A power transmission line external damage hidden danger early warning method based on a visualization device

Abstract

The application discloses a kind of based on visual device transmission line external damage hidden danger early warning method, and the method steps are as follows: S1 obtains the transmission line image containing large construction machinery and pre-processes;S2 is to the pre-processed image Shadow synthesis obtains enhanced image, reduces enhanced image, and obtains fusion feature image by fusing the features of three;S3 is constructed and is trained target recognition model with fusion feature image, and the optimal model is obtained by optimizing parameters using optimization algorithm;S4 visual device obtains real-time image, and optimal model identification, if mechanical is identified, then progress S5;S5 millimeter wave radar scans three-dimensional point cloud data, projection two-dimensional is compared with identification result, and if it coincides, then progress S6;S6 preset safety distance, calculate the height of machinery in combination with point cloud data and projection area, and if it exceeds safety distance, then alarm.The application can realize all-weather unattended, accurately identify machinery, filter invalid early warning, reduce missed detection, reduce the risk of early warning failure, avoid line trip-out and other accidents, and protect the safety of transmission line.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission line protection, in particular to a power transmission line external damage hidden danger early warning method based on a visual device. BACKGROUND

[0002] At present, most power failure accidents are caused by external damage to the power grid, so ensuring the safety of power transmission lines has become a key social problem that needs to be solved urgently. At present, a large number of power transmission and distribution networks pass through urban and rural areas, and are inevitably disturbed by human production and living activities, resulting in a large number of safety hazards around overhead power transmission lines. For example, large construction machinery (cranes, excavators, dump trucks, etc.) used in construction sites are difficult to accurately control the safety distance from overhead power transmission lines during operation, and are easy to break through the safety threshold, thereby causing power transmission and distribution line tripping, damage to related power facilities, and other accidents.

[0003] However, the existing monitoring technology for preventing external damage still has obvious technical defects and cannot meet the actual needs:

[0004] Limitations of traditional video monitoring technology: Traditional video monitoring technology relies on monitoring personnel to monitor video images in real time in the monitoring center. Once the monitoring personnel leaves the display, the monitoring ability of the on-site situation will be lost, and continuous monitoring without human intervention cannot be realized, and there is a risk of monitoring blind spots and response lag.

[0005] Defects of existing intelligent video monitoring technology:

[0006] 1. Low warning accuracy: Most existing intelligent video monitoring technology only identifies and warns moving objects, while the construction site environment around the power transmission line is complex, often with personnel, ordinary vehicles, birds, and other non-threatening moving objects, which easily triggers system false alarms, resulting in low alarm accuracy, many invalid warnings, and difficulty in accurately locking the core safety hazard source.

[0007] 2. Poor adaptability to complex scenes: The power transmission line coverage area often involves complex environments such as mountains, rivers, and suburbs, and challenging working conditions such as extreme lighting (such as direct sunlight, backlight, and weak light at night), obstructions (such as tree obstructions and construction material stacking obstructions), and long-distance shooting (such as construction machinery on the road far end or hillside) are easily encountered in the monitoring scene. Existing intelligent video monitoring technology is prone to broken segmentation errors in such scenarios, specifically manifested as blurred road edge and mechanical contour definition, small construction equipment or mechanical key component missed detection, resulting in failure to accurately extract the complete form and position information of large machinery, further causing misjudgment of the safety distance between the large machinery and the power transmission line, and increasing the risk of early warning failure. SUMMARY

[0008] In view of the deficiencies of the prior art, the application provides a power transmission line external damage hidden danger early warning method based on a visual device, and aims at solving the problems in the background art.

[0009] To achieve the above object, the application provides the following technical scheme: a power transmission line external damage hidden danger early warning method based on a visual device, comprising the following steps:

[0010] Step S1: acquire a plurality of power transmission line images containing large construction machinery and perform preprocessing;

[0011] Step S2: perform shadow synthesis on the preprocessed images to form enhanced images; reduce the enhanced images to obtain reduced images; perform feature extraction and fusion on the preprocessed images, the enhanced images and the reduced images to obtain a fusion feature map;

[0012] Step S3: construct a target recognition model; train the target recognition model based on the fusion feature map, and simultaneously, optimize the model parameters by using an optimization algorithm to obtain an optimal target recognition model;

[0013] Step S4: acquire real-time power transmission line images through a visual device, and obtain real-time fusion feature maps by performing step S2 on the real-time power transmission line images; input the real-time fusion feature maps into the optimal target recognition model to identify large construction machinery in the images; when the large construction machinery is identified, proceed to step S5, otherwise, proceed to the next identification cycle;

[0014] Step S5: scan a target power transmission line region by using a millimeter wave radar unit to acquire three-dimensional point cloud data of the large construction machinery; compare the three-dimensional point cloud data with the large construction machinery identified by the optimal target recognition model in a two-dimensional space; when the three-dimensional point cloud data and the large construction machinery identified by the optimal target recognition model do not coincide in the two-dimensional space, proceed to the next identification cycle, otherwise, proceed to step S6;

[0015] Step S6: preset a conductor cross arm safety distance, combine the three-dimensional point cloud data and a projection area on a two-dimensional image coordinate to calculate the height of the identified large construction machinery; when the height of the identified large construction machinery exceeds the preset safety distance, report an alarm, otherwise, proceed to the next identification cycle.

[0016] Further, the specific process of performing shadow synthesis on the preprocessed images to form enhanced images is as follows:

[0017] Step S2.31: set the height of the circumscribed rectangle of the large and medium-sized machinery in each preprocessed image as ; and each preprocessed image contains:

[0018] ​Mechanical material label ;

[0019] Artificial light source parameters in night scene, including light source direction and light source intensity ;

[0020] Step S2.32: scene classification and parameter extraction

[0021] Classify the scene and extract the key parameters for each pre-processed image:

[0022] Calculate the average gray value of the brightness histogram of the pre-processed image , combined with the real-time illumination angle collected by the illumination angle sensor , into strong light direct scene, backlight scene and night weak light scene

[0023] Segment the pre-processed image using the U-Net semantic segmentation model to output the occlusion mask , wherein represents the pixel is the occlusion area, represents the pixel is the non-occlusion area

[0024] Fuse the occlusion mask and the dynamic occlusion area mask to obtain the real-time occlusion mask ;

[0025] Step S2.33: basic shadow template generation

[0026] Based on the height of the circumscribed rectangle of the large machinery in the pre-processed image and the real-time illumination angle , calculate the length of the shadow pixel projected by the large machinery in the corresponding pre-processed image ;

[0027] Calculate the average gray value of the large machinery area in the pre-processed image ;

[0028] Based on the average gray value of the large machinery area in the pre-processed image , calculate the shadow gray value under different scenes ;

[0029] Based on the shadow pixel length , construct a two-dimensional pixel matrix on the side of the large machinery in the pre-processed image, and set the gray value of each pixel in the matrix to , forming a basic shadow template ;

[0030] Step S2.34: Occlusion shadow mask superposition;

[0031] The real-time occlusion mask is superposed with the basic shadow template to obtain a shadow template with occlusion characteristics by pixel point multiplication.

[0032] Step S2.35: The shadow template is fused with the preprocessed image using a pixel-level weighted fusion algorithm to generate an enhanced image , which is expressed as:

[0033] ;

[0034] In the formula, represents the fusion weight.

[0035] Further, the specific process for obtaining the dynamic occlusion area mask is as follows: the pixel changes of adjacent 5 frames of images are compared using an inter-frame difference method to extract the dynamic occlusion area mask , which is expressed as:

[0036] ;

[0037] In the formula, represents the gray value of the i-th frame image at the pixel ; represents the gray value of the i+1-th frame image at the pixel ;

[0038] Further, the fusion weight is expressed as:

[0039] ;

[0040] In the formula, represents the shadow gray value; represents the average gray value of the large mechanical region.

[0041] Further, in a night weak light scene, when there is only 1 high-intensity light source, the basic shadow template is still used, and when there are ≥2 high-intensity light sources, the basic shadow template is replaced with a double-light-source shadow template ; the high-intensity light source is defined as: ;

[0042] The specific process for obtaining the double-light-source shadow template is as follows:

[0043] According to the light source intensity Sort the light sources and select the first two high-intensity light sources as the main light sources, denoted as light source 1 and light source 2;

[0044] Assign shadow weights according to the proportion of light source intensity:

[0045] ;

[0046] In the formula, This represents the shadow weight of light source 1; Indicates the intensity of light source 1; This represents the shadow weight of light source 2; This indicates the intensity of light source 2; ;

[0047] According to the direction of the light source Adjust the shadow casting angle to generate a dual-light source shadow template. , is represented as:

[0048] ;

[0049] In the formula, This represents the base shadow template generated based on light source 1; This represents the base shadow template generated based on light source 2.

[0050] Furthermore, the optimization algorithm employs an improved RL-TLBO algorithm, and the target recognition model uses the YOLOv8 model. The improved RL-TLBO algorithm is used to optimize the parameters of the YOLOv8 model to obtain the optimal target recognition model. The specific process is as follows:

[0051] Step S3.1: Configure the hyperparameters of the improved RL-TLBO algorithm, including population size. Maximum number of iterations Initial learning rate and initial value of inertia weight The convolutional kernel weights, feature fusion weights of the intermediate processing network, classification weights of the detector head, and anchor box regression weights of the YOLOv8 model backbone network are extracted as the YOLOv8 model parameters to be optimized. ;

[0052] Step S3.2: Construct the weighted fitness function for the shadow scene, expressed as:

[0053] ;

[0054] In the formula, Indicates parameters fitness value; Indicates the first Large construction machinery in shadow-like scenes The score reflects the accuracy of mechanical recognition. ; Indicates the false alarm rate for non-threat targets; This indicates the total number of missed detections of large machinery in the shadowed scene; , , All of these represent recognition weights, used to balance optimization across different scenarios; This indicates the weight of the false alarm rate on the fitness level. This represents the penalty coefficient for missed detections;

[0055] Step S3.3: Initial population generation using shadow feature-guided Logistic mapping;

[0056] Step S3.31: Using a Logistic mapping that incorporates shadow brightness features, generate a random seed for the initial population, represented as:

[0057] ;

[0058] In the formula, Indicates the first The chaotic value of the next iteration; Indicates the first The chaotic value of the next iteration; Indicates the average brightness of the shaded area; Indicates the Logistic mapping parameters;

[0059] Step S3.32: Linearly map the generated chaotic values ​​to the YOLOv8 model parameters. Within the defined scope, the initial population of individuals is generated;

[0060] Step S3.4: Dynamically update shadow scene parameters based on SAC reinforcement learning;

[0061] Step S3.41: Update parameters for the teaching phase; select the individual with the highest overall fitness in the current population as the teacher individual. ; Calculate the current population average parameters Update individual temporary parameters:

[0062] ;

[0063] In the formula, Indicates the first During the nth iteration, the 1st The first individual Updated values ​​for each parameter; Represents a random number; Indicates the first During the nth iteration, the 1st The first individual original value of the i-th parameter; denotes a teaching factor; denotes a population size; denotes the i-th individual at the j-th iteration; denotes the i-th individual at the j-th iteration; denotes the i-th individual at the j-th iteration; denotes the average value of the i-th parameter in the population average parameter denotes the average value of the i-th parameter in the population average parameter denotes the average value of the i-th parameter in the population average parameter

[0064] Step S3.42: SAC reinforcement learning state and reward definition;

[0065] Step S3.43: Dynamically update the inertia weight: the SAC agent maximizes the cumulative reward , denotes a discount factor, and the output optimal inertia weight adjustment amount , updates the current inertia weight , denotes the inertia weight at the j-th iteration; denotes the inertia weight at the j-th iteration; denotes the inertia weight at the j-th iteration; denotes the inertia weight at the j-th iteration; denotes the discount factor at the j-th iteration; denotes the immediate reward at the j-th iteration; denotes the immediate reward at the j-th iteration; Step S3.44: Student stage parameter update: randomly select two different individuals in the population

[0066] , , combine the updated inertia weight , and update the final parameter; Step S3.5: Local search;

[0067] Step S3.6: Improve the RL-TLBO algorithm convergence judgment and optimal target identification model generation;

[0068] Step S3.61: Convergence judgment: the improved RL-TLBO algorithm is determined to be converged when the following two conditions are met:

[0069] the overall fitness change amount of the continuous

[0070] generations , denotes the population optimal fitness value at the j-th iteration, denotes the population optimal fitness value at the j-th iteration, denotes the first convergence threshold;

[0071] ​​​The fitness variation of the three types of shadow scenarios satisfies , represents the optimal fitness value of the first type of scenario in the i-th iteration, represents the optimal fitness value of the second type of scenario in the i-th iteration, represents the optimal fitness value of the third type of scenario in the i-th iteration, represents the optimal fitness value of the first type of scenario in the i-th iteration, represents the optimal fitness value of the second type of scenario in the i-th iteration, represents the optimal fitness value of the third type of scenario in the i-th iteration, represents the second convergence threshold;

[0072] Step S3.62: Load the final converged optimal parameters into the YOLOv8 model and verify, represents the optimal parameter individual in the population in the i-th iteration; Step S3.63: If the verification is passed, output the corresponding YOLOv8 model as the optimal target recognition model, otherwise return to step S3.3 to adjust the Logistic mapping parameters

[0073] and re-execute the optimization process.

[0074] Further, the specific process of SAC reinforcement learning state and reward definition is as follows:

[0075] Define the state , including the diversity of the population in the i-th iteration , the fitness improvement rate in the i-th iteration , and the shadow scenario fitness deviation in the i-th iteration ; The population diversity is represented as:

[0076] ;

[0077] ;

[0078] ;

[0079] In the formula, represents the probability weight of ; represents the fitness value of ;

[0080] The fitness improvement rate is represented as:

[0081] ;

[0082] In the formula, represents the fitness value of ; ​​​the optimal parameter individual in the population at the i-th iteration; the optimal parameter individual in the population at the i-th iteration; the fitness value of the i-th iteration; the optimal parameter individual in the population at the i-th iteration; the optimal parameter individual in the population at the i-th iteration; the optimal parameter individual in the population at the i-th iteration;

[0083] shadow scene fitness deviation is represented as:

[0084]

[0085] wherein, the fitness value of the i-th iteration under the direct light scene; the fitness value of the i-th iteration under the backlight scene; the fitness value of the i-th iteration under the night weak light scene; the reward at the i-th iteration is defined as:

[0086] the reward at the i-th iteration is defined as:

[0087] Further, the specific process of the local search is as follows:

[0088] Step S3.51: Set a local search triggering condition, which triggers the local search when any local search condition is met; the local search triggering condition is:

[0089] after each iteration;

[0090] the fitness of a certain type of shadow scene continuously <first preset improvement rate threshold>;

[0091] Step S3.52: YOLOv8 model parameter sensitive point positioning: focus on the parameter sensitive point that has the greatest impact on shadow identification; Step S3.53: Add Gaussian disturbance to the parameter sensitive point to generate multiple candidate solutions;

[0092] Step S3.54: Evaluate the fitness of the

[0093] Step S3.54: Evaluate the fitness of the

[0094] Step S3.54: Evaluate the fitness of the Step S3.54: Evaluate the fitness of the Step S3.55: If the same parameter sensitive point is continuously selected for

[0095] ​​​​​​​​No significant improvement in generational fitness If the current learning rate is less than the second preset boost rate threshold, then... Decline exponentially.

[0096] Furthermore, the enhanced image is reduced in size to obtain the reduced image. The specific process is as follows:

[0097] Step S2.11: Obtain the baseline monitoring distance of the visualization device ;

[0098] Step S2.12: Based on the baseline monitoring distance The actual monitoring distance of the visualization device Calculate the reduction ratio ;

[0099] Step S2.13: Enhance the image using bilinear interpolation. According to the reduction ratio The image is reduced in size to obtain a preliminary reduced image. ;

[0100] Step S2.14: Initially reduce the image size Sobel edge sharpening is used to obtain a smaller image. .

[0101] Furthermore, the specific process of extracting and fusing features from the preprocessed image, the enhanced image, and the reduced image to obtain the fused feature map is as follows:

[0102] Step S2.21: Obtain the preprocessed image Image enhancement and reduce image size Real-time scene type tags ;

[0103] Step S2.22: Process the preprocessed images respectively Image enhancement and reduce image size Extraction is performed to obtain the preprocessed image feature map. Enhanced image feature maps and reduced image feature map ;

[0104] Step S2.23: with The center coordinates of large machinery in China Based on, , Mapped to the same coordinate system;

[0105] Step S2.24: Based on scene type label Dynamically adjust the preprocessed image feature map Enhanced image feature maps and reduced image feature map The weights;

[0106] Step S2.25: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] , , Concatenate them into a feature matrix according to the channel dimension. ;

[0107] Step S2.26: Utilize the preprocessed image feature map weight Enhanced image feature maps weight Reduced image feature map weight For the characteristic matrix Weighted fusion is performed to generate the final fused feature map. .

[0108] Compared with existing technologies, the present invention has the following advantages:

[0109] (1) This invention addresses the problems of traditional video surveillance, such as "reliance on manual labor, blind spots and delayed response", and existing intelligent video surveillance, such as "low accuracy of early warning and poor adaptability to complex scenarios". Through the synergistic effect of shadow synthesis, YOLOv8 model optimization and improved RL-TLBO algorithm, it realizes all-weather unattended monitoring, can accurately identify large construction machinery around transmission lines, filter invalid early warnings, reduce missed detections, accurately determine safe distances, reduce the risk of early warning failure, effectively avoid accidents such as line tripping and damage to power facilities, and ensure the safety of transmission lines.

[0110] (2) This invention generates dynamic shadow templates by scene classification and generates enhanced images by weighted fusion, which solves the problem of shadows being confused with machinery / background under extreme lighting; it calculates the reduction ratio by monitoring distance and sharpens to generate a reduced image, which solves the problem of blurred details in long-distance images, improves the image quality under complex lighting and long-distance scenes, and lays the foundation for feature extraction; this invention dynamically adjusts the image feature weights according to scene type labels, avoids single feature defects by fusion, ensures complete and accurate feature extraction under different scenes, reduces the problem of missed detection of key mechanical parts and blurred contour definition, and improves the adaptability to complex scenes.

[0111] (3) This invention optimizes model parameters by dividing the shadow scene fitness function, shadow feature initial population, SAC reinforcement learning parameter update and local search, solves the problems of many false alarms and many missed detections in shadow scenes in existing models, improves the recognition accuracy of large construction machinery, and filters out invalid warnings of non-threat targets.

[0112] (4) The present invention projects the radar three-dimensional point cloud onto a two-dimensional image and compares and verifies it with the model recognition result, filters out image misjudgments (such as misjudgment of birds or ordinary vehicles), further reduces invalid warnings, and improves the accuracy of target recognition; combines the three-dimensional point cloud and the two-dimensional projection area to accurately calculate the mechanical height, compares it with the preset safe distance to alarm, avoids distance misjudgment caused by single image size estimation, reduces warning failure, and reduces accident risk. Attached Figure Description

[0113] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0114] like Figure 1 As shown, the present invention provides a technical solution: a method for early warning of external damage hazards to transmission lines based on a visualization device, comprising the following steps:

[0115] Step S1: Acquire several images of power transmission lines containing large construction machinery and perform preprocessing.

[0116] Step S2: Perform shadow synthesis on the preprocessed image to form an enhanced image; reduce the size of the enhanced image to obtain a reduced image; extract and fuse features from the preprocessed image, the enhanced image, and the reduced image to obtain a fused feature map.

[0117] Step S3: Construct a target recognition model; train the target recognition model based on the fused feature map, and at the same time, use an optimization algorithm to optimize the model parameters to obtain the optimal target recognition model.

[0118] Step S4: Acquire real-time transmission line images through a visualization device (camera). Pass the real-time transmission line images through step S2 to obtain a real-time fused feature map. Input the real-time fused feature map into the optimal target recognition model to identify large construction machinery in the image. When large construction machinery is identified, proceed to step S5; otherwise, proceed to the next recognition cycle.

[0119] Step S5: Use a millimeter-wave radar unit to scan the target transmission line area and obtain three-dimensional point cloud data of large construction machinery; project the three-dimensional point cloud data onto two-dimensional image coordinates and compare it with the large construction machinery identified by the optimal target recognition model. If the three-dimensional point cloud data and the large construction machinery identified by the optimal target recognition model do not overlap in two-dimensional space, proceed to the next recognition cycle; otherwise, proceed to step S6.

[0120] Step S6: Set a safe distance for the conductor crossarm, and calculate and identify the height of the large construction machinery by combining the three-dimensional point cloud data and the projected area on the two-dimensional image coordinates; if the height of the identified large construction machinery exceeds the preset safe distance, report an alarm; otherwise, proceed to the next identification cycle.

[0121] wherein the specific process of reducing the enhanced image to obtain the reduced image is:

[0122] Step S2.11: obtaining the reference monitoring distance of the visualization device .

[0123] Step S2.12: calculating the reduction ratio according to the reference monitoring distance and the actual monitoring distance of the visualization device , which can be expressed as:

[0124] ;

[0125] wherein, the reference reduction ratio is represented by .

[0126] Step S2.13: reducing the enhanced image by the reduction ratio to obtain a preliminary reduced image .

[0127] Step S2.14: performing Sobel edge sharpening on the preliminary reduced image to obtain the reduced image , which can be expressed as:

[0128] ;

[0129] ;

[0130] .

[0131] wherein, represents the Sobel edge gradient operator in the horizontal direction; represents the Sobel edge gradient operator in the vertical direction; represents the pixel coordinate; represents the sharpening weight coefficient, which is used to control the intensity of edge sharpening.

[0132] wherein the specific process of extracting and fusing features from the preprocessed image, the enhanced image and the reduced image to obtain the fused feature map is:

[0133] Step S2.21: obtaining the real-time scene type label of the preprocessed image , the enhanced image and the reduced image : : complex light / shading; : long distance;​ : normal scene).

[0134] Step S2.22: Extract the pre-processed image feature map , the enhanced image feature map and the reduced image feature map by the ResNet50 network respectively. .

[0135] Step S2.23: Map , to the same coordinate system by affine transformation with the large machine center coordinate in as the reference.

[0136] Step S2.24: Dynamically adjust the weights of the pre-processed image feature map , the enhanced image feature map and the reduced image feature map according to the scene type label , which can be expressed as:

[0137] ;

[0138] wherein , , represent the weights of , , respectively. .

[0139] Step S2.25: Concatenate , , into a feature matrix in the channel dimension.

[0140] Step S2.26: Perform weighted fusion on the feature matrix , , to generate the final fusion feature map , which can be expressed as:

[0141] ;

[0142] wherein represents the channel index of ; represents the channel index of .​​​ represents the channel index of represents the channel index of

[0143] The specific process of shadow synthesis on the pre-processed image to form an enhanced image is as follows:

[0144] Step S2.31: Set the height of the outer rectangle of the large machine in each pre-processed image to ; each pre-processed image contains:

[0145] Machine material label (based on image texture feature classification: metal material , such as crane arm; non-metallic material , such as excavator cab plastic parts);

[0146] Artificial light source parameters (light source direction , indicating the horizontal angle between the light source and the machine; light source intensity , such as the intensity of construction vehicle lights, street light intensity) under night scene.

[0147] Step S2.32: scene classification and parameter extraction;

[0148] Scene classification and key parameter extraction are performed on each pre-processed image:

[0149] Illumination scene classification: calculate the average gray value of the brightness histogram of the pre-processed image , combined with the real-time illumination angle (sun / light source and horizontal direction angle) collected by the illumination angle sensor, and divided into three categories of shadow scenes:

[0150] Strong light direct scene: > 220 and > 60°;

[0151] Backlight scene: / > 2 and < 30°, represents the average gray value of the power tower / sky area in the pre-processed image, represents the average gray value of the large machine area in the pre-processed image;

[0152] Weak light scene at night: < 50.

[0153] Obstruction type identification: the pre-processed image is segmented by using a U-Net semantic segmentation model, and an obstruction mask ​​​​wherein, represents a pixel is an occluded region (such as trees / construction materials), represents a pixel is a non-occluded region.

[0154] The occlusion mask is fused with the dynamic occluded region mask to obtain a real-time occlusion mask , which can be represented as:

[0155] .

[0156] wherein the specific process for obtaining the dynamic occluded region mask is as follows:

[0157] By using the inter-frame difference method, the pixel changes of adjacent 5 frames of images are compared to extract the dynamic occluded region mask , which can be represented as:

[0158] ;

[0159] In the formula, represents the gray value of the i-th frame of image at the pixel ; represents a gray difference threshold value, which can be adjusted according to the illumination; represents a pixel is a dynamic occluded region. Step S2.33: Generation of a basic shadow template;

[0160] According to the scene classification, a dynamic shadow template matching the physical law is constructed:

[0161] Based on the height of the circumscribed rectangle of the large machine in the preprocessed image

[0162] and the real-time illumination angle , the length of the shadow pixel projected by the large machine in the corresponding preprocessed image is calculated , which can be represented as:

[0163] .

[0164] The average gray value of the large machine region in the preprocessed image is calculated :

[0165] ;

[0166] In the formula, represents the total number of pixels of the circumscribed rectangle of the large machine. ​​

[0167] Average gray value of large machinery area in pre-processed image Calculate the shadow gray value under different scenes (Adjust based on scene brightness to avoid too dark or too light shadows):

[0168] If it is a strong light direct scene: high ambient brightness, machinery is easy to reflect light, take low gray coefficient to avoid confusion between shadow and reflection, the shadow gray value is represented as: ;

[0169] If it is a backlight scene: bright background, dark foreground machinery, take medium gray coefficient to balance brightness, the shadow gray value is represented as: ;

[0170] If it is a weak light scene at night: overall low brightness, add light compensation to avoid shadow and night color fusion, the shadow gray value is represented as: .

[0171] The range of is checked, if <0, take 0, if >255, take 255, to ensure that the shadow brightness is within the visually distinguishable range.

[0172] Based on the length of the shadow pixel , a two-dimensional pixel matrix is constructed on the side of the large machinery in the pre-processed image (opposite to the light source direction), the gray value of each pixel in the matrix is set to the corresponding position , forming a basic shadow template .

[0173] Step S2.34: superimpose the occlusion shadow mask;

[0174] Multiply the real-time occlusion mask and the basic shadow template by pixel points to get the shadow template with occlusion characteristics , which can be represented as:

[0175] .

[0176] Step S2.35: use a pixel-level weighted fusion algorithm to fuse the shadow template and the pre-processed image to generate an enhanced image , which can be represented as:

[0177] ;

[0178] In the formula, represents the fusion weight.

[0179] Wherein, the fusion weight Can be expressed as:

[0180] ;

[0181] In the formula, Indicates the shadow gray scale; Indicates the large machinery area average gray scale; when the metal material is, With the decrease of shadow gray scale, it increases slightly (retains the reflection details), and the value range is 0.72-0.9; when the non-metal material is, With the increase of shadow and machinery gray scale difference, it increases slightly (suppresses background interference), and the value range is 0.3-0.5.

[0182] Wherein, in the night weak light scene, if there is only 1 high intensity light source, the basic shadow template If there are ≥2 high intensity light sources ( ), the basic shadow template Is replaced by the double light source shadow template .

[0183] The specific process of obtaining the double light source shadow template Is as follows:

[0184] According to the light source intensity Sorting, taking the first 2 high intensity light sources as the main light sources, denoted as light source 1 and light source 2.

[0185] According to the light source intensity ratio, the shadow weight is distributed to avoid secondary light source interference:

[0186] ;

[0187] In the formula, Indicates the shadow weight of light source 1; Indicates the intensity of light source 1; Indicates the shadow weight of light source 2; Indicates the intensity of light source 2; , the high intensity light source corresponds to higher shadow weight.

[0188] According to the light source direction Adjust the shadow projection angle (the shadow direction is opposite to the light source direction), generate the double light source shadow template , which can be expressed as:

[0189] ;

[0190] In the formula, Indicates the basic shadow template generated based on light source 1; Indicates the basic shadow template generated based on the light source 2; ensure that the shadow direction matches the actual light source, avoid disordered superposition.

[0191] The optimization algorithm adopts an improved RL-TLBO algorithm, and the target identification model adopts a YOLOv8 model; the improved RL-TLBO algorithm is used to optimize the YOLOv8 model parameters to obtain an optimal target identification model, and the specific process is as follows:

[0192] Step S3.1: Configure the improved RL-TLBO algorithm hyperparameters, including the population size , the maximum number of iterations , the initial learning rate , and the initial value of the inertia weight ; extract the convolution kernel weight in the YOLOv8 model backbone, the feature fusion weight of the intermediate processing network (Neck), and the classification weight and anchor frame regression weight of the detection head (Head) as the YOLOv8 model parameters to be optimized .

[0193] Step S3.2: Construct a weighted fitness function for the shadow scene, which quantifies the comprehensive performance of the YOLOv8 model in the dynamic shadow scene, and can be expressed as:

[0194] ;

[0195] In the formula, , represents the fitness value of the parameter ; , represents the score of the large construction machinery in the th shadow scene, reflecting the identification accuracy of the machinery, ; , represents the false alarm rate of non-threatening targets, reflecting the probability that non-threatening targets (personnel / vehicles / flying birds) are misjudged as machinery, avoiding invalid early warning; , represents the total number of missed detections of large machinery in the shadow scene, which is used to summarize the number of missed detections of machinery in the three types of shadow scenes, and the model is forced to reduce the missed detection rate through the penalty term; , , , all represent identification weights for balancing the optimization of each scene; , represents the weight of the false alarm rate on the fitness; , represents the missed detection penalty coefficient, which is used to adjust the strength of the missed detection penalty.

[0196] Step S3.3: Shadow feature-oriented Logistic mapping initial population generation;

[0197] Step S3.31: Generate random seeds of the initial population using the Logistic map with shadow luminance features, which makes the population parameters naturally adapt to the luminance differences of the shadow scene, which can be expressed as:

[0198]

[0199] wherein, represents the chaotic value of the th iteration; represents the chaotic value of the th iteration; represents the shadow luminance disturbance term, which adjusts the chaotic sequence by the normalized luminance value, making the population parameters more sensitive to the shadow luminance changes; represents the average luminance of the shadow area; represents the Logistic map parameter.

[0200] Step S3.32: Linearly map the generated chaotic values to the standard range of YOLOv8 model parameters to generate initial population individuals:

[0201]

[0202] wherein, represents the th parameter value of the th individual; , respectively represent the upper and lower limits of the th parameter of the individual; represents the input chaotic value.

[0203] wherein, the legal range of YOLOv8 model parameters is: convolution kernel weight , batch normalization (BN) layer parameter , detection head anchor frame regression parameter , classification confidence weight .

[0204] wherein, 500 enhanced images are used for rapid pre-evaluation of the initial population, the fitness of each individual is calculated, and the top 30% high fitness individuals are retained; the remaining individuals are re-generated through the above Logistic map to ensure that the initial population has basic shadow scene recognition ability.

[0205] Step S3.4: Update the parameters of the shadow scene dynamically based on SAC reinforcement learning;

[0206] Step S3.41: Parameter update in the teaching phase;

[0207] ​​The individual with the highest overall fitness in the current population is selected as the teacher. ; Calculate the current population average parameters: Update individual temporary parameters:

[0208] ;

[0209] In the formula, Indicates the first During the nth iteration, the 1st The first individual Updated values ​​for each parameter; Represents a random number; Indicates the first During the nth iteration, the 1st The first individual The original values ​​of each parameter; Indicates teaching factors, At the same time, strengthen local search. At that time, expand the search scope; Indicates population size; Indicates the first During the nth iteration, the 1st Individual; Represents the population mean parameter The Middle The average value of each parameter.

[0210] Step S3.42: Definition of SAC reinforcement learning state and reward (to address the optimization bottleneck of YOLOv8 model in shadow scenes, expand the state and reward dimensions of reinforcement learning).

[0211] Definition of the first The state at the next iteration , including the Population diversity at the next iteration , No. Fitness improvement rate at the next iteration and the Shadow scene fitness deviation in the next iteration ;

[0212] Population diversity It can be represented as:

[0213] ;

[0214] ;

[0215] In the formula, express The probability weights; express The fitness value.

[0216] Adaptability Improvement Rate It can be represented as:

[0217] ;

[0218] In the formula, express fitness value; Indicates the first The individual with the best parameters in the population at the next iteration; express fitness value; Indicates the first The individual with the best parameters in the population at the next iteration.

[0219] Shadow scene adaptation bias It can be represented as:

[0220] ;

[0221] In the formula, Indicates the first The fitness value in the scene under direct strong light at the next iteration; Indicates the first The fitness value in the backlight scene at the next iteration; Indicates the first The fitness value in low-light nighttime scenes during the next iteration.

[0222] Definition of the first Rewards at the next iteration (Guiding the algorithm's optimization direction through differentiated rewards) can be expressed as:

[0223] ;

[0224] Step S3.43: Dynamically update inertia weights: The SAC agent maximizes cumulative rewards. , This represents the discount factor, and the output is the optimal inertia weight adjustment. Update the current inertia weights:

[0225] ;

[0226] In the formula, Indicates the first Inertia weights in the next iteration; Indicates the first Inertia weights in the next iteration; Indicates the first Discount factor at the next iteration; Indicates the first Instantaneous reward at the next iteration.

[0227] Step S3.44: Student phase parameter update: randomly select two different individuals in the population 、 , update the final parameters in combination with the updated inertia weight :

[0228] ;

[0229] In the formula, represents the value of the th parameter of the th individual at the th iteration; represents a random number; represents the value of the th parameter of the th individual randomly selected from the population at the th iteration; represents the value of the th parameter of the th individual randomly selected from the population at the th iteration; represents the fitness value of ; represents the fitness value of .

[0230] Step S3.5: Perform local search;

[0231] Step S3.51: Set local search trigger conditions, trigger local search when any local search condition is met; the local search trigger conditions are:

[0232] Every 15 iterations are completed;

[0233] The fitness of a certain shadow scene is continuously improved by 3 generations <The first preset improvement rate threshold.

[0234] Step S3.52: YOLOv8 model parameter sensitive point positioning: focus on the parameter sensitive points that have the greatest impact on shadow identification, including:

[0235] The last two layers of convolution kernel weights in the backbone network (affecting high-level shadow feature extraction);

[0236] The up-sampling fusion weight of the intermediate processing network (Neck) (affecting multi-scale shadow feature fusion);

[0237] The anchor frame regression weight of the detection head (Head) (affecting the prediction of the mechanical boundary frame under the shadow).

[0238] Step S3.53: Add Gaussian disturbance to the parameter sensitive point to generate 5 candidate solutions:

[0239] ;

[0240] In the formula, represents the value of the i-th parameter in the i-th generated candidate solution, ; represents the value of the i-th parameter of the optimal parameter individual in the population at the j-th iteration; represents a Gaussian distribution with a mean of 0 and a variance of .

[0241] Step S3.54: Evaluate the fitness of the 5 candidate solutions , select the candidate solution with the highest fitness , and update the current optimal parameters:

[0242] ;

[0243] In the formula, represents the fitness value of ; represents the fitness value of .

[0244] Step S3.55: If the fitness has not been significantly improved for 5 generations < second preset improvement rate threshold), the current learning rate is exponentially decayed to ensure focusing on local fine optimization in the later period:

[0245] ;

[0246] In the formula, represents the learning rate at the j-th iteration; represents the learning rate at the j-th iteration. Step S3.6: Improve the convergence judgment and optimal target identification model generation of the RL-TLBO algorithm;

[0247] Step S3.61: Convergence judgment: the improved RL-TLBO algorithm is determined to converge when the following two conditions are met:

[0248]

[0249] the overall fitness change for consecutive generations, represents the value of the i-th parameter of the optimal parameter individual in the population at the j-th iteration; ​​​​​​The optimal fitness value of the population at the next iteration. Indicates the first The optimal fitness value of the population at the next iteration. This represents the first convergence threshold;

[0250] The fitness variation of all three types of shadow scenes satisfies , Indicates the first During the nth iteration, the 1st The optimal fitness value for a given scenario. Indicates the first During the nth iteration, the 1st The optimal fitness value for a given scenario. This represents the second convergence threshold.

[0251] Step S3.62: The finally converged optimal parameters... Load the YOLOv8 model and verify the preset metrics. Indicates the first In the next iteration, the individual with the best parameters in the population; preset indicators include:

[0252] Standard indicators: F1 score for large construction machinery recognition ≥ 0.90, false alarm rate for non-threat targets ≤ 6%, and inference time per image ≤ 180ms (retaining the lightweight characteristics of the original model).

[0253] Specific indicators for shadow scenes: accuracy rate of mechanical contour recognition in shadow areas ≥85%, false negative rate of key components (crane boom, bucket) under backlight occlusion shadows ≤9%, and false positive rate of mechanical components under low light shadows at night ≤5%.

[0254] Step S3.63: If all preset indicators are met, output the corresponding YOLOv8 model as the optimal target recognition model; otherwise, return to step S3.3 and adjust the Logistic mapping parameters. Then, re-execute the optimization process.

[0255] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for early warning of external damage hazards of a power transmission line based on a visualization device, characterized in that, The method comprises the following steps: Step S1: obtaining a plurality of transmission line images containing large construction machinery and preprocessing; Step S2: performing shadow synthesis on the preprocessed images to form enhanced images; and reducing the enhanced images to obtain reduced images; Performing feature extraction and fusion on the preprocessed images, the enhanced images and the reduced images to obtain a fused feature map; Step S3: constructing a target recognition model; training the target recognition model based on the fused feature map, and simultaneously optimizing the model parameters by using an optimization algorithm to obtain an optimal target recognition model; Step S4: obtaining real-time transmission line images through a visual device, and obtaining real-time fused feature maps by processing the real-time transmission line images through step S2; inputting the real-time fused feature maps into the optimal target recognition model to identify the large construction machinery in the images; when the large construction machinery is identified, proceeding to step S5; otherwise, proceeding to the next identification cycle; Step S5: scanning a target transmission line area by using a millimeter wave radar unit to obtain three-dimensional point cloud data of the large construction machinery; comparing the three-dimensional point cloud data with the large construction machinery identified by the optimal target recognition model in a two-dimensional space; when the three-dimensional point cloud data and the large construction machinery identified by the optimal target recognition model do not coincide in the two-dimensional space, proceeding to the next identification cycle; otherwise, proceeding to step S6; Step S6: presetting a conductor cross arm safety distance, combining the three-dimensional point cloud data and the projected area on the two-dimensional image coordinates, and calculating the height of the identified large construction machinery; when the height of the identified large construction machinery exceeds the preset safety distance, reporting an alarm; otherwise, proceeding to the next identification cycle; The specific process of performing shadow synthesis on the preprocessed images to form enhanced images is as follows: Step S2.31: Set each pre-processed image The height of the circumscribed rectangle of the medium and large-sized machine is ; each pre-processed image contains: Mechanical material tag ; Artificial light source parameters in a night-time scene, including light source direction and light source intensity ; Step S2.32: scene classification and parameter extraction; Classifying the scenes of each preprocessed image and extracting key parameters: Computing the average gray value of the brightness histogram of the pre-processed image , in combination with the real-time illumination angle collected by the illumination angle sensor , into a strong light direct scene, a backlight scene and a night weak light scene A U-Net semantic segmentation model is used to segment the preprocessed image, and an occlusion mask is output wherein, represents a pixel is an occluded region, represents a pixel is a non-occluded region; Occluder mask With dynamic occluded region mask Fusing, resulting in real-time occlusion mask ; Step S2.33: basic shadow template generation; based on a height of a bounding rectangle of the large machine in the pre-processed image with a real-time light angle , calculating a length of a shadow pixel projected by the large machine in the corresponding pre-processed image ; Computing average gray value of large machinery region in pre-processed image ; based on the average gray value of the large machinery region in the pre-processed image calculate the shadow gray value under different scenes ; Based on the length of the shadow pixels , the gray value of each pixel in the matrix is set as the gray value of the corresponding position in the image , and a basic shadow template is formed ; Step S2.34: occlusion shadow mask superposition; apply real-time occlusion mask to the base shadow template pixel-wise multiplication to get the shadow template with occlusion features ; Step S2.35: using a pixel-level weighted fusion algorithm, fusing the shadow template with the pre-processed image to generate an enhanced image , denoted as: ； In the formula, denotes the fusion weight. 2.The transmission line external damage hidden danger early warning method based on a visualization device according to claim 1, characterized in that: A dynamic occlusion region mask is obtained The specific process is as follows: by using an interframe difference method, pixel changes of adjacent five frames of images are compared, a dynamic occlusion region mask is extracted, and the dynamic occlusion region mask is represented as follows: The specific process is as follows: by using an interframe difference method, pixel changes of adjacent five frames of images are compared, a dynamic occlusion region mask is extracted, and the dynamic occlusion region mask is represented as follows: The specific process is as follows: by using an interframe difference method, pixel changes of adjacent five frames of images are compared, a dynamic ； wherein represents the first frame image at a pixel having a gray value; represents a gray difference threshold. 3.The transmission line external damage hidden danger early warning method based on a visualization device according to claim 2, characterized in that: Fusion weight is represented as: ； In the formula, indicates the shadow gray scale; indicates the large machine area average gray scale.

4. The method for early warning of external damage hazards of power transmission lines based on a visualization device according to claim 3, characterized in that: In a night-time low light scene, when there is only 1 high intensity light source, still use the base shadow template When there are ≥ 2 high intensity light sources, replace the base shadow template with a dual light source shadow template High intensity light source is defined as: ​ Obtaining a dual light source shadow mask The detailed process is as follows: According to the light source intensity Ranking, take the top 2 high-intensity light sources as the main light sources, recorded as light source 1 and light source 2; According to the light source intensity proportion, the shadow weight is distributed: ； wherein denotes the shadow weight of light source 1; denotes the intensity of light source 1; denotes the shadow weight of light source 2; denotes the intensity of light source 2; ; According to the direction of the light source Adjusting the shadow projection angle to generate a dual light source shadow template , is represented as: ； In the formula, denotes a base shadow mask generated based on light source 1 ; denotes a base shadow mask generated based on light source 2.

5. The method for early warning of external damage hazards of power transmission lines based on a visualization device according to claim 4, characterized in that: The optimization algorithm uses an improved RL-TLBO algorithm, and the target recognition model uses a YOLOv8 model; the parameters of the YOLOv8 model are optimized by using the improved RL-TLBO algorithm to obtain the optimal target recognition model, and the specific process is as follows: Step S3.1: configuring improved RL-TLBO algorithm hyperparameters, including population size , maximum number of iterations , initial learning rate , and inertia weight initial value ; extracting convolution kernel weights in the YOLOv8 model backbone network, feature fusion weights in the intermediate processing network, classification weights and anchor box regression weights of the detection head as YOLOv8 model parameters to be optimized ; Step S3.2: constructing a shadow scene weighted fitness function, which is expressed as: ； In the formula, represent the fitness value of the parameters ; represent the score of large construction machinery in the shadow scene of the first class, reflecting the identification accuracy of the machinery, ; represent the false alarm rate of non-threatening targets; represent the total number of missed detections of large machinery in the shadow scene; , , all represent identification weights for balancing the optimization of each scene; represent the weight of the control false alarm rate on the fitness; represent the missed detection penalty coefficient; Step S3.3: shadow feature oriented Logistic mapping initial population generation; Step S3.31: generating a random seed of the initial population by using a Logistic mapping integrated with shadow brightness features, which is expressed as: ； wherein represents the chaotic value of the first iteration; represents the chaotic value of the first iteration; represents the average luminance of the shadow region; represents the Logistic map parameter; Step S3.32: The generated chaotic values are mapped to the YOLOv8 model parameters through a linear mapping within the specification range, to generate initial population individuals; Step S3.4: dynamic updating of shadow scene parameters based on SAC reinforcement learning; Step S3.41: teaching stage parameter updating; select the individual with the highest comprehensive fitness in the current population as the teacher individual ; calculate the average parameters of the current population , update the temporary parameters of the individual ； wherein represents the th iteration, the updated value of the th parameter of the th individual; represents a random number; represents the th iteration, the original value of the th parameter of the th individual; represents a teaching factor; represents a population size; represents the th iteration, the th individual; represents the average value of the th parameter in the population average parameter ; Step S3.42: SAC reinforcement learning state and reward definition; Step S3.43: Dynamically update inertia weights: The SAC agent maximizes the cumulative reward. , This represents the discount factor, and the output is the optimal inertia weight adjustment. Update the current inertia weights , Indicates the first Inertia weights in the next iteration; Indicates the first Inertia weights in the next iteration; Indicates the first Discount factor at the next iteration; Indicates the first Instant rewards at the next iteration; Step S3.44: Student phase parameter update: randomly select two different individuals from the population , , combine the updated inertia weight , update the final parameters; Step S3.5: local search; Step S3.6: convergence judgment of the improved RL-TLBO algorithm and generation of the optimal target recognition model; Step S3.61: convergence judgment: the improved RL-TLBO algorithm is determined to converge when the following two conditions are met simultaneously: successively change in overall fitness , denotes the population optimal fitness value at the th iteration, denotes the population optimal fitness value at the th iteration, denotes a first convergence threshold value; The fitness variation of the three types of shadow scenes satisfies , represents the optimal fitness value of the first type of scene at the i-th iteration, represents the optimal fitness value of the second type of scene at the i-th iteration, represents the optimal fitness value of the third type of scene at the i-th iteration, represents the optimal fitness value of the first type of scene at the i-th iteration, represents the optimal fitness value of the second type of scene at the i-th iteration, represents the optimal fitness value of the third type of scene at the i-th iteration, represents a second convergence threshold value; Step S3.62: loading the final converged optimal parameters Load the YOLOv8 model and verify, represents the optimal parameter individual in the population at the th iteration; Step S3.63: If the verification passes, output the corresponding YOLOv8 model as the optimal target recognition model, otherwise return to step S3.3 to adjust the Logistic mapping parameters , and re-execute the optimization process.

6. The method for early warning of external damage hazards of power transmission lines based on a visualization device according to claim 5, characterized in that: The specific process of SAC reinforcement learning state and reward definition is as follows: Definition of states , including the diversity of the population at the first iteration , the fitness improvement rate at the first iteration , and the fitness deviation of the shadow scenario at the first iteration ; Population diversity is represented as: ； ； wherein represents a probability weight of represents a fitness value of .​ Rate of fitness increase is represented as: ； wherein denotes the fitness value of the individual denotes the fitness value of the individual denotes the best individual in the population at the i-th iteration denotes the best individual in the population at the i-th iteration denotes the fitness value of the individual denotes the fitness value of the individual denotes the best individual in the population at the i-th iteration denotes the best individual in the population at the i-th iteration Shadow scene fitness bias is represented as: ； In the formula, represents the fitness value of the scene under direct light at the first iteration; represents the fitness value of the scene under direct light at the first iteration; represents the fitness value of the scene under backlight at the first iteration; represents the fitness value of the scene under backlight at the first iteration; represents the fitness value of the scene under night weak light at the first iteration; represents the fitness value of the scene under night weak light at the first iteration; Definition of the first Reward at the second iteration is expressed as: 。 7. The method of claim 6, wherein the method further comprises: The specific process of local search is as follows: Step S3.51: Set the local search trigger condition, which is triggered when any of the local search conditions is met; the local search trigger condition is: per completed generation iteration; fitness continuity of certain shadow scenarios generation promotion rate <first preset promotion rate threshold value; Step S3.52: YOLOv8 model parameter sensitive point positioning: focusing on the parameter sensitive point that has the greatest impact on shadow identification; Step S3.53: Add Gaussian perturbation to the parameter sensitive point to generate multiple candidate solutions; Step S3.54: evaluating the fitness of the candidate solution , selecting the candidate solution with the highest fitness , updating the current best parameters; Step S3.55: If consecutive fitness is not significantly improved, i.e. if the second preset improvement rate threshold is reached, the current learning rate is exponentially decayed. 8.The method of visual device-based external damage early warning of power transmission lines according to claim 7, characterized in that: The specific process of reducing the enhanced image to obtain a reduced image is as follows: Step S2.11: Acquire the reference monitoring distance of the visualization device ; Step S2.12: Monitoring the distance according to the reference and the actual monitoring distance of the visualization device Calculating the reduction scale ; Step S2.13: applying a bilinear interpolation to the enhanced image in a reduced scale reducing to obtain a preliminary reduced image ; Step S2.14: performing Sobel edge sharpening on the preliminary down-scaled image Sobel edge sharpening to obtain a down-scaled image .

9. The method of claim 8, wherein the method further comprises: The specific process of feature extraction and fusion of the preprocessed image, the enhanced image and the reduced image to obtain a fused feature map is as follows: Step S2.21: obtaining the pre-processed image , an enhanced image , and a down-sized image of the real-time scene type label ; Step S2.22: respectively extracting the pre-processed image , the enhanced image and the reduced image to obtain pre-processed image feature maps , enhanced image feature maps and reduced image feature maps ; Step S2.23: mapping the large mechanical center coordinates in to the same coordinate system with as reference. , ​ Step S2.24: dynamically adjusting weights of the pre-processed image feature maps according to the scene type label , enhancing the image feature maps , and reducing the image feature maps , and the image feature maps ​ Step S2.25: concatenating the feature matrices along the channel dimension into one feature matrix , , ;​ Step S2.26: using the pre-processed image feature map to weight the image feature map , enhancing the image feature map to weight the image feature map , reducing the image feature map to weight the image feature map , and weighting fusion of the feature matrix to generate the final fusion feature map .

Images