Safety protection wearing detection method for 0.4 kV low-voltage uninterruptible power operation personnel based on image recognition

Abstract

The invention discloses a safety protection wearing detection method for 0.4 kV low-voltage uninterruptible power operation personnel based on image recognition, and relates to the technical field of electric power safety monitoring and intelligent image recognition. Acquiring continuous image frame data of a working area where the low-voltage uninterruptible power operation personnel are located in real time and preprocessing the continuous image frame data; by introducing a thermal disturbance characteristic index extraction and deep learning judgment mechanism, response weights of a recognition network to different scale characteristics are dynamically adjusted, and intelligent perception and recognition robustness enhancement of image disturbance in a high-temperature or strong-sunlight environment are realized. According to the method, identification interference caused by image edge jitter and structure distortion due to hot air disturbance can be effectively inhibited, dependence on high-frequency details is reduced, and the identification capability of low and medium frequency structures is enhanced, so that the accuracy and stability of safety protection wearing detection of operators are improved, and key potential safety hazards such as wearing misjudgment and missing detection are avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power safety monitoring, in particular to a 0.4kV low-voltage live-line working personnel safety protection wearing detection method based on image recognition. BACKGROUND

[0002] The "0.4kV low-voltage live-line working personnel safety protection wearing detection based on image recognition" refers to a method for 0.4kV low-voltage distribution line live-line working scene, using artificial intelligence image recognition technology to automatically detect and determine the wearing compliance of various safety protection equipment (including but not limited to safety helmet, protective mask, insulating gloves, insulating shoes, insulating clothes, etc.) worn by field workers. The detection method acquires image data of workers in real time by laying image acquisition devices (such as fixed or portable cameras) at the working site, and combines a target detection model based on deep learning (such as a convolutional neural network CNN) to identify and analyze the personnel protection wearing state in the image, to determine whether all protective equipment is worn, whether the wearing position is correct, and whether there is a non-standard wearing or missing safety risk. This technical means can be used for pre-inspection control before work, or for dynamic monitoring during work, and has the advantages of non-contact, intelligence, real-time, etc., significantly improving the safety control ability and work compliance rate in the process of low-voltage live-line work.

[0003] The existing technology has the following disadvantages: In a high-temperature or strong-sunlight working environment, especially when there are heat source obstacles (such as transformers, power distribution cabinets) or bare metal conductors between the camera and the target worker, the density gradient caused by the temperature difference in the ambient air will cause nonlinear disturbance to the light propagation path, forming a local thermal air disturbance phenomenon. This disturbance will cause a slight "water ripples" type distortion in the picture acquired by the image acquisition device, that is, dynamic jitter, deformation or blur phenomenon occurs in the target contour edge in the image. Since this kind of distortion mainly comes from the change of the refractive index of the air medium itself, it cannot be effectively inhibited by conventional image enhancement or filtering means, and it is easy to cause false matching in the personnel positioning and key part feature extraction process of the image recognition model based on deep learning, which shows as key point drift, shielding misjudgment, equipment recognition failure, etc. Although this problem is difficult to detect in a normal temperature static scene, it has a high probability of inducing conditions in typical low-voltage live-line working scenes such as summer, outdoor, bare distribution environment, or near high-temperature electrical equipment. Once it happens, it will seriously affect the recognition accuracy of the safety protection equipment wearing state of the workers, and weaken the judgment ability of the system on the work compliance, which may cause "wearing missing missed detection" and "illegal work without warning" and other major safety risks.

[0004] Therefore, a 0.4kV low-voltage live-line working personnel safety protection wearing detection method based on image recognition is needed. SUMMARY

[0005] In view of the problems in the prior art, the present application provides a 0.4kV low-voltage live-line working personnel safety protection wearing detection method based on image recognition, which can effectively suppress the recognition interference caused by image edge jitter and structure distortion due to hot air disturbance, reduce the dependence on high-frequency details, enhance the mid-low frequency structure recognition capability, thereby improving the accuracy and stability of the working personnel safety protection wearing detection, and avoiding key safety hazards such as wearing misjudgment and missed detection. The specific technical solutions are as follows: A 0.4kV low-voltage live-line working personnel safety protection wearing detection method based on image recognition, comprising the following steps: Step 1: Real-time acquisition of continuous image frame data of the working area where the low-voltage live-line working personnel is located through the industrial-grade high-definition camera equipment deployed at the working site; Step 2: Preprocessing of the real-time acquired image data, and extracting key indicators reflecting the hot air disturbance characteristics of the local image from the preprocessed image data, and comprehensively analyzing the extracted key indicators to quantify the image structure abnormalities caused by the hot air disturbance phenomenon; Step 3: Inputting the disturbance characteristic indicators after quantitative analysis into the deep learning model trained in advance using historical image data, and comprehensively analyzing the input disturbance characteristics through the model to determine whether there is a local hot disturbance phenomenon in the current image; Step 4: When it is determined that the image data has hot air disturbance, dynamically adjusting the feature response parameters of the recognition network, reducing the recognition sensitivity to high-frequency detail features, and at the same time improving the recognition sensitivity of the mid-low frequency region, to suppress the amplification interference of the edge discrimination result caused by the hot disturbance effect.

[0006] Preferably, the specific steps of step 1 are as follows: Step 11: Starting the image acquisition channel through the connected camera equipment, and starting to acquire continuous image frame data of the target working area based on the set time interval or trigger mechanism; Step 12: The camera equipment transmits the acquired image frames to the on-site computing module in the form of image stream, and uses a high-speed interface protocol to ensure the integrity and low delay of image transmission; Step 13: The image stream is uniformly cached in the computing module and sequentially arranged according to the time stamp to construct an image frame sequence reflecting the dynamic characteristics of the working process; Step 14: Real-time scheduling management of the image frame sequence in the cache to provide a continuous frame input interface for subsequent recognition tasks.

[0007] Preferably, the specific steps of step 2 are as follows: Step 21: the key indicators of extraction include the time sequence variance of the texture features of the target region in the continuous frames and the fluctuation degree of the image local sharpness focus over time; Step 22: the time sequence variance of the texture features of the target region in the continuous frames and the fluctuation degree of the image local sharpness focus over time are comprehensively analyzed under the detection window to generate the time sequence texture disturbance indicator and the boundary continuity disturbance indicator respectively; Step 23: the image structure abnormalities caused by the hot air disturbance phenomenon are quantified by the time sequence texture disturbance indicator and the boundary continuity disturbance indicator.

[0008] Preferably, in the step 22, the specific steps of comprehensively analyzing the time sequence variance of the texture features of the target region in the continuous frames under the detection window to generate the time sequence texture disturbance indicator are as follows: In the detection window, the texture direction features of the target region are extracted from the continuous image frames, the direction gradient structure entropy is used to represent the complexity of the texture direction distribution of each frame, for each frame of the image, the gradient direction set is set , the distribution probability of the texture gradient intensity in each direction is respectively counted, and the direction structure entropy value is calculated, the calculation expression is as follows: In the formula, is the texture gradient probability distribution value of the frame in the direction , is a set of preset gradient direction angle set, is the direction structure entropy value; After obtaining the sequence of direction structure entropy values of the continuous frames, the change degree between the adjacent frames in the sequence is analyzed, the fluctuation intensity of the texture of the target region in the time dimension is evaluated, and the time sequence texture disturbance indicator for identifying the local disturbance state of the image is generated, the calculation expression is as follows: , in the formula, is the time sequence texture disturbance indicator, is the number of image frames, is the direction structure entropy jump amplitude, and the calculation formula is as follows: , is the direction structure entropy value of the previous frame image, is the direction structure entropy value of the previous frame image, is the high-frequency texture coverage factor in the frame.

[0009] Preferably, in the step 22, the specific steps of comprehensively analyzing the time sequence variance of the texture features of the target region in the continuous frames under the detection window to generate the time sequence texture disturbance indicator are as follows: Under the detection window, after acquiring the continuous frame data of the target region of the image, the edge detection algorithm is used to extract the boundary contour information in each frame of the image, and the boundary curvature response value at the pixel level is calculated for the boundary region of each frame of the image, and the calculation expression is as follows: , wherein, is the boundary curvature response intensity, indicating the “bending degree” intensity of the pixel point on the boundary structure in the image, is the pixel point coordinate, is the Laplace second-order derivative of the image grayscale, indicating the second-order change rate of the image grayscale at the pixel point , is the change rate of the grayscale of each pixel point in the image, indicating the “change trend” of the image, is the image first-order gradient amplitude; Based on the curvature response graph of the continuous frame image, the curvature response difference value between adjacent frames at the same boundary position is extracted, and the boundary continuity disturbance index is calculated by combining the region weighting mechanism, and the calculation expression is as follows: , wherein, is the boundary continuity disturbance index, is the curvature difference value set, , wherein, is the inter-frame curvature difference value, and the calculation formula is as follows: , is the boundary curvature response intensity of the pixel point in the image of the frame, is the boundary curvature response intensity of the pixel point in the image of the frame, is the maximum value of the curvature difference value of all boundary points in the same image region between continuous frames, is the effective boundary region in the curvature response definition domain, is the position weighting coefficient, is the small change amount along the horizontal direction of the image, is the small change amount along the vertical direction of the image.

[0010] Preferably, it further comprises: The time-series texture disturbance index and the boundary continuity disturbance index after quantitative analysis are input into a deep learning model trained in advance using historical image data, a disturbance risk coefficient is generated through the model, and whether the current image has a local thermal disturbance phenomenon is judged based on the disturbance risk coefficient.

[0011] Preferably, the disturbance risk coefficient is compared with the pre-set disturbance risk coefficient reference threshold to determine whether the current image has local thermal disturbance phenomenon, and the judgment logic is as follows: If the disturbance risk coefficient is greater than the pre-set disturbance risk coefficient reference threshold, it is determined that the current image has local thermal disturbance phenomenon; if the disturbance risk coefficient is greater than the pre-set disturbance risk coefficient reference threshold, it is determined that the current image does not have local thermal disturbance phenomenon.

[0012] Preferably, when it is determined that the image data has thermal air disturbance, the feature response parameters of the recognition network are dynamically adjusted to reduce the recognition sensitivity to high-frequency detail features and improve the recognition sensitivity of medium and low-frequency regions. The specific steps are as follows: When it is determined that the image data has thermal air disturbance, a high-frequency suppression factor for regulating the high-frequency feature response sensitivity is constructed according to the deviation relationship between the disturbance risk coefficient and the pre-set disturbance risk coefficient reference threshold, and the formula is as follows: , wherein, is the disturbance enhancement response coefficient, is the disturbance risk coefficient, is the disturbance risk coefficient reference threshold, is the hyperbolic tangent function, is the high-frequency feature suppression factor; In order to ensure that the recognition process does not fail as a whole due to high-frequency suppression, and to improve the discrimination ability of stable regions under disturbance scene, a medium and low-frequency enhancement factor is constructed, and the calculation expression is as follows: , wherein, is the medium and low-frequency enhancement factor, is the low-frequency enhancement coefficient, is the disturbance response smoothing factor; The high-frequency feature suppression factor and the medium and low-frequency enhancement factor are applied to the multi-scale feature fusion module in the recognition network respectively to realize dynamic weighting of different scale feature responses, and finally the fused output feature map is output, and the expression is as follows: , wherein, is the fused output feature map, is the high-frequency feature response component, is the medium-frequency feature response component, is the low-frequency feature response component.

[0013] Preferably, the pre-processing of the real-time acquired image data in step 2 specifically includes the following steps: Firstly, gray scale conversion is performed to convert the color image into a single-channel gray scale image to reduce the calculation complexity and highlight the structural information of the target region; Then, a denoising filter algorithm is applied to suppress sensor noise, dust point interference or background clutter and improve image edge definition. Finally, contrast adaptive adjustment is implemented to enhance the edge and texture contrast under the premise of ensuring that local features are not lost Compared with the prior art, the present application has the following beneficial effects: The present application can effectively improve the stability and accuracy of the recognition system in typical disturbance environments such as high temperature or strong sunlight, especially in the case of image local edge jitter, structure distortion and other interference caused by hot air disturbance. Through feature index quantization and intelligent discrimination, real-time perception of the disturbed image is realized, and the response weight of the recognition network to multi-scale features is further dynamically adjusted, thereby significantly reducing the dependence on high-frequency noise, strengthening the recognition ability of low-frequency stable structure, avoiding key point misjudgment and wearing misjudgment caused by edge drift, ensuring that the recognition result of the operating personnel protection state has high robustness and engineering adaptability, and effectively solving the key safety risk problems such as "missing detection of wearing loss" or "no alarm for illegal operation". BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or prior art description. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual proportion.

[0015] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] It should be understood that when used in the present specification and the appended claims, the terms "comprise" and "include" indicate the presence of described features, integers, steps, operations, elements, and / or components, but do not exclude one or more other features, integers, steps, operations, elements, components, and / or sets thereof.

[0018] It should also be understood that the terms used in the specification of the present application are only for the purpose of describing particular embodiments and are not intended to limit the present application. As used in the specification and the appended claims of the present application, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0019] It should be further understood that the term "and / or" used in the specification and the appended claims of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0020] Embodiments Please refer to Figure 1 The present application provides an image recognition-based 0.4kV low-voltage live-line working personnel safety protection and wearing detection method, comprising the following steps: Step 1, through the industrial-grade high-definition camera equipment deployed in the working site, real-time acquisition of continuous image frame data of the working area where the low-voltage live-line working personnel are located; Through the industrial-grade high-definition camera equipment deployed in the working site, real-time acquisition of continuous image frame data of the working area where the low-voltage live-line working personnel are located can be realized. The specific commonly used industrial-grade high-definition camera equipment includes GigE industrial camera, USB3.0 high-speed industrial camera, PoE network high-definition camera, area array / linear array industrial CCD or CMOS camera, which has the characteristics of high resolution (such as 1920x1080, 2592x2048), high frame rate (≥30fps), wide dynamic range (WDR) and anti-strong light interference, etc., and is suitable for outdoor complex power operation environment. This kind of equipment is usually connected with edge computing device or field industrial computer through standard industrial image acquisition protocol (such as GigE Vision, GenICam), and uses image stream channel to continuously acquire dynamic image frames of the working personnel. The main role of the acquisition data is to provide high-quality and stable input source for subsequent image recognition-based working personnel positioning, safety protection equipment wearing detection, abnormal behavior judgment and other algorithms, to ensure the real-time, continuous and accurate perception ability of the identification system to the working state, which is the basic link to build the low-voltage live-line working intelligent identification system.

[0021] The acquisition device in this embodiment should have the functions of automatic exposure, wide dynamic range, automatic focusing, etc., to ensure that clear and detailed image information can be effectively obtained under different lighting conditions, especially in high-temperature and strong-light environment.

[0022] The purpose of this step is to obtain the basic data required for subsequent disturbance detection, feature extraction and intelligent recognition processing. The camera should be reasonably deployed at different angles to reduce the reflection interference of high-temperature equipment or metal components and obtain the best visual coverage effect. This step directly determines the input quality of the subsequent recognition system and ensures accurate and reliable subsequent analysis.

[0023] By deploying industrial-grade high-definition camera equipment at the work site, real-time continuous image frame data of the work area of the low-voltage uninterrupted power operation personnel is collected, and the specific steps are as follows: First, the image acquisition channel is opened through the connected industrial-grade high-definition camera equipment, and continuous image frame data of the target work area is obtained based on the set time interval or trigger mechanism (such as timed acquisition or operation start signal); second, the camera equipment transmits the collected image frames to the on-site computing module in the form of image stream, uses high-speed interface protocol (such as GigE Vision or USB3 Vision) to ensure the integrity and low delay of image transmission; third, the image stream is uniformly cached in the computing module and sequentially arranged according to the time stamp to build an image frame sequence reflecting the dynamic characteristics of the operation process; finally, the image frame sequence in the cache is managed in real time to provide a continuous frame input interface for subsequent recognition tasks, and to realize stable perception and timing capture of information such as the posture change and protective clothing state of the operation personnel.

[0024] The real-time acquired image data is preprocessed, and key indicators reflecting the local hot air disturbance characteristics of the image are extracted from the preprocessed image data. The extracted key indicators are comprehensively analyzed to quantify the image structure abnormalities caused by the hot air disturbance phenomenon. Among them, the real-time acquired image data is preprocessed, which specifically includes but is not limited to the following operations: First, the gray scale conversion is performed to convert the color image into a single-channel gray scale image to reduce the computational complexity and highlight the structural information of the target area; Subsequently, a denoising filter algorithm (such as median filter, Gaussian filter) is applied to suppress sensor noise, dust point interference or background clutter, and improve image edge definition; then the histogram equalization is used to enhance the uniformity of the overall brightness distribution of the image, so that the detail area is more easily identified in the dark or bright scene; Finally, the contrast adaptive adjustment (such as CLAHE algorithm) is implemented to enhance the edge and texture contrast on the premise of not losing local features, thereby improving the feature extraction capability of the subsequent recognition model for safety helmets, gloves, work clothes and other equipment.

[0025] The core role of the preprocessing process is to improve image quality and feature separability, to the greatest extent eliminate the adverse effects of environmental light changes, image blur or abnormal collection on intelligent recognition accuracy, and to ensure the stability and robustness of downstream recognition tasks in complex outdoor operating environments.

[0026] The role of this step is to improve the quality and consistency of the original data, making subsequent feature extraction more accurate and effective, and laying a solid foundation for thermal disturbance analysis.

[0027] Step 2, from the pre-processed image data, extract key indicators reflecting the local thermal air disturbance features of the image, the extracted indicators include the time series variance of the target region texture features in consecutive frames and the fluctuation degree of the image local sharpness focus over time, and the time series variance of the target region texture features in consecutive frames and the fluctuation degree of the image local sharpness focus over time are analyzed under the detection window to generate time series texture disturbance indicators and boundary continuity disturbance indicators, respectively, and the image structure anomalies caused by thermal air disturbance phenomena are quantified through time series texture disturbance indicators and boundary continuity disturbance indicators.

[0028] When the target region texture features in consecutive frames show high-frequency, light-amplitude and irregular time series variance fluctuations, it can effectively indicate that there is a local thermal air disturbance feature in the image. This is because thermal air disturbance is essentially caused by local temperature gradient in the air, which in turn causes uneven changes in refractive index, and then causes small deflection of light propagation path in a very short time scale. This disturbance does not cause large changes in the overall structure of the image, but it can cause slight random deformation at the texture level, especially in areas with rich edges or dense details, which is manifested as slight misalignment or blurred reconstruction of texture patterns between multiple frames. Therefore, if texture analysis (such as using local binary pattern LBP or Gabor filter) is performed on this area and the variation of texture features is calculated on the time axis, a non-periodic, unstable and small amplitude jitter pattern will often be observed. This is significantly different from the normal texture change pattern caused by object movement or light changes, and has a unique disturbance fingerprint. High-frequency light-amplitude fluctuations are the projection of thermal disturbance randomness and locality in image space, so this indicator has high thermal disturbance sensitivity and discrimination value, and is an important auxiliary basis for thermal disturbance feature quantification.

[0029] The specific steps for generating time series texture disturbance indicators by comprehensively analyzing the time series variance of the target region texture features in consecutive frames under the detection window are as follows: Within the detection window, extract the texture direction features of the target region from consecutive image frames, use Directional Gradient Structure Entropy (DGSE) to represent the complexity of the texture direction distribution of each frame, and for each frame, set the gradient direction set (e.g., 0°, 45°, 90°, 135°), respectively, statistically analyze the distribution probability of texture gradient intensity in each direction, and calculate the directional structure entropy value. The calculation expression is as follows: In the formula, It is the first Frame in direction The texture gradient probability distribution value on the image represents the probability distribution value of the texture gradient on the image. In the frame, along the direction within the detection window The gradient intensity pixel value is the proportion of the total gradient value in the region. It is used to measure the concentration or directional dominance of the texture in a specified direction and is the core probabilistic basis of entropy calculation. And for all ,satisfy , It is a set of preset gradient orientation angles used as the reference orientation dimension when calculating gradients or texture responses in an image. Each It represents an angle in a specific direction, and the unit is usually "degree" or "radian". It is the directional structure entropy value, representing the first... The complexity of the texture structure or the degree of disorder in the directional distribution of the target region in a frame image across multiple dimensions is used to quantify the unevenness of the texture distribution in different directions of a local region within the image frame. The more dispersed the directions and the more unstable the texture structure, the more complex the texture structure. The higher the value, the more likely the area is to be disturbed by hot air disturbances; Obtaining the directional structure entropy values ​​of several consecutive frames After sequencing, the degree of change between adjacent frames in the sequence is analyzed to evaluate the fluctuation intensity of the target region texture in the time dimension, thereby generating a temporal texture perturbation index for identifying the local perturbation state of the image. The calculation expression is as follows: In the formula, This is a temporal texture perturbation index used to measure the intensity of texture perturbation in a target region across consecutive image frames. A larger value indicates that the local texture is more unstable and the thermal perturbation effect is more significant. It is the number of image frames. It is the magnitude of the directional structural entropy jump, representing the first... The change in the entropy of the texture orientation structure of the target region between frames is used to characterize the discontinuity and abrupt change in the texture orientation structure between frames. The calculation formula is as follows: , It is the directional structure entropy value of the previous frame image. It is the first The high-frequency texture coverage factor in the frame describes the structural complexity of the region, reflecting the sensitivity of the region to image disturbance, The greater, the more edges, textures, and details in the region, which is a sensitive area that is easy to be amplified by thermal disturbance; conversely, The smaller, the smoother or more uniform the region, the less structural change, and the less susceptible to disturbance, with a value range of 0-1.

[0030] By constructing the texture direction structure entropy sequence and combining the local jump volatility, the image local thermal air disturbance characteristics in high temperature environment are accurately modeled and quantified, effectively improving the judgment accuracy and stability of the image recognition system in dynamic disturbance background.

[0031] The greater the time sequence texture disturbance index generated by the time sequence variance of the texture features of the target region in the continuous frames under the detection window, the higher the possibility that the local region of the image is affected by thermal air disturbance. This index is based on the measurement of the change variance of the texture features (such as gray texture, edge direction, local pattern, etc.) of the target region in the continuous frame image sequence, reflecting the stability of the texture in the time dimension. If the region is affected by thermal disturbance, the instability of air refractive index will cause light deflection, resulting in small and random texture jitter, and thus irregular fluctuations of texture features between frames, which will increase the time sequence variance and increase the disturbance index. On the contrary, if the texture remains stable between multiple frames, it means that the region is not affected by thermal disturbance, and the image formation path is stable, so the disturbance index value is relatively low.

[0032] The low-frequency but irregular focus "drift" phenomenon of the image local sharpness on the time axis can clearly indicate that the image region has thermal air disturbance characteristics. This is because thermal air disturbance is essentially caused by the local fluctuation of the refractive index of the air medium around the high-temperature equipment, which causes the light path received by the camera during imaging to change slightly and dynamically. This change is not a sharp jump, but a slow, non-periodic but continuous disturbance to the local structure of the image, especially the sharpness of the focus area. Since the focus "drift" is manifested as a region gradually changing from clear to blurred and then recovering to clear in consecutive frames, and there is no fixed rhythm, this fluctuation is difficult to attribute to lens focusing or device jitter, but is a typical spatial non-uniform imaging distortion caused by thermal disturbance. Therefore, the spatial selectivity (only in the local), time low-frequency (change is not sharp but continuous) and nonlinear trend of the focus fluctuation constitute the structural evidence for judging the occurrence of thermal disturbance, which can be used as an important feature input for intelligent recognition of thermal disturbance.

[0033] The specific steps of generating the boundary continuity disturbance index by comprehensively analyzing the degree of fluctuation of the image local sharpness focus over time in the detection window are as follows: After the continuous frame data of the target region of the image is acquired under the detection window, the boundary contour information in each frame of the image is extracted based on an edge detection algorithm, and the boundary curve is subjected to fitting processing to enhance the structural continuity. For the boundary region of each frame of the image, the boundary curvature response value at the pixel level is calculated, and the calculation expression is as follows: In the formula, is the boundary curvature response intensity, indicating the “bending degree” of the pixel point in the image on the boundary structure, is the pixel point coordinate, that is, the geometric variation trend of the point on the boundary curve, used to identify the potential structural mutation point on the image boundary, which will produce a drift in time under thermal disturbance, and the value range is 0-1, is the Laplacian second-order derivative of the image gray scale, indicating the second-order variation rate of the gray scale of the pixel point in the image, which is usually realized through a Laplacian operator, used to reflect the curvature trend of the gray scale value in the local region of the image, that is, the “steepness of the edge, when the change at the boundary is violent, the second-order derivative value is large; and in the flat region, the value is approximately zero, is the gray scale variation rate of each pixel point in the image, indicating the “variation trend” of the image, is the image first-order gradient amplitude, indicating the first-order gradient module length of the gray scale of the pixel point in the image, indicating the rate of variation of the gray scale value, and used to suppress the interference of the background low-contrast region on the boundary response; The edge detection algorithm is a basic method in image processing, which is used to identify regions with a sharp change in gray scale in an image. These regions usually correspond to the outlines, structural boundaries or shape edges of objects. In the image, the edge represents the position where the pixel value changes sharply in space, so the structural features of the image can be extracted through edge detection. In the above steps, the main role of the edge detection algorithm is to extract the boundary contour of the worker and the equipment worn by the worker from the continuous image frames, to provide basic data support for the subsequent calculation of the boundary curvature response and the boundary continuity disturbance index. Typical existing edge detection algorithms include the Canny algorithm (with high accuracy and noise resistance), the Sobel operator (simple calculation and highlighting of edge directionality), the Prewitt operator, the Laplacian operator, etc. Among them, the Canny algorithm has good edge extraction robustness in a high-temperature disturbance environment due to its multi-stage processing mechanism (including noise suppression, gradient calculation, double-threshold detection, etc.), and is suitable for boundary feature extraction in complex power operation images. The boundary data extracted by the edge detection algorithm can be used as an important input for the spatial features such as curvature change and structural fracture in thermal disturbance recognition.

[0034] Based on the curvature response map of consecutive frame images, the curvature response difference at the same boundary location between adjacent frames is extracted, and the boundary continuity perturbation index is calculated by combining a region weighting mechanism. The calculation expression is as follows: In the formula, This is a boundary continuity perturbation index. A larger value indicates significant and concentrated structural perturbation at the local boundary, possibly due to thermal disturbance. A smaller value indicates stable image boundaries and good continuity. It is a set of curvature difference values, representing the set of differences in the curvature response values ​​of pixels on the same boundary between adjacent frames in a continuous image. ,in, It is the inter-frame curvature difference, representing the pixel value. The change in boundary curvature response between the current frame and the previous frame is calculated using the following formula: , It is the first Frame image pixels Boundary curvature response intensity at the point, It is the first Frame image pixels Boundary curvature response intensity at the point, It is the maximum value of the curvature difference among all boundary points in the same image region between consecutive frames. It is the effective boundary region within the domain of curvature response. This is a location-weighted coefficient, a regional weighting coefficient set based on the structural sensitivity or visual importance of boundary points. It is used to highlight the role of high curvature, corners, or abrupt shape changes in the overall disturbance assessment, and its value ranges from 0 to 1. It is along the horizontal direction of the image ( Insignificant changes on the axis It is along the vertical direction of the image ( A tiny change on the axis.

[0035] This disturbance index effectively characterizes the degree of disruption to local continuity caused by thermal disturbance by measuring the ratio between the maximum local disturbance amplitude and the overall disturbance energy distribution.

[0036] The greater the boundary continuity disturbance index generated by comprehensively analyzing the degree of fluctuation of the image local sharpness focus over time in the detection window, the greater it can be used as one of the criteria for the image local presence of hot air disturbance features. This index analyzes the geometric continuity changes of the image edge profile in the time series, especially the time stability of the boundary curvature, smoothness and topological structure. When hot air disturbance occurs, the dynamic non-uniform change of air density will cause slight bending of light during propagation, resulting in slight "shaking", "twisting" or "fuzzy fluctuation" of the target profile edge in the camera image. These disturbances usually do not change the overall target shape, but destroy the local continuity of the boundary, causing irregular changes in the boundary curvature between consecutive frames, and thus causing the boundary continuity disturbance index to rise. Therefore, the greater the index value, the stronger the non-stationary features in the image boundary, indicating that the hot disturbance is more significant; on the contrary, if the index value remains low for a long time, it indicates that the boundary continuity is stable and the image is not significantly disturbed by hot air.

[0037] The disturbance feature index after quantitative analysis is input into a deep learning model (such as a convolutional neural network CNN or a CNN-LSTM fusion network) trained in advance using historical image data, and the model analyzes the input disturbance features to determine whether there is a local hot disturbance phenomenon in the current image. Step 3, input the time series texture disturbance index and boundary continuity disturbance index after quantitative analysis into the deep learning model trained in advance using historical image data, generate a disturbance risk coefficient through the model, and determine whether there is a local hot disturbance phenomenon in the current image based on the disturbance risk coefficient.

[0038] The "deep learning model trained in advance using historical image data" means that before the system runs, a large number of collected and labeled image data are used as training samples, and the internal relationship between the hot air disturbance features and the disturbance results is modeled through a deep neural network, thereby constructing an image disturbance recognition model with generalization ability. In this process, first, a training data set is constructed, which consists of multiple image frame sequences containing hot disturbance and non-disturbance states, and each image frame needs to be labeled with whether it has a disturbance phenomenon and the corresponding disturbance level label (e.g. no disturbance, mild disturbance, moderate disturbance, severe disturbance). Then, the key index features related to hot disturbance (such as time series texture disturbance index, boundary continuity disturbance index, edge drift index, etc.) are extracted from each image, and these quantitative features are used as input vectors to the deep learning network for training. Common model structures can use fully connected neural networks (DNN), convolutional neural networks (CNN), or long short-term memory networks (LSTM) with stronger time series modeling capabilities, etc. Through algorithms such as back propagation and gradient descent, the model parameters are iteratively optimized, so that it learns to judge whether the image has a disturbance and the disturbance degree according to the input disturbance index.

[0039] After the model is trained, it has the ability to intelligently evaluate new image data in the actual running environment. In the actual detection process, when the system extracts the disturbance feature indicators of a frame or continuous frame image in real time, these features are input into the trained deep learning model. The model outputs a disturbance risk coefficient reflecting the current disturbance intensity according to the discrimination rules learned in the training process. The coefficient is a continuous value between 0 and 1, and the higher the value, the more likely the image is affected by thermal disturbance. The system compares the disturbance risk coefficient with the set threshold value. If it exceeds the set threshold value (such as 0.6), it is determined that the current image has thermal disturbance phenomenon, and then the robustness adjustment mechanism of the recognition network is triggered. Otherwise, the normal recognition path is maintained. The biggest advantage of this model is its adaptability and high fault tolerance. It can continuously expand the training set and continuously optimize through iteration, significantly improving the system's ability to identify non-explicit disturbance phenomena, especially suitable for image recognition stability protection in complex, dynamic, and high-temperature field environments such as power operations.

[0040] Compare the disturbance risk coefficient with the pre-set disturbance risk coefficient reference threshold to determine whether the current image has local thermal disturbance phenomenon. The judgment logic is as follows: If the disturbance risk coefficient is greater than the pre-set disturbance risk coefficient reference threshold, it is determined that the current image has local thermal disturbance phenomenon. If the disturbance risk coefficient is greater than the pre-set disturbance risk coefficient reference threshold, it is determined that the current image does not have local thermal disturbance phenomenon.

[0041] Step 4. When it is determined that the image data has thermal air disturbance, dynamically adjust the feature response parameters of the recognition network to reduce the recognition sensitivity to high-frequency detailed features (such as edges, small textures), and increase the recognition sensitivity to medium and low-frequency regions (such as overall shape, color blocks, and coarse line contours), to suppress the amplification interference of thermal disturbance effect on edge discrimination results and improve the recognition confidence of stable structures. The above steps aim to achieve robust perception and discrimination optimization of image information under thermal air disturbance by dynamically adjusting the recognition network's sensitivity to features at different frequency levels. In high-temperature or strong sunlight environments, edge regions and detailed textures in images are most susceptible to thermal air disturbance, manifesting as unstructured noise such as edge jitter, texture misalignment, and local blurring. If the recognition model maintains high sensitivity to high-frequency features (such as sharp edges and small-scale textures), it is highly likely to misjudge false edge information caused by disturbances as real features, leading to incorrect target localization results or distorted judgments of equipment status. Therefore, when the system detects thermal disturbance in the image, it needs to actively reduce the model's response to such high-frequency information, so that the recognition process no longer relies on unstable edge cues. Simultaneously, the model's sensitivity to mid- and low-frequency regions is increased, shifting the focus of recognition to more stable information sources in the image, such as the overall outline of the worker, the main color blocks of clothing, and the shape and structure of the safety helmet. These features are less deformed under thermal disturbance and more stably reflect the true state of the target. Through this sensitivity reconstruction mechanism, the system can effectively suppress the amplification and propagation of disturbance effects in the recognition results, avoid over-responding to abnormal edges, and increase the judgment weight of stable structures. Thus, while ensuring recognition robustness, it enhances the reliable recognition capability of the worker's wearing status, reduces the risk of false detection and missed detection, and ensures that reliable visual recognition support can still be provided under extreme working conditions.

[0042] When hot air disturbance is detected in the image data, the specific steps for dynamically adjusting the feature response parameters of the recognition network to reduce the recognition sensitivity of high-frequency detail features (such as edges and fine textures) while increasing the recognition sensitivity of mid-to-low frequency regions (such as overall shape, color blocks, and thick outlines) are as follows: When hot air disturbance is detected in the image data, a high-frequency suppression factor is constructed to regulate the sensitivity of high-frequency feature responses based on the deviation between the disturbance risk coefficient and a preset disturbance risk coefficient reference threshold. The formula is as follows: In the formula, It is the disturbance enhancement response coefficient, controlling the disturbance difference. In hyperbolic tangent function The magnification factor in the figure, taking values ​​of The larger the value, The faster the response to changes in disturbance risk, This is the disturbance risk coefficient; a higher value indicates a greater degree of disturbance. The value ranges from 0 to 1. It is a reference threshold for the disturbance risk coefficient. It is the hyperbolic tangent function. It is a high-frequency feature suppression factor, with a value of 0-1. The stronger the perturbation, the stronger the feature suppression factor. A value close to 0 indicates that the participation of high-frequency features is at its lowest. This step uses The core function of the (hyperbolic tangent) function is to perform nonlinear compression and dynamic smoothing mapping on the difference between the disturbance risk coefficient and the reference threshold, so as to realize the flexible response control of the identification model to changes in the disturbance intensity. The function exhibits an S-shaped curve characteristic, with an output range of (−1, 1). Its change is gradual when the input value is close to 0, and it rapidly approaches its extreme value when the input value is large or small. This characteristic makes it highly suitable for control scenarios where "weak disturbance → slight adjustment, strong disturbance → rapid suppression" is required. Specifically, regarding the high-frequency feature suppression mechanism involved in this invention... The function can map the over-threshold deviation of the disturbance risk coefficient into a smoothly decreasing suppression factor. This avoids frequent fluctuations caused by the model's oversensitivity when the disturbance approaches the threshold, and rapidly reduces the recognition participation of high-frequency features when high-intensity disturbances occur. This achieves a stable, controllable, and gradual sensing and adjustment mechanism for the model in response to disturbances. Therefore, The introduction of functions not only has mathematical rationality, but also provides a dual guarantee of smooth control and nonlinear response for the practical engineering deployment of the model.

[0043] when When the value is close to 0, it indicates that there is a strong disturbance in the image. This step will output a high-frequency suppression factor close to 0, which is used to reduce the model's response intensity to high-frequency information such as edges and fine textures, thereby effectively suppressing the erroneous edge enhancement and misjudgment behavior caused by thermal disturbance.

[0044] To ensure that the recognition process does not fail as a whole due to high-frequency suppression, and to simultaneously improve the ability to distinguish stable regions under perturbed scenarios, a mid-to-low frequency enhancement factor is constructed, and its calculation expression is as follows: In the formula, It is a low-to-mid frequency enhancement factor used to adjust the response intensity of low-to-mid frequency feature channels in the recognition network, with a value range of... When a thermal disturbance is detected, by increasing The numerical values ​​are used to enhance the focus on low-frequency structures (such as large color areas and the outline of the main body), thereby improving the confidence of wearable recognition in distinguishing stable regions. It is the low-frequency enhancement coefficient, control The maximum enhancement level is determined by the master control coefficient that adjusts the output gain intensity, ranging from 0.5 to 3. The larger the value, the greater the enhancement of the low-frequency response under disturbance. It is a disturbance response smoothing factor used to control the disturbance strength (i.e., The curvature and slope of the enhancement factor increase are in the range of 1-10, and are set according to the requirements of the system for distinguishing the intensity level of the disturbance; Through the enhancement mechanism, the perception sensitivity of the model to the low-frequency structural features such as the overall shape, color block, and main contour in the image can be significantly improved, thereby improving the target contour recognition accuracy in a high disturbance environment.

[0045] The high-frequency feature suppression factor and the medium-low frequency enhancement factor are applied to the multi-scale feature fusion module in the recognition network, respectively, to realize dynamic weighting of the responses of different scale features, and finally output a feature map, expressed as follows: In the formula, is the fusion output feature map, representing the final multi-scale feature response output of the recognition network in the heat disturbance detection state, which is directly input to the subsequent recognition module (such as the wearing detection module and the target classifier) for identifying the personnel target and the state of the safety protection equipment in the image, is the high-frequency feature response component, which is the edge, texture, contour details, and other high-frequency features extracted by the shallow layer (such as the first-3 layers of convolution) of the recognition network, is the medium-frequency feature response component, which is the medium-scale structural feature extracted by the middle layer of the recognition network, taking into account a certain spatial resolution and semantic expression ability, and is used to support the structural modeling of the target and improve the basic judgment of the equipment shape by the model. In this step, the response sensitivity is not adjusted, and the structural integrity is maintained, is the low-frequency feature response component, which is the low-frequency, semantic-level feature response output by the deep layer (such as the last 1-2 layers) of the recognition network, mainly describing the overall shape, category information, and contour trend of the target, and showing strong stability under heat disturbance, which is suitable for being used as the basis for judging the overall state of wearing (such as the presence or absence of safety helmets, the color block of work clothes, etc.).

[0046] By introducing the high-frequency suppression factor and the low-frequency enhancement factor, the participation proportion of multi-scale features in the fusion process is dynamically adjusted, thereby constructing a feature response mechanism with self-adaptive ability to the heat disturbance environment in the recognition model. This step can effectively suppress the misleading amplification of the disturbance on the edge discrimination result, and at the same time, enhance the recognition confidence of the model to the stable structure area in the image, thereby improving the overall recognition robustness and accuracy.

[0047] The application can effectively improve the stability and accuracy of the recognition system in typical disturbance environments such as high temperature or strong sunlight, especially in the case of interference such as local edge jitter and structure distortion of the image caused by hot air disturbance. The disturbance image is perceived in real time through feature index quantization and intelligent discrimination, and the response weight of the recognition network to multi-scale features is further dynamically adjusted, thereby significantly reducing the dependence on high-frequency noise, strengthening the recognition ability of low-frequency stable structure, avoiding key point misjudgment and wearing misjudgment caused by edge drift, ensuring that the recognition result of the operating personnel protection state has high robustness and engineering adaptability, and effectively solving the key safety risk problems such as "missing detection of wearing" or "no alarm for illegal operation".

[0048] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation. The preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0049] Those skilled in the art can realize that the units of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components of the examples have been generally described in the above description. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0050] In the embodiments provided by the present application, it should be understood that the division of units is only a logical functional division, and actual implementation can have another division manner, for example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc.

[0051] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0052] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-0nly Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0053] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the specification of the present application.

Claims

1. A method for detecting the protective clothing worn by personnel engaged in 0.4kV low-voltage live-line work based on image recognition, characterized in that, Includes the following steps: Step 1: Use industrial-grade high-definition cameras deployed at the work site to collect continuous image frame data of the work area where the low-voltage uninterrupted power supply workers are located in real time; Step 2: Preprocess the real-time acquired image data, and extract key indicators reflecting the characteristics of local thermal air disturbance from the preprocessed image data. Perform comprehensive analysis on the extracted key indicators to quantify the image structure anomalies caused by thermal air disturbance. Step 3: Input the perturbation feature index after quantitative analysis into the deep learning model that has been trained using historical image data. The model performs a comprehensive analysis of the input perturbation features to determine whether there is a local thermal perturbation phenomenon in the current image. Step 4: When it is determined that there is thermal air disturbance in the image data, the feature response parameters of the recognition network are dynamically adjusted to reduce the recognition sensitivity of high-frequency detail features and increase the recognition sensitivity of mid- and low-frequency regions, so as to suppress the amplification interference of thermal disturbance effect on edge discrimination results.

2. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step 11: Open the image acquisition channel through the connected camera device, and start acquiring continuous image frame data of the target work area based on the set time interval or trigger mechanism; Step 12: The camera device transmits the captured image frames to the on-site computing module in real time as an image stream, using a high-speed interface protocol to ensure the integrity and low latency of image transmission; Step 13: The image stream is uniformly cached in the calculation module and arranged sequentially according to timestamps to construct an image frame sequence that reflects the dynamic characteristics of the operation process; Step 14: Perform real-time scheduling and management of the image frame sequence in the cache to provide a continuous frame input interface for subsequent recognition tasks.

3. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 1, characterized in that, The specific steps of step 2 are as follows: Step 21: The key indicators extracted include the temporal variance of the texture features of the target region in consecutive frames and the degree of fluctuation of the local sharpness and focus of the image over time; Step 22: Perform a comprehensive analysis of the temporal variance of the target region's texture features in consecutive frames and the degree of fluctuation of the image's local sharpness and focus over time within the detection window to generate temporal texture perturbation index and boundary continuity perturbation index respectively; Step 23: Quantify the image structure anomalies caused by thermal air disturbance using temporal texture perturbation index and boundary continuity perturbation index.

4. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 3, characterized in that, In step 22, the specific steps for generating a temporal texture perturbation index by comprehensively analyzing the temporal variance of the target region's texture features in consecutive frames within a detection window are as follows: Within the detection window, texture orientation features of the target region are extracted from consecutive image frames. The complexity of the texture orientation distribution in each frame is characterized by the orientation gradient structure entropy. For each frame in the image, a set of gradient orientations is defined. The distribution probability of texture gradient intensity in each direction is statistically analyzed, and the directional structure entropy value is calculated. The calculation expression is as follows: In the formula, It is the first Frame in direction Texture gradient probability distribution values ​​on the surface. It is a set of preset gradient direction angles. It is the entropy value of the directional structure; Obtaining the directional structure entropy values ​​of several consecutive frames After sequencing, the degree of change between adjacent frames in the sequence is analyzed to evaluate the fluctuation intensity of the target region texture in the time dimension, thereby generating a temporal texture perturbation index for identifying the local perturbation state of the image. The calculation expression is as follows: ， In the formula, It is a temporal texture perturbation index. It is the number of image frames. It is the magnitude of the directional structural entropy jump, calculated using the following formula: , It is the directional structure entropy value of the previous frame image. It is the first High-frequency texture coverage factor in the frame.

5. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 3, characterized in that, In step 22, the specific steps for generating a boundary continuity perturbation index by comprehensively analyzing the fluctuation of local image sharpness focus over time within the detection window are as follows: Within the detection window, after acquiring continuous frame data of the target region of the image, boundary contour information in each frame is extracted based on the edge detection algorithm. For the boundary region of each frame, the boundary curvature response value at the pixel level is calculated, and the calculation expression is as follows: ， In the formula, It is the boundary curvature response intensity, representing the pixel in the image. The "bending" strength at the boundary structure These are pixel coordinates. It is the second Laplacian derivative of the image grayscale, representing the image at a given pixel. The second-order rate of change of gray level at that location. It is each pixel in the image. The grayscale change rate represents the "trend of change" of the image. It is the magnitude of the first-order gradient of the image; Based on the curvature response map of consecutive frame images, the curvature response difference at the same boundary location between adjacent frames is extracted, and the boundary continuity perturbation index is calculated by combining a region weighting mechanism. The calculation expression is as follows: ， In the formula, It is a boundary continuity disturbance index. It is a set of curvature differences. ,in, It is the difference in curvature between frames, and the calculation formula is as follows: , It is the first Frame image pixels Boundary curvature response intensity at the point, It is the first Frame image pixels Boundary curvature response intensity at the point, It is the maximum value of the curvature difference among all boundary points in the same image region between consecutive frames. It is the effective boundary region within the domain of curvature response. These are position-weighted coefficients. It is a tiny change along the horizontal direction of the image. It is a tiny change along the vertical direction of the image.

6. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 3, characterized in that, Also includes: The temporal texture perturbation index and boundary continuity perturbation index, after quantitative analysis, are input into a deep learning model pre-trained using historical image data. The model generates a perturbation risk coefficient, and the presence of local thermal perturbation in the current image is determined based on the perturbation risk coefficient.

7. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 6, characterized in that, The step involves comparing the disturbance risk coefficient with a pre-set reference threshold to determine whether the current image exhibits local thermal disturbance. The determination logic is as follows: If the disturbance risk coefficient is greater than the preset disturbance risk coefficient reference threshold, it is determined that there is a local thermal disturbance in the current image; if the disturbance risk coefficient is greater than the preset disturbance risk coefficient reference threshold, it is determined that there is no local thermal disturbance in the current image.

8. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 7, characterized in that, The specific steps for dynamically adjusting the feature response parameters of the recognition network when hot air disturbance is detected in the image data, reducing the recognition sensitivity to high-frequency detail features while increasing the recognition sensitivity in the mid-to-low frequency region, are as follows: When hot air disturbance is detected in the image data, a high-frequency suppression factor is constructed to regulate the sensitivity of high-frequency feature responses based on the deviation between the disturbance risk coefficient and a preset disturbance risk coefficient reference threshold. The formula is as follows: ， In the formula, It is the disturbance enhancement response coefficient. It is the disturbance risk coefficient. It is a reference threshold for the disturbance risk coefficient. It is the hyperbolic tangent function. It is a high-frequency characteristic inhibitory factor; To ensure that the recognition process does not fail as a whole due to high-frequency suppression, and to simultaneously improve the ability to distinguish stable regions under perturbed scenarios, a mid-to-low frequency enhancement factor is constructed, and its calculation expression is as follows: ， In the formula, It is a mid-to-low frequency enhancement factor. It is the low-frequency enhancement coefficient. It is a disturbance response smoothing factor; High-frequency feature suppression factor and mid-to-low frequency enhancement factor These are applied to the multi-scale feature fusion module in the recognition network to achieve dynamic weighting of feature responses at different scales, ultimately fusing and outputting a feature map, as shown in the following expression: ， In the formula, It is a fused output feature map. It is a high-frequency characteristic response component. It is the mid-frequency characteristic response component. It is a low-frequency characteristic response component.

9. The method for detecting the protective clothing worn by 0.4kV low-voltage live-line workers based on image recognition according to claim 1, characterized in that: Step 2, which involves preprocessing the real-time acquired image data, specifically includes the following steps: First, grayscale conversion is performed to convert the color image into a single-channel grayscale image, in order to reduce computational complexity and highlight the structural information of the target area; Then, a denoising filtering algorithm is applied to suppress sensor noise, dust interference, or background clutter, thereby improving the image edge sharpness. Next, histogram equalization is used to enhance the overall brightness distribution balance of the image, making detailed areas easier to identify in excessively dark or bright scenes. Finally, adaptive contrast adjustment is implemented to enhance edge and texture contrast while ensuring that local features are not lost.

Images