physical authenticity of candidate targets from the model's decision logic level, effectively identifying false targets generated by misjudgment of background textures.

[0041] In one feasible implementation, the step of verifying the physical authenticity of the candidate target based on the decision response characteristics includes: Extract the first semantic feature distribution inside the bounding box of the candidate target; According to the preset extraction rules, the distribution of second semantic features of the outer edge region adjacent to the bounding box is extracted; Calculate the semantic boundary difference degree between the first semantic feature distribution and the second semantic feature distribution. The semantic boundary difference degree is used to measure the transition continuity between the candidate target and the surrounding scene at the semantic level. When the semantic boundary difference is lower than a preset difference threshold, the candidate target is determined to have failed the physical authenticity verification.

[0042] It should be noted that the first semantic feature distribution refers to the statistical distribution of semantic categories obtained after extracting semantic features from each pixel or spatial position within the bounding box of the candidate target. This first semantic feature distribution reflects the semantic composition of the region within the bounding box. The preset extraction rule refers to pre-configured rules for determining the spatial extent of the outer edge region, including but not limited to using the ring-shaped region formed by expanding the bounding box outward by a preset pixel width as the outer edge region, or using the region enclosed by shifting each edge of the bounding box outward by a preset distance as the outer edge region. The second semantic feature distribution refers to the statistical distribution of semantic categories obtained after extracting semantic features from each pixel or spatial position within the outer edge region adjacent to the bounding box according to the preset extraction rule. This second semantic feature distribution reflects the semantic composition of the surrounding scene of the candidate target. The semantic boundary difference degree refers to the quantified value of the distribution difference between the first semantic feature distribution and the second semantic feature distribution. This degree measures the semantic continuity of the transition between the candidate target and its surrounding scene. A higher semantic boundary difference degree indicates a more drastic change in semantic features between the inner and outer edge regions of the bounding box, meaning a clear semantic boundary exists between the candidate target and its surrounding scene. Conversely, a lower semantic boundary difference degree indicates a smoother and more continuous transition in semantic features between the inner and outer edge regions of the bounding box. The preset difference threshold is a pre-defined critical value for semantic boundary difference degree, used to distinguish the differences in semantic boundary transition characteristics between real physical entities and spatial artifacts.

[0043] Understandably, this implementation utilizes the fundamental differences in semantic boundary transition characteristics between real physical entities and spatial artifacts to verify physical authenticity. For real physical entities, such as construction machinery on grass, there is a significant semantic abrupt change between the semantic features of the machinery inside its bounding box and the semantic features of the grass in the adjacent area outside the bounding box, exhibiting a high degree of semantic boundary difference. For false targets formed by water reflections or optical artifacts, since the reflection and the surrounding water surface are semantically continuous and smoothly transitioned, the semantic feature distribution inside the bounding box is highly similar to the semantic feature distribution of the outer edge region, exhibiting a low degree of semantic boundary difference. This step, by quantitatively calculating the semantic boundary difference between the first semantic feature distribution and the second semantic feature distribution, and comparing the semantic boundary difference with a preset difference threshold, realizes the logical judgment of whether the candidate target possesses physical entity boundary features, effectively identifying false targets with smooth semantic transitions, such as reflections and mirrors.

[0044] In one feasible implementation, the step of calculating the semantic boundary difference degree between the first semantic feature distribution and the second semantic feature distribution includes: The first semantic feature distribution and the second semantic feature distribution are normalized respectively to obtain the first semantic probability distribution and the second semantic probability distribution; Calculate the cross-entropy between the first semantic probability distribution and the second semantic probability distribution, and calculate the initial semantic boundary difference degree based on the cross-entropy; Based on the initial semantic boundary difference, the boundary region of the bounding box is subjected to topological structure reasoning according to the preset graph convolutional network, and the final semantic boundary difference is output. The preset graph convolutional network is trained on a training set containing physical entity samples and non-physical entity samples, and the boundary semantic features of the non-physical entity samples have smooth transition characteristics.

[0045] It should be noted that normalization refers to converting the statistical values ​​of each category in the first and second semantic feature distributions into probability values ​​within the range of zero to one, so that the first and second semantic feature distributions can be compared under a unified dimension. The first semantic probability distribution refers to the probability distribution representing each semantic category within the bounding box region, obtained after normalization of the first semantic feature distribution. The second semantic probability distribution refers to the probability distribution representing each semantic category within the outer edge region, obtained after normalization of the second semantic feature distribution. Cross-entropy is an information-theoretic metric between the first and second semantic probability distributions. Cross-entropy measures the degree of difference between the two probability distributions; the more similar the two probability distributions, the smaller the value of the cross-entropy. The initial semantic boundary difference degree refers to a preliminary estimate of the semantic boundary difference degree calculated directly based on the cross-entropy, without further reasoning or optimization. The pre-trained graph convolutional network (GCNN) is a graph neural network model used for topological inference of image regions. The GCNN treats each spatial location within the bounding box as a graph node, constructs edge connections based on the semantic feature similarity between these nodes, and infers the semantic topological structure of the boundary region through graph convolution operations. The training set of the GCNN includes physical entity samples and non-physical entity samples. Physical entity samples are image regions of real physical objects and their surrounding scenes, while non-physical entity samples are image regions of spatial artifacts such as water reflections or optical reflections and their surrounding scenes. Non-physical entity samples exhibit smooth transition characteristics in semantic features within the boundary region. Topological inference refers to the process by which the GCNN, after modeling the graph structure of the bounding box's boundary region, infers whether there are semantically abrupt structural changes in the boundary region by aggregating the semantic feature information of adjacent nodes. The final semantic boundary difference is a semantic boundary difference metric value output by the GCNN after performing topological inference on the boundary region, incorporating spatial topological information.

[0046] Understandably, this implementation obtains the initial semantic boundary difference degree through normalization and cross-entropy calculation, preliminarily quantifying the degree of difference in semantic distribution between the inner and outer edge regions of the bounding box. Based on this, by introducing a pre-defined graph convolutional network to perform topological structure inference on the boundary region, this step can further capture the spatial structure information of the boundary region, distinguishing between semantic abrupt changes caused by real physical entities and pseudo-differences caused by noise or local texture fluctuations. Since the graph convolutional network has learned the essential differences in topological structure between real boundary abrupt changes and smooth transition artifacts from physical entity samples and non-physical entity samples during the training phase, the final semantic boundary difference degree output by its inference can more accurately reflect the physical boundary characteristics of the candidate target boundary region, providing a more reliable quantitative basis for subsequent physical authenticity determination.

[0047] In one feasible implementation, the step of obtaining the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and performing temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results includes: Obtain the temporal historical frame corresponding to the image to be detected, and obtain the historical scene semantic analysis result corresponding to the temporal historical frame; The image to be detected and the temporal history frame are input into a preset offset prediction network to obtain the deformation offset vector between the image to be detected and the temporal history frame. Based on the deformation offset vector, the historical scene semantic analysis result is geometrically transformed, the historical scene semantic analysis result is registered to the coordinate space of the image to be detected, and the registered historical scene semantic analysis result is written into a preset temporal buffer. Calculate the first global scene entropy of the current frame in the time-series buffer, and obtain the second global scene entropy of adjacent historical frames in the time-series buffer; When the fluctuation of the first global scene entropy relative to the second global scene entropy exceeds a preset fluctuation threshold, an abnormal state signal is generated, the scene semantic analysis result of the most recent valid frame in the time-series buffer is latched, and the writing operation of the analysis result of the current frame to the time-series buffer is suspended.

[0048] It should be noted that a temporal history frame refers to a continuous historical image frame that was acquired before the image to be detected and belongs to the same monitoring point as the image to be detected. The temporal history frame and the image to be detected constitute a continuous image sequence in the time dimension. The historical scene semantic analysis result refers to the stored scene semantic analysis output obtained by the preset scene semantic analysis model after performing semantic parsing on the temporal history frame, including but not limited to the semantic segmentation mask or semantic feature map corresponding to the temporal history frame. The preset offset prediction network refers to a pre-trained neural network model used to predict the spatial offset between adjacent image frames. The offset prediction network can extract corresponding feature points from two image frames and estimate the spatial deformation relationship between frames. The deformation offset vector refers to the spatial position offset of each pixel or feature point between the image to be detected and the temporal history frame output by the offset prediction network. The deformation offset vector consists of a horizontal offset component and a vertical offset component. Geometric transformation refers to the operation of transforming the spatial coordinates of the historical scene semantic analysis results according to the deformation offset vector, including but not limited to translation, rotation, or affine transformation, so that the semantic information of each spatial position in the historical scene semantic analysis results is accurately aligned with the corresponding spatial position in the image to be detected. Registration refers to mapping the geometrically transformed historical scene semantic analysis results to the coordinate space of the image to be detected, so that the historical scene semantic analysis results and the image to be detected maintain a pixel-by-pixel correspondence in spatial position. The preset temporal buffer refers to a storage area pre-allocated in memory or video memory for temporarily storing the scene semantic analysis results of multiple consecutive frames. The temporal buffer stores the scene semantic analysis results of the current frame and historical frames in chronological order. Global scene entropy refers to the scene semantic complexity metric obtained by globally statistically analyzing the scene semantic analysis results of a certain frame. The global scene entropy reflects the uniformity of the distribution of scene semantic categories in the image of that frame. The more uniform the distribution of scene semantic categories, the higher the global scene entropy; the more uniform the distribution of scene semantic categories, the lower the global scene entropy. The first global scene entropy refers to the global scene entropy value corresponding to the current frame. The second global scene entropy refers to the global scene entropy value corresponding to adjacent historical frames in the temporal buffer. Fluctuation amplitude refers to the absolute value or relative rate of change of the difference between the first and second global scene entropies. The preset fluctuation threshold is a pre-set threshold value for fluctuation amplitude, used to determine whether an abnormal overall image change has occurred in the current frame. The abnormal state signal is an identifier signal generated when the fluctuation amplitude exceeds the preset fluctuation threshold; the abnormal state signal indicates that the scene semantic analysis result of the current frame may be distorted overall due to strong light coverage, lens obstruction, or severe gimbal shaking. Latching refers to fixing the scene semantic analysis result of the most recent valid frame in the temporal buffer in a currently readable state, preventing it from being overwritten by the analysis results of subsequent frames.A valid frame refers to an image frame in the temporal buffer that has not been marked by an abnormal state signal and has been determined to be reliable. Pausing the writing operation of the analysis result of the current frame to the temporal buffer means preventing the scene semantic analysis result of the current frame from being stored in the temporal buffer, so as to avoid contaminating the reliable temporal semantic information already stored in the temporal buffer with semantic information that is determined to be distorted.

[0049] Understandably, this implementation introduces temporal preprocessing before the physical authenticity verification step. By performing offset correction and registration on the scene semantic analysis results of temporal historical frames, it effectively overcomes the spatial position deviation between consecutive frames caused by micro-movements of the camera gimbal or environmental vibrations, ensuring that the semantic information of historical frames can be accurately mapped to the coordinate space of the current frame. Simultaneously, this implementation automatically detects abnormal scenes such as the lens being completely covered by strong light or the overall image collapsing by monitoring the fluctuation amplitude of the global scene entropy between the current frame and historical frames. When an abnormal state is detected, this implementation triggers the latching protection mechanism of the temporal buffer, pausing the writing of distorted frames and maintaining the scene semantic analysis results of the most recent valid frame. This prevents abnormal frames from contaminating the temporal enhanced semantic information relied upon by the subsequent physical authenticity verification step, thereby improving the robustness and stability of the entire risk target detection method in complex outdoor environments.

[0050] For example, to help understand the implementation process of the power transmission channel risk target detection method obtained by combining this embodiment with the above embodiments, please refer to... Figure 2 , Figure 2 A schematic diagram of the overall structure of a power transmission channel risk target detection method is provided. Specifically, the power transmission channel risk target detection method provided in this application typically involves multi-task collaborative reasoning and logical verification in its overall implementation process. First, a preset target detection model is used to quickly reason about the acquired power transmission channel image to be detected in order to locate and preliminarily determine candidate targets that may pose a risk. This target detection model can adopt the You Only Look Once version 8 (YOLOv8, target detection model) with an anchor-free mechanism and a decoupled detection head. Its backbone network extracts multi-scale features and then fuses them through a feature pyramid network and a path aggregation network structure to finally output the bounding boxes and category labels of the candidate targets.

[0051] In one embodiment, to identify the physical authenticity of candidate targets at the model decision level, the system introduces a pre-defined scene semantic analysis model to parse the power transmission channel image. This model is built on a Masked-attention MaskTransformer (Mask2Former) architecture to acquire semantic information of each spatial location in the power transmission channel scene. Simultaneously, the internal decision information used by the scene semantic analysis model to identify the region to be detected as a candidate target is extracted, and this decision information is quantified as decision response features. Specifically, the attention distribution output by the scene semantic analysis model during semantic parsing can be organized to generate a normalized attention heatmap. This heatmap reflects the semantic category response weights assigned by the scene semantic analysis model to each spatial location, representing the degree of attention the model pays during judgment. Then, based on the bounding box coordinates of the candidate targets in the image to be detected obtained during the target detection stage, a local attention sub-map registered with the bounding box coordinate space is extracted from the normalized attention heatmap as a decision response feature. Since the decision response feature is directly derived from the basis for the model's category determination, it is possible to trace whether the model relied on reasonable contextual semantic cues when determining candidate targets.

[0052] For example, to identify the physical authenticity of candidate targets at the model decision level, the system introduces a pre-defined scene semantic analysis model to parse the power transmission channel image. This model is built on a Masked-attention MaskTransformer (Mask2Former) architecture to acquire semantic information of each spatial location in the power transmission channel scene. Simultaneously, it extracts the internal decision information used by the scene semantic analysis model when identifying the image region to be detected as a candidate target, and quantifies this decision information as decision response features. Specifically, the attention distribution output by the scene semantic analysis model during semantic parsing can be organized to generate a normalized attention heatmap. Let the original semantic category response feature map output by the scene semantic analysis model be... ,in For semantic category indexing, This serves as the spatial location index. First, the feature map is linearly normalized to obtain the normalized semantic category response weights for each spatial location: ; in, The scene semantic analysis model targets semantic categories. In spatial location The original activation response value at the location; and Representing the current semantic category respectively The minimum and maximum values ​​of the original activation response at each spatial location within the entire feature map space; That is, the spatial location in the normalized attention heatmap for the semantic category semantic category response weights, A larger numerical value indicates a higher level of attention the model pays to that semantic category at that location. This heatmap reflects the semantic category response weights assigned by the scene semantic analysis model to each spatial location, representing the level of attention the model pays during decision-making. Then, based on the bounding box coordinates of the candidate targets in the image to be detected obtained during the target detection stage, a local attention submap registered with the bounding box coordinate space is extracted from the normalized attention heatmap as the decision response feature. Let the bounding box region of the candidate target be... Define indicator functions When spatial position Located in the bounding box The value is 1 if the time interval is within the interval, and 0 otherwise. The local attention subgraph can then be represented as: ; in, This refers to the local attention subgraph obtained after spatial registration and cropping. This subgraph retains the local decision information relied upon by the scene semantic analysis model when identifying the region as the candidate target. Since this decision response feature directly originates from the basis for the model's category determination, it allows us to trace whether the model relied on reasonable contextual semantic cues when determining the candidate target.

[0053] In one implementation, the physical authenticity verification of candidate targets can be achieved by analyzing the distribution characteristics of semantic category response weights in the decision response features. Specifically, the scene semantic category of the region where the candidate target is located is first determined based on the scene semantic analysis model; then, the concentration of semantic category response weights at each spatial location in the decision response features is calculated, and this concentration is used as a singularity index of the decision semantics. This singularity index reflects whether the scene semantic analysis model's decision basis is highly concentrated on a single semantic category when judging candidate targets. Subsequently, the degree of deviation between the singularity index of the decision semantics and the preset benchmark singularity index corresponding to the scene semantic category is calculated. If the deviation exceeds a preset deviation threshold, it indicates that there is a significant contradiction between the scene semantic analysis model's judgment logic for candidate targets and the expected judgment logic for real physical targets. In this case, the candidate target is determined to have failed the physical authenticity verification and is filtered out. Thus, this embodiment effectively identifies false alarms that rely solely on background texture rather than physical features by verifying the concentration deviation of the decision basis within the model.

[0054] For example, the physical authenticity verification of candidate targets can be achieved by analyzing the distribution characteristics of semantic category response weights in the decision response features. Specifically, firstly, the scene semantic category of the region where the candidate target is located is determined according to the scene semantic analysis model; then, the concentration of semantic category response weights at each spatial location in the decision response features is calculated, and this concentration is used as the singularity index of the decision semantics. This singularity index reflects whether the decision basis of the scene semantic analysis model is highly concentrated on a single semantic category when judging candidate targets. Subsequently, the degree of deviation between the singularity index of the decision semantics and the preset benchmark singularity index corresponding to the scene semantic category is calculated. If the deviation exceeds a preset deviation threshold, it indicates that there is a significant contradiction between the scene semantic analysis model's judgment logic for candidate targets and the expected judgment logic for real physical targets. In this case, the candidate target is determined to have failed the physical authenticity verification and is filtered out.

[0055] Specifically, let the local attention subgraph obtained according to the aforementioned steps be... For this bounding box region Total number of pixels contained First, the mean of the semantic category response weights for each spatial location within the region is calculated to determine the average response level for that region. ; in, Represents the bounding box region Internal space location The corresponding normalized semantic category response weights; This represents the total number of spatial locations contained within the bounding box region; This represents the average of the semantic category response weights for each spatial location within the bounding box region.

[0056] Next, based on this mean, the concentration of the semantic category response weights is calculated; in this embodiment, the variance value is used. To quantify this concentration: ; in, This is the singularity index of the decision semantics. The smaller the value, the more concentrated the semantic category response weight distribution is when the scene semantic analysis model determines the region, that is, the more singular the decision semantics.

[0057] After obtaining the singularity index of decision semantics, a preset benchmark singularity index corresponding to the semantic category of the scenario is obtained. The degree of deviation between the two is calculated. In this embodiment, the absolute value difference is used as the quantification of the degree of deviation: ; in, This refers to the degree of deviation. It is a benchmark singleness index obtained in advance based on statistical analysis of real physical target samples.

[0058] Ultimately, when the degree of deviation Greater than the preset deviation threshold When this occurs, the candidate target is determined to have failed the physical authenticity verification. The specific determination criteria can be expressed as follows: ; Thus, this embodiment effectively identifies false alarms that rely solely on background texture rather than physical features by verifying the concentration deviation of decision-making criteria within the model.

[0059] In another implementation, the physical authenticity verification of candidate targets can also be achieved by measuring the boundary difference of semantic features between the candidate target region and the surrounding scene. In this case, the first semantic feature distribution inside the bounding box of the candidate target is extracted, and simultaneously, according to a preset extraction rule, the second semantic feature distribution of the outer edge region adjacent to the bounding box is extracted. The outer edge region is typically a ring-shaped region surrounding the outer edge of the bounding box. Then, the semantic boundary difference between the first and second semantic feature distributions is calculated. This semantic boundary difference is used to measure the semantic continuity between the candidate target and the surrounding scene. When the semantic boundary difference is lower than a preset difference threshold, the candidate target is determined to have failed the physical authenticity verification. In the actual monitoring scenario of power transmission channels, real physical targets (such as construction machinery) usually have distinctly different semantic features from their attached background (such as grassland), exhibiting a significant semantic difference at the boundary; while false targets formed by spatial artifacts such as water reflections or optical reflections have a smooth and continuous transition between their internal semantic features and the external water background, i.e., a lower semantic boundary difference.

[0060] In the specific implementation of calculating the aforementioned semantic boundary difference, the first semantic feature distribution and the second semantic feature distribution can be normalized to obtain the first semantic probability distribution and the second semantic probability distribution. Then, the cross-entropy between the first semantic probability distribution and the second semantic probability distribution is calculated, and the initial semantic boundary difference is obtained based on this cross-entropy. Based on this, to more accurately capture the spatial structure information of the boundary region, a pre-defined graph convolutional network can be used to perform topological structure reasoning on the boundary region of the bounding box based on the initial semantic boundary difference. The graph convolutional network treats each spatial location within the boundary region as a graph node, constructs edge connections based on the semantic feature similarity between nodes, and aggregates the features of adjacent nodes through graph convolution operations to output the final semantic boundary difference. When training the graph convolutional network, the training set includes physical entity samples and non-physical entity samples; among them, non-physical entity samples refer to spatial artifacts such as water reflections or optical reflections and image regions of the surrounding scene. Such samples typically exhibit smooth transition characteristics of semantic features in the boundary region. By incorporating graph convolutional networks for spatial topology reasoning, it is possible to further distinguish between local pseudo-differences caused by noise and boundary semantic mutations caused by real physical entities, thereby improving the accuracy of physical authenticity verification.

[0061] For example, the physical authenticity verification of candidate targets can also be achieved by measuring the boundary difference of semantic features between the candidate target region and the surrounding scene. In this case, the first semantic feature distribution inside the bounding box of the candidate target is extracted, and simultaneously, according to a preset extraction rule, the second semantic feature distribution of the outer edge region adjacent to the bounding box is extracted. The outer edge region is typically a ring-shaped region surrounding the outer edge of the bounding box. Then, the semantic boundary difference between the first and second semantic feature distributions is calculated. This semantic boundary difference is used to measure the semantic continuity between the candidate target and the surrounding scene. When the semantic boundary difference is lower than a preset difference threshold, the candidate target is determined to have failed the physical authenticity verification. In the actual monitoring scenario of power transmission channels, real physical targets (such as construction machinery) usually have distinctly different semantic features from their attached background (such as grassland), exhibiting a significant semantic difference at the boundary; while false targets formed by spatial artifacts such as water reflections or optical reflections have a smooth and continuous transition between their internal semantic features and the external water background, i.e., a lower semantic boundary difference.

[0062] In the specific implementation of calculating the aforementioned semantic boundary difference, let the internal region of the bounding box containing the candidate target be denoted as . The distribution of the first semantic features corresponding to this region is as follows: At the same time, according to the preset expansion step size Extend the bounding box outward to construct an annular outer edge region surrounding the outer edge of the bounding box. And extract the second semantic feature distribution corresponding to the region. The distribution of the first semantic features is as follows. Second semantic feature distribution After normalization, the first semantic probability distribution is obtained. Second semantic probability distribution .

[0063] Among them, for the semantic category set First semantic probability distribution The middle belongs to the first Probability components of the category for: ; Similarly, the second semantic probability distribution The middle belongs to the first Probability components of the category for: ; In the above formula, Indicates internal area The middle belongs to the first The total number of pixels for semantic features. Indicates internal area Total number of pixels; Indicates the outer edge region The middle belongs to the first The total number of pixels for semantic features. Indicates the outer edge region The total number of pixels.

[0064] Furthermore, the first semantic probability distribution is calculated. With the second semantic probability distribution Cross-entropy between : ; And obtain the initial semantic boundary difference degree based on the cross-entropy. , that is to say .

[0065] Building upon this, in order to more accurately capture the spatial structure information of the boundary region, this embodiment also uses the initial semantic boundary difference degree. The method utilizes a pre-defined graph convolutional network to perform topological reasoning on the boundary region of the bounding box. The graph convolutional network treats each spatial location within the boundary region as a graph node, and the node set is denoted as . And extract each node The corresponding semantic feature vectors. Edge connections are constructed based on the similarity of the semantic feature vectors between nodes, where nodes... and Edge weights between The similarity is measured using the cosine similarity of feature vectors. Graph convolutional networks aggregate feature information from neighboring nodes through multiple layers of graph convolution operations, and combine this with the initial semantic boundary difference. Output the final semantic boundary difference. .

[0066] When training a graph convolutional network, the training set includes physical entity samples and non-physical entity samples. Non-physical entity samples refer to spatial artifacts such as water reflections or optical reflections and image regions of their surrounding scene. These samples typically exhibit smooth transitions in semantic features at boundary regions. By incorporating a graph convolutional network for spatial topology reasoning, it is possible to further distinguish between local pseudo-differences caused by noise and semantic abrupt changes at boundaries caused by real physical entities, thereby improving the accuracy of physical authenticity verification.

[0067] Ultimately, the final semantic boundary difference will be... Difference threshold from preset The comparison and judgment criteria can be expressed as follows: ; When the semantic boundary difference is lower than a preset difference threshold, the candidate target is determined to have failed the physical authenticity verification.

[0068] In one implementation, prior to the aforementioned physical authenticity verification step, a temporal enhancement and abnormal frame blocking step is included. In power transmission channel monitoring scenarios, fluctuations in outdoor light and camera stability, such as the lens being completely covered by strong light or the pan-tilt unit shaking violently, often cause complete distortion of the overall scene semantic analysis of a single frame image. To solve this problem, temporal historical frames from the time series preceding the image to be detected, along with the corresponding historical scene semantic analysis results, can be obtained. The image to be detected and the temporal historical frames are input together into a preset offset prediction network to obtain the deformation offset vector between the image to be detected and the temporal historical frames. Since outdoor monitoring equipment typically experiences slight vibrations, this deformation offset vector is used to characterize the spatial positional deviation between images. Based on the deformation offset vector, the historical scene semantic analysis results are geometrically transformed and registered to the coordinate space of the image to be detected, thereby achieving accurate alignment of the semantic information of historical frames in the coordinate space of the current frame. The registered historical scene semantic analysis results are written into a preset temporal buffer. Subsequently, the first global scene entropy of the current frame in the time-series buffer is calculated, and the second global scene entropy of adjacent historical frames in the time-series buffer is obtained. The first and second global scene entropies are used to reflect the overall uniformity of the distribution of scene semantic categories in the corresponding frames. If the fluctuation amplitude of the first global scene entropy relative to the second global scene entropy exceeds a preset fluctuation threshold, it indicates that the image of the current frame has undergone a drastic overall change, generating an abnormal state signal. Once the abnormal state signal is triggered, the system latches the scene semantic analysis result of the most recent valid frame in the time-series buffer and suspends the writing operation of the analysis result of the current frame to the time-series buffer. This latching mechanism avoids the contamination of the stable time-series historical scene information by erroneous semantic information of the current frame that is distorted by strong light or other interference, effectively improving the stability and reliability of the entire power transmission channel risk target detection method in complex and harsh outdoor environments.

[0069] In one implementation, the power transmission channel risk target detection method provided in this application has the following overall architecture: Figure 3 As shown in the figure, this method first uses You Only Look Once version 8 (YOLOv8, an object detection model) as the base detector. YOLOv8 employs an anchor-free mechanism and a decoupled detection head, improving detection accuracy while maintaining high inference speed. In the power transmission channel hazard detection task, this detector is responsible for quickly locating and classifying potential hazard targets from the input image, outputting the target category and its bounding box information, providing basic detection results for subsequent semantic filtering.

[0070] After obtaining preliminary target detection results, the system introduces a pre-defined scene semantic analysis model to perform pixel-level semantic understanding on the same image to be detected. This model is built on a Masked-attention Mask Transformer (Mask2Former) architecture to achieve pixel-level semantic understanding and temporal stability enhancement for power transmission scenarios. This scene semantic analysis module generates a highly reliable temporal scene semantic map through a three-pronged strategy: customized semantic labels, robustness optimization in complex environments, and temporal stability optimization. Specifically, the temporal stability optimization uses a sliding window mechanism to fuse segmentation results from multiple consecutive frames, dynamically correcting single-frame errors and improving the spatiotemporal consistency of scene understanding. Figure 4 As shown, this temporal scene semantic map can accurately reflect the distribution and changes of various elements (such as the sky, water surface, towers, etc.) in the power transmission channel scene at different time points, providing reliable prior information with spatiotemporal continuity for subsequent physical logic verification.

[0071] Furthermore, to address the issue of discrepancies between target detection results and the physical logic of real-world scenes, the system constructs a semantic logic filtering module. (Refer to...) Figure 5 This module constructs a matching rule base for scenarios and hazards based on common-sense physical logic, and performs semantic consistency verification on YOLOv8 detection results. Specific filtering rules include: for airspace, when the sky area within the detection box of fire or a specified mechanical target is greater than 70%, it is judged as a false alarm and filtered; for water areas, when the water area within the detection box of smoke, fire, or a specified mechanical target is greater than 70%, it is judged as a false alarm and filtered; for the interior of towers, when the tower area within the detection box of a specified mechanical target is greater than 70%, it is judged as a false alarm and filtered. Through the above pixel area ratio logic judgment based on common-sense physical logic, spatial artifacts such as water reflections and cloud reflections can be effectively filtered out from the alarm list.

[0072] To achieve high-precision semantic segmentation results to support subsequent semantic filtering logic, this paper employs a multi-task joint optimization strategy in the semantic segmentation module, comprehensively considering pixel-level classification accuracy and region boundary consistency. Specifically, the total loss function is a weighted combination of Dice loss and cross-entropy loss. ; in, This represents the total loss function value of the semantic segmentation model; This represents pixel-level cross-entropy loss, which aims to optimize the classification confidence of pixels in each category, and has good optimization sensitivity, especially for key hazard categories with a small proportion (such as fireworks and construction machinery). This represents the Dice loss, which aims to effectively alleviate the class imbalance problem and improve the segmentation accuracy of the target boundary region by maximizing the Dice coefficient between the predicted mask and the true mask. and These represent the weighting coefficients of the cross-entropy loss and the Dice loss, respectively, used to adjust the relative weights of the two types of loss terms in the joint optimization.

[0073] Dice loss The specific calculation method is shown in the following formula: ; In this formula, This represents the total number of semantic categories to be segmented; Indicates the current category index, with a value range from arrive ; Represents the spatial location index of a pixel in an image; Indicates spatial location Location, Target Category The actual label is usually 0 or 1; The scene semantic analysis model represents the spatial location. Location, Target Category The predicted probability value. This loss function strongly constrains the model's boundary segmentation accuracy by calculating the overlap between the predicted probability and the true label in terms of spatial scale.

[0074] During training, to address the issue of sparse samples for key hazard categories in power transmission scenarios, this embodiment further introduces category-weighted cross-entropy loss, i.e., the cross-entropy loss... The loss terms of each category are inversely weighted according to their frequency of occurrence in the training set. By increasing the error contribution weight of low-frequency target categories, the model's ability to identify small targets and rare hidden danger categories is enhanced.

[0075] In the engineering implementation of a specific embodiment, a dedicated semantic segmentation dataset for power transmission channels was constructed to achieve accurate semantic understanding of the transmission channel scene. This dataset systematically collected 1464 high-resolution images from actual power transmission line inspection sites, covering various typical geographical and environmental scenes such as mountainous areas, forest areas, and urban and rural streets. All images were semantically annotated at the pixel level by professional annotators according to unified standards, with nine annotation categories: background, sky, towers, conductors, water surface, road surface, vegetation, streetlights, watermarks, and buildings. Model training experiments were conducted using an Intel(R) Xeon(R) Platinum 8458P processor and an NVIDIA A800 graphics card, employing the PyTorch 2.1.2 deep learning framework. The semantic segmentation module used the Mask2Former architecture as its base, with the Swin Transformer as the backbone network, an embedding dimension of 96, and a window size of 7. A transfer learning strategy was adopted, using pre-trained weights from the Cityscapes dataset, which has similarities to the power transmission channel scene, for model initialization. The optimizer used is Adaptive Moment Estimation with Weight Decay (AdamW), with an initial learning rate of 1×10⁻⁶. -4 The backbone network learning rate was set to 0.1 times the initial learning rate, and a poly strategy was used for learning rate decay. The total number of training iterations was 15,000, and the batch size was 4.

[0076] Furthermore, to verify the effectiveness of each module, this embodiment designed an ablation experiment of the system, with three configurations: Configuration A is a basic YOLOv8 model without any semantic optimization; Configuration B is a model with a static semantic map; and Configuration C is the method of this invention with a fully integrated temporal semantic logic filtering mechanism. Experimental results show that Configuration C significantly reduces the false alarm rate compared to Configuration A, while effectively improving the recall rate, proving that the temporal semantic logic filtering mechanism significantly suppresses invalid alarms while ensuring the detection capability of real potential problems. Configuration B has a relatively high false alarm rate, and Configuration C achieves further significant optimization on this basis, verifying the effectiveness of the temporal fusion mechanism in correcting single-frame segmentation jitter and reducing false alarms and missed filtering caused by instantaneous errors. Figure 6 The visualization results further confirm that, compared to the basic YOLOv8 model and the contrastive model (SegFormer), the method in this embodiment can completely eliminate false positives and retain all valid hazards in near, mid, and far-field testing environments, demonstrating the best detection performance. Through the above process, the system effectively filters out non-realistic risk targets and ultimately outputs real hazard alarm results that meet the scene logic.

[0077] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the transmission channel risk target detection method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0078] This application also provides a power transmission channel risk target detection device, please refer to... Figure 7 The power transmission channel risk target detection device includes: Detection module 10 is used to acquire the image to be detected of the power transmission channel, perform target detection on the image to be detected, and obtain at least one candidate target; The enhancement module 20 is used to obtain the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and to perform temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results. The extraction module 30 is used to perform semantic parsing on the image to be detected based on a preset scene semantic analysis model, and extract the decision response features of the region where the candidate target is located by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. The verification module 30 is used to verify the physical authenticity of the candidate targets based on the decision response characteristics, so as to filter out non-real risk targets and obtain risk target detection results.

[0079] The power transmission channel risk target detection device provided in this application, employing the power transmission channel risk target detection method described in the above embodiments, can solve the technical problem of low operation and maintenance efficiency of power transmission channels due to frequent false alarms. Compared with the prior art, the beneficial effects of the power transmission channel risk target detection device provided in this application are the same as those of the power transmission channel risk target detection method provided in the above embodiments, and other technical features in the power transmission channel risk target detection device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0080] This application provides a power transmission channel risk target detection device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the power transmission channel risk target detection method in the first embodiment described above.

[0081] The following is for reference. Figure 8The diagram illustrates a structural schematic suitable for implementing the transmission channel risk target detection device in the embodiments of this application. The transmission channel risk target detection device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, tablets, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), vehicle terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital televisions and desktop computers. Figure 8 The power transmission channel risk target detection device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0082] like Figure 8 As shown, the power transmission channel risk target detection device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the power transmission channel risk target detection device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the power transmission channel risk target detection equipment to communicate wirelessly or wiredly with other devices to exchange data. Although power transmission channel risk target detection equipment with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0083] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0084] The power transmission channel risk target detection device provided in this application, employing the power transmission channel risk target detection method described in the above embodiments, can solve the technical problem of low operation and maintenance efficiency of power transmission channels caused by frequent false alarms. Compared with the prior art, the beneficial effects of the power transmission channel risk target detection device provided in this application are the same as those of the power transmission channel risk target detection method provided in the above embodiments, and other technical features of this power transmission channel risk target detection device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0085] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0087] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the power transmission channel risk target detection method in the above embodiments.

[0088] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0089] The aforementioned computer-readable storage medium may be included in the power transmission channel risk target detection equipment; or it may exist independently and not be assembled into the power transmission channel risk target detection equipment.

[0090] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the power transmission channel risk target detection device, cause the power transmission channel risk target detection device to: acquire a target image of the power transmission channel, perform target detection on the target image, and obtain at least one candidate target; The historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected are obtained, and the historical scene semantic analysis results are temporally registered and abnormal frames are filtered to obtain temporally stable scene semantic analysis results. Based on a preset scene semantic analysis model, the image to be detected is semantically parsed, and the decision response features of the region where the candidate target is located are extracted by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. Based on the decision response characteristics, the physical authenticity of the candidate targets is verified to filter out non-real risk targets and obtain the risk target detection results.

[0091] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0092] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0093] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0094] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described power transmission channel risk target detection method. This solves the technical problem of low operation and maintenance efficiency of power transmission channels due to frequent false alarms. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the power transmission channel risk target detection method provided in the above embodiments, and will not be repeated here.

[0095] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the power transmission channel risk target detection method described above.

[0096] The computer program product provided in this application can solve the technical problem of low operation and maintenance efficiency of power transmission channels caused by frequent false alarms. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the power transmission channel risk target detection method provided in the above embodiments, and will not be repeated here.

[0097] All acquisition of signals, information, or actions in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the relevant device owner.

[0098] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the scope of protection of this application.

Claims

1. A method for detecting risk targets in power transmission channels, characterized in that, The method includes: Acquire an image of the power transmission channel to be detected, perform target detection on the image to be detected, and obtain at least one candidate target; The historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected are obtained, and the historical scene semantic analysis results are temporally registered and abnormal frames are filtered to obtain temporally stable scene semantic analysis results. Based on a preset scene semantic analysis model, the image to be detected is semantically parsed, and the decision response features of the region where the candidate target is located are extracted by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. The step of extracting the decision response features of the region where the candidate target is located includes: The image to be detected is semantically parsed according to a preset scene semantic analysis model, and the semantic parsing results are organized to generate a normalized attention heatmap. Based on the bounding box coordinates of the candidate target in the image to be detected, a local attention sub-map that is registered with the bounding box coordinate space is extracted from the normalized attention heatmap as the decision response feature; Based on the decision response characteristics, the physical authenticity of the candidate targets is verified to filter out non-real risk targets and obtain the risk target detection results; The normalized attention heatmap includes the semantic category response weights assigned by the scene semantic analysis model to each spatial location in the image to be detected. The step of verifying the physical authenticity of the candidate target based on the decision response features includes: Based on the scene semantic analysis model, the scene semantic category of the region where the candidate target is located in the image to be detected is determined; Based on the semantic category response weights of each spatial location in the decision response features, the concentration of the semantic category response weights is calculated, and the concentration is used as the singleness index of decision semantics on which the scene semantic analysis model determines the region as the candidate target. Calculate the degree of deviation between the singleness index of the decision semantics and the preset benchmark singleness index corresponding to the scene semantic category; When the deviation exceeds a preset deviation threshold, the candidate target is determined to have failed the physical authenticity verification.

2. The method as described in claim 1, characterized in that, The step of verifying the physical authenticity of the candidate target based on the decision response characteristics includes: Extract the first semantic feature distribution inside the bounding box of the candidate target; According to the preset extraction rules, the distribution of second semantic features of the outer edge region adjacent to the bounding box is extracted; Calculate the semantic boundary difference degree between the first semantic feature distribution and the second semantic feature distribution. The semantic boundary difference degree is used to measure the transition continuity between the candidate target and the surrounding scene at the semantic level. When the semantic boundary difference is lower than a preset difference threshold, the candidate target is determined to have failed the physical authenticity verification.

3. The method as described in claim 2, characterized in that, The step of calculating the semantic boundary difference between the first semantic feature distribution and the second semantic feature distribution includes: The first semantic feature distribution and the second semantic feature distribution are normalized respectively to obtain the first semantic probability distribution and the second semantic probability distribution; Calculate the cross-entropy between the first semantic probability distribution and the second semantic probability distribution, and calculate the initial semantic boundary difference degree based on the cross-entropy; Based on the initial semantic boundary difference, the boundary region of the bounding box is subjected to topological structure reasoning according to the preset graph convolutional network, and the final semantic boundary difference is output. The preset graph convolutional network is trained on a training set containing physical entity samples and non-physical entity samples, and the boundary semantic features of the non-physical entity samples have smooth transition characteristics.

4. The method as described in claim 1, characterized in that, The steps of obtaining the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and performing temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results include: Obtain the temporal historical frame corresponding to the image to be detected, and obtain the historical scene semantic analysis result corresponding to the temporal historical frame; The image to be detected and the temporal history frame are input into a preset offset prediction network to obtain the deformation offset vector between the image to be detected and the temporal history frame. Based on the deformation offset vector, the historical scene semantic analysis result is geometrically transformed, the historical scene semantic analysis result is registered to the coordinate space of the image to be detected, and the registered historical scene semantic analysis result is written into a preset temporal buffer. Calculate the first global scene entropy of the current frame in the time-series buffer, and obtain the second global scene entropy of adjacent historical frames in the time-series buffer; When the fluctuation of the first global scene entropy relative to the second global scene entropy exceeds a preset fluctuation threshold, an abnormal state signal is generated, the scene semantic analysis result of the most recent valid frame in the time-series buffer is latched, and the writing operation of the analysis result of the current frame to the time-series buffer is suspended.

5. A power transmission channel risk target detection device, characterized in that, The device includes: The detection module is used to acquire the image to be detected of the power transmission channel, perform target detection on the image to be detected, and obtain at least one candidate target; The enhancement module is used to obtain the historical scene semantic analysis results of the temporal historical frames corresponding to the image to be detected, and to perform temporal registration and abnormal frame filtering on the historical scene semantic analysis results to obtain temporally stable scene semantic analysis results. The extraction module is used to perform semantic parsing on the image to be detected based on a preset scene semantic analysis model, and extract the decision response features of the region where the candidate target is located by combining the time-stable scene semantic analysis results. The decision response features are used to characterize the internal decision information on which the scene semantic analysis model determines the region as the candidate target. The step of extracting the decision response features of the region where the candidate target is located includes: The image to be detected is semantically parsed according to a preset scene semantic analysis model, and the semantic parsing results are organized to generate a normalized attention heatmap. Based on the bounding box coordinates of the candidate target in the image to be detected, a local attention sub-map that is registered with the bounding box coordinate space is extracted from the normalized attention heatmap as the decision response feature; The verification module is used to verify the physical authenticity of the candidate targets based on the decision response characteristics, so as to filter out non-real risk targets and obtain risk target detection results. The normalized attention heatmap includes the semantic category response weights assigned by the scene semantic analysis model to each spatial location in the image to be detected. The step of verifying the physical authenticity of the candidate target based on the decision response features includes: Based on the scene semantic analysis model, the scene semantic category of the region where the candidate target is located in the image to be detected is determined; Based on the semantic category response weights of each spatial location in the decision response features, the concentration of the semantic category response weights is calculated, and the concentration is used as the singleness index of decision semantics on which the scene semantic analysis model determines the region as the candidate target. Calculate the degree of deviation between the singleness index of the decision semantics and the preset benchmark singleness index corresponding to the scene semantic category; When the deviation exceeds a preset deviation threshold, the candidate target is determined to have failed the physical authenticity verification.

6. A power transmission channel risk target detection device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the transmission channel risk target detection method as described in any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the power transmission channel risk target detection method as described in any one of claims 1 to 4.

8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the power transmission channel risk target detection method as described in any one of claims 1 to 4.

Images