Unmanned aerial vehicle intelligent shooting control method and system for distribution network inspection

Abstract

This invention discloses an intelligent shooting control method and system for power distribution network inspection using unmanned aerial vehicles (UAVs), belonging to the field of intelligent control technology. The method includes: acquiring perception evaluation parameters of the current captured image and the component importance weights corresponding to the current captured target; analyzing and obtaining a comprehensive shooting quality deviation index; inputting the comprehensive shooting quality deviation index and component importance weights into a pre-trained shooting parameter adjustment strategy generation model to obtain a shooting parameter adjustment strategy; controlling the UAV gimbal camera to perform adjustment actions according to the shooting parameter adjustment strategy and re-acquiring the captured image; inputting the re-acquired image into a pre-trained image quality evaluation model to obtain a quality pass probability; and performing adversarial incremental optimization of the shooting parameter adjustment strategy generation model based on the quality pass probability and the comprehensive shooting quality deviation index until the quality pass probability meets a preset condition. This invention effectively improves the shooting quality and detection reliability of power distribution network inspection images.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to an intelligent shooting control method and system for power distribution network inspection using unmanned aerial vehicles (UAVs). Background Technology

[0002] With the widespread application of drone technology in power distribution network inspection, automated shooting using drones equipped with gimbal cameras has become the mainstream approach. In existing technologies, drone shooting control mainly relies on manual remote control or preset fixed shooting parameters; some solutions also adjust based on simple image brightness or sharpness indicators.

[0003] However, existing shooting control methods have many technical shortcomings. Distribution network equipment has a complex structure, and different components have significantly different requirements for image details. Existing methods do not implement differentiated shooting strategies for key distribution network components such as insulators and clamps, resulting in unstable imaging quality and low accuracy in component defect identification. At the same time, the shooting parameter decision model lacks an online optimization mechanism, making it difficult to adapt to distribution network inspection scenarios with complex terrain and variable lighting, resulting in insufficient shooting quality and detection reliability of distribution network inspection images. Summary of the Invention

[0004] This invention provides an intelligent shooting and control method and system for distribution network inspection using unmanned aerial vehicles (UAVs), aiming to solve the technical problems of insufficient shooting quality and detection reliability of distribution network inspection images in the prior art.

[0005] In view of the above problems, the present invention provides a method and system for intelligent shooting and control of drones for power distribution network inspection.

[0006] In a first aspect, the present invention provides an intelligent shooting and control method for unmanned aerial vehicles (UAVs) for power distribution network inspection, including: Obtain the perception evaluation parameters of the currently captured image, wherein the perception evaluation parameters include the target space occupancy ratio, texture clarity attenuation degree, and illumination zone contrast value; Obtain the importance weights of the components corresponding to the current shooting target; Based on the perception evaluation parameters and the importance weights of the components, the comprehensive shooting quality deviation index is analyzed and obtained; The comprehensive shooting quality deviation index and the component importance weights are input into a pre-trained shooting parameter adjustment strategy generation model to obtain a shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment and metering adjustment. The drone gimbal camera is controlled to perform adjustment actions according to the shooting parameter adjustment strategy, and the shooting image is re-acquired; The re-acquired footage is input into a pre-trained image quality assessment model to obtain the probability of passing quality assessment. Based on the quality pass probability and the comprehensive shooting quality deviation index, the shooting parameter adjustment strategy generation model is subjected to adversarial incremental optimization until the quality pass probability meets the preset conditions.

[0007] Secondly, this invention provides an intelligent drone shooting control system for power distribution network inspection, comprising: The perception parameter acquisition module is used to acquire the perception evaluation parameters of the current shooting image, wherein the perception evaluation parameters include the target space occupancy ratio, texture clarity attenuation degree, and illumination zone contrast value. The component weight acquisition module is used to acquire the importance weight of the components corresponding to the current shooting target; The deviation index calculation module is used to analyze and obtain the comprehensive shooting quality deviation index based on the perception evaluation parameters and the importance weights of the components. The adjustment strategy generation module is used to input the comprehensive shooting quality deviation index and the component importance weight into the pre-trained shooting parameter adjustment strategy generation model to obtain the shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment and metering adjustment. The camera adjustment framing module is used to control the UAV gimbal camera to perform adjustment actions according to the shooting parameter adjustment strategy, and to reacquire the shooting image; The quality probability assessment module is used to input the re-acquired captured images into the pre-trained image quality assessment model to obtain the probability of quality passing. The adversarial incremental optimization module is used to perform adversarial incremental optimization on the shooting parameter adjustment strategy generation model based on the quality pass probability and the comprehensive shooting quality deviation index, until the quality pass probability meets the preset conditions.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides an intelligent shooting control method and system for power distribution network inspection using unmanned aerial vehicles (UAVs). By quantifying multi-dimensional perception and evaluation parameters and component importance weights, a comprehensive shooting quality deviation index is generated. Relying on an intelligent model, precise zoom, focus, and metering adjustment strategies are output, which can adaptively optimize camera shooting parameters in real time, effectively improving the clarity and target integrity of the captured images. Furthermore, through adversarial incremental optimization and continuous model iteration, the scene adaptability is enhanced, solving the problems of poor shooting quality and weak adaptability in existing technologies, and ensuring the accuracy and stability of power distribution network inspection shooting. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the intelligent shooting and control method for power distribution network inspection provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the UAV intelligent shooting control system for power distribution network inspection provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: Perception parameter acquisition module 11, component weight acquisition module 12, deviation index calculation module 13, adjustment strategy generation module 14, camera adjustment framing module 15, quality probability evaluation module 16, and adversarial incremental optimization module 17. Detailed Implementation

[0010] This invention provides an intelligent shooting control method and system for distribution network inspection using unmanned aerial vehicles (UAVs), which addresses the technical problems of insufficient image shooting quality and detection reliability in existing distribution network inspection technologies.

[0011] Example 1, as Figure 1 As shown, this invention provides an intelligent drone shooting control method for power distribution network inspection, the method comprising: S100: Obtain the perception evaluation parameters of the currently captured image, wherein the perception evaluation parameters include the target space occupancy ratio, texture clarity attenuation degree, and illumination zone contrast value.

[0012] In this embodiment of the invention, perception evaluation parameters of the currently captured image are obtained. These parameters include target space occupancy ratio, texture clarity attenuation, and illumination zone contrast value. In power distribution network inspection scenarios, when drones capture images of key power components such as insulators, clamps, and crossarms, the reasonableness of the target occupancy ratio, texture clarity, and illumination distribution uniformity are core factors determining image quality and subsequent defect identification accuracy. If image quality is evaluated solely based on a single indicator or subjective judgment, imaging defects cannot be accurately quantified, nor can data support be provided for adaptive adjustment of camera parameters. Therefore, it is necessary to quantify and extract perception evaluation parameters from three dimensions: target space occupancy ratio, texture clarity attenuation, and illumination zone contrast value. This achieves a standardized representation of the perceived quality of the captured image, providing a data foundation for subsequent comprehensive image quality assessment.

[0013] Step S100 in the method provided in this embodiment of the invention includes: Obtain the pixel area value of the target detection box in the current shooting frame as the detection box area, and obtain the total pixel value of the current shooting frame as the total area of ​​the frame; The target space occupancy ratio is obtained by dividing the area of ​​the detection frame by the total area of ​​the screen. Edge detection is performed on the image region within the target detection box to obtain the average edge intensity as the current edge intensity; the reference edge intensity is obtained according to the preset target type-distance-reference edge intensity three-dimensional mapping table, and the texture sharpness attenuation is calculated, where texture sharpness attenuation = 1 - (current edge intensity / reference edge intensity); The current captured image is divided into a preset number of grid regions, and the average brightness value of each grid region is obtained; the standard deviation of the average brightness values ​​of all grid regions is calculated and recorded as the brightness standard deviation, and the arithmetic mean of the average brightness values ​​of all grid regions is calculated as the average brightness of the entire image. The brightness standard deviation is calculated and divided by the average brightness of the entire screen to obtain the contrast value of the illumination zone.

[0014] First, the pixel area of ​​the target detection box in the current captured image is obtained as the detection box area, and the total pixel value of the current captured image is obtained as the total image area. The target detection box refers to the rectangular positioning border formed in the captured image by the distribution network inspection target identified by the target detection algorithm. The detection box area is the total number of pixels covered by the target detection box. The total image area is the total number of pixels in the current drone-captured image.

[0015] Specifically, the target detection module on the drone identifies power distribution network components in the captured image, outputs the pixel size of the target detection frame, and calculates the area of ​​the detection frame. The camera's imaging resolution parameters are then retrieved to calculate the total number of pixels in the captured image, i.e., the total area of ​​the image. For example, if the drone captures an image with a resolution of 1920×1080, the total area of ​​the image is 1920×1080=2073600 pixels; the size of the detected insulator string target detection frame is 384×270, and the detection frame area is 384×270=103680 pixels.

[0016] Next, the area of ​​the detection frame is divided by the total area of ​​the image to obtain the target space occupancy ratio. The target space occupancy ratio refers to the proportion of pixel area occupied by the inspected target in the captured image, used to determine whether the target's proportion in the image is reasonable. The ratio obtained by dividing the area of ​​the detection frame by the total area of ​​the image is the target space occupancy ratio. For example, a target space occupancy ratio of 103680 / 2073600 = 0.05 means that the insulator string target occupies 5% of the captured image.

[0017] Next, edge detection is performed on the image region within the target detection box to obtain the average edge intensity, which is taken as the current edge intensity. A reference edge intensity is obtained according to a preset target type-distance-reference edge intensity three-dimensional mapping table, and texture sharpness attenuation is calculated: Texture sharpness attenuation = 1 - (current edge intensity / reference edge intensity). Texture sharpness attenuation represents the degree of attenuation of the target region's texture sharpness compared to the ideal sharpness; a larger value indicates more severe texture blurring. The average edge intensity is the arithmetic mean of the intensity of all edge pixels after edge detection within the target detection box. The reference edge intensity refers to the ideal edge intensity standard value obtained by looking up the target type-distance-reference edge intensity three-dimensional mapping table, which is established through experimental calibration, based on the target type and the distance between the UAV and the target.

[0018] The target type-distance-reference edge intensity three-dimensional mapping table is pre-set according to the following method: For each target type, multiple standard images with clear focus and uniform illumination are acquired at different distances, the average edge intensity within the target detection box is extracted, and after statistical analysis, the median or average value is taken as the reference edge intensity of the target type at that distance.

[0019] Specifically, edge detection is performed on the image region within the target detection box, and the average edge intensity is calculated to obtain the current edge intensity. A reference edge intensity is obtained by combining the target type and the measured distance data from the UAV. Texture sharpness attenuation is calculated as 1 - (current edge intensity / reference edge intensity). When the current edge intensity is greater than or equal to the reference edge intensity, the texture sharpness attenuation is 0, indicating no sharpness attenuation. For example, when performing edge detection on the insulator string detection box region, the current edge intensity is 40; when the UAV measures a distance of 10 meters, the reference edge intensity is 50, and the texture sharpness attenuation is 1 - 40 / 50 = 0.2.

[0020] Then, the currently captured image is divided into a preset number of grid regions, and the average brightness value of each grid region is obtained. The standard deviation of the average brightness values ​​of all grid regions is calculated and denoted as the brightness standard deviation. The arithmetic mean of the average brightness values ​​of all grid regions is then calculated as the average brightness of the entire image. A grid region refers to several rectangular sub-regions that evenly divide the captured image. The brightness standard deviation is the dispersion of the average brightness of each grid region. The average brightness of the entire image is the arithmetic mean of the average brightness of all grid regions.

[0021] Specifically, the captured image is divided into a preset number of grid areas, and the average brightness of each grid is calculated. Then, the standard deviation and arithmetic mean of the average brightness of all grids are calculated. For example, if the captured image is evenly divided into 10×10 grid areas, the calculated brightness standard deviation is 15, and the average brightness of the entire image is 75.

[0022] Finally, the brightness standard deviation is calculated and divided by the average brightness of the entire image to obtain the illumination zone contrast value. The illumination zone contrast value is the ratio of the brightness difference in different areas of the captured image to the average brightness of the entire image, representing the uniformity of illumination in the image. Dividing the brightness standard deviation by the average brightness of the entire image yields the illumination zone contrast value. For example, the illumination zone contrast value = 15 / 75 = 0.2.

[0023] In this embodiment of the invention, quantitative perception evaluation parameters are obtained through standardized calculation from three dimensions: target proportion, texture clarity, and illumination uniformity. This enables a digital and objective evaluation of the image quality captured during power distribution network inspections, avoiding errors caused by subjective judgment. The calculation logic of each parameter aligns with the imaging requirements of power distribution network component inspections, taking into account both target positioning and detail clarity. This provides accurate and multi-dimensional basic data for subsequent calculation of the comprehensive shooting quality deviation index, ensuring the targeted and effective adjustment of subsequent shooting parameters.

[0024] S200: Obtain the importance weight of the component corresponding to the current shooting target.

[0025] In this embodiment of the invention, the importance weight of the component corresponding to the current shooting target is obtained. The impact of faults in components such as insulator strings, clamps, and pins involved in distribution network inspections on the safe operation of the distribution network varies significantly. The imaging quality of key components directly determines the accuracy of defect identification and the effectiveness of the inspection. If a uniform shooting control standard is applied to all inspected components, there will be problems such as insufficient shooting accuracy for highly important components and excessive resource allocation for less important components. Therefore, it is necessary to determine the type of component to be shot based on UAV positioning information, and then match it with a preset importance weight to achieve quantitative differentiation of component importance, providing a differentiated weight basis for subsequent shooting quality evaluation and parameter adjustment.

[0026] Step S200 in the method provided in this embodiment of the invention includes: The component type corresponding to the current shooting point is obtained based on the current coordinates of the drone, wherein the component type includes insulator string, wire clamp, pin, bolt and crossarm; Based on the component type, the corresponding importance weight is retrieved from a pre-established component type-importance weight mapping table and used as the component importance weight.

[0027] First, based on the drone's current coordinates, the component type corresponding to the current shooting point is obtained. This component type includes insulator strings, clamps, pins, bolts, and crossarms. The drone's current coordinates refer to the real-time spatial latitude and longitude coordinates collected by the BeiDou / GPS positioning module onboard the drone, used to accurately locate the drone's inspection position. The component types are classifications of distribution network line inspection components, specifically including five categories: insulator strings, clamps, pins, bolts, and crossarms.

[0028] Specifically, the system retrieves the real-time spatial positioning coordinates uploaded by the drone and compares them with a pre-stored database of spatial coordinates for distribution network components to pinpoint the inspection target corresponding to the current shooting point, thereby determining the type of component. For example, if the drone's real-time coordinates are 108.9°E, 34.3°N, after importing these coordinates into the distribution network component coordinate database for matching, the system determines that the component type corresponding to the current shooting point is an insulator string.

[0029] Secondly, based on the component type, the corresponding importance weight is retrieved from a pre-established component type-importance weight mapping table and used as the component importance weight. The component importance weight is a quantified value pre-set based on the fault severity level and maintenance priority of the distribution network component; a higher value indicates a higher level of component importance. The component type-importance weight mapping table is a pre-established one-to-one correspondence data table between component types and their corresponding importance weights, based on power grid operation and maintenance specifications and the impact of component faults.

[0030] Specifically, using the identified component type as the query index, a search and matching process is performed in a pre-built component type-importance weight mapping table. The corresponding weight data in the table is directly retrieved, which is the component importance weight of the current target being photographed. For example, in the pre-built mapping table, insulator strings correspond to a weight of 0.9, wire clamps to 0.8, pins to 0.6, bolts to 0.5, and crossarms to 0.7; if the current component type is an insulator string, querying the mapping table yields a component importance weight of 0.9.

[0031] In this embodiment of the invention, the type of inspection component is accurately matched by the drone coordinates, and the quantitative importance weight is quickly obtained by combining the preset mapping table, thereby realizing the standardized differentiation of the importance of different components of the power distribution network. This provides weight parameters for the subsequent weighted calculation of the comprehensive shooting quality deviation index, allowing the shooting control logic to tilt towards high-importance components, avoiding the average allocation of resources, ensuring that the shooting quality of key power distribution network components meets the inspection requirements first, and improving the pertinence and practicality of the overall inspection.

[0032] S300: Based on the perception evaluation parameters and the importance weights of the components, analyze and obtain the comprehensive shooting quality deviation index.

[0033] In this embodiment of the invention, a comprehensive shooting quality deviation index is obtained based on the perceived evaluation parameters and the importance weights of the components. The numerical magnitudes and evaluation dimensions of target space occupancy ratio, texture clarity attenuation, and illumination zone contrast values ​​are different, making it impossible to directly sum them to reflect the overall shooting quality deviation. Furthermore, the importance of different network components varies, leading to different requirements for shooting quality. Simply calculating the original parameters directly would result in distorted shooting quality evaluation due to inconsistent dimensions and unreasonable weights, failing to provide accurate quantitative guidance for subsequent shooting parameter adjustments. Therefore, it is necessary to normalize each perceived evaluation parameter, assign reasonable weights based on historical data, and incorporate component importance weights for weighted calculation. Finally, a quantitative index that comprehensively reflects the degree of shooting quality deviation from the ideal state is obtained, providing a unified evaluation basis for generating adjustment strategies.

[0034] Step S300 in the method provided in this embodiment of the invention includes: Multiply the texture clarity attenuation by the first weighting coefficient to obtain the first deviation component; Calculate the reciprocal of the target space occupancy ratio, and divide the reciprocal of the target space occupancy ratio by the upper limit of the reciprocal of the occupancy ratio to obtain the normalized value of the occupancy ratio; Multiply the occupancy ratio normalization value by the second weighting coefficient to obtain the second deviation component; The maximum value of the contrast value of the illumination zone is obtained from the historical inspection data in advance, and is used as the upper limit of the contrast value. The contrast value of the illumination zone is divided by the upper limit of the contrast value to obtain the contrast normalization value. The third deviation component is obtained by multiplying the contrast normalization value by the component importance weight, and then multiplying it by the third weighting coefficient. The first deviation component, the second deviation component, and the third deviation component are added together to obtain the comprehensive shooting quality deviation index, wherein the sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is 1, and each weight coefficient is obtained according to the frequency ratio of shooting failure caused by the corresponding parameter in the historical inspection.

[0035] First, the texture sharpness attenuation is multiplied by a first weighting coefficient to obtain the first deviation component. The first deviation component is a shooting quality deviation component composed of the texture sharpness attenuation, used to characterize the quality deviation caused by target texture blur. The first weighting coefficient is a weighting coefficient determined based on the frequency of shooting failures caused by unclear textures in historical inspections, and its sum with the second and third weighting coefficients is 1. Multiplying the texture sharpness attenuation obtained in S100 by the preset first weighting coefficient yields the first deviation component.

[0036] For example, the calculated texture clarity attenuation is 0.2; according to historical inspection data, the frequency of image capture failure due to texture blur is 40%, so the first weighting coefficient is set to 0.4; the first deviation component = 0.2 × 0.4 = 0.08.

[0037] Next, the reciprocal of the target space occupancy ratio is calculated, and then divided by the upper limit of the reciprocal of the occupancy ratio to obtain the normalized occupancy ratio value. The reciprocal of the occupancy ratio is the inverse of the target space occupancy ratio; the smaller the proportion, the larger the reciprocal, indicating a greater degree of deviation of the target from the center of the frame. The upper limit of the reciprocal of the occupancy ratio is the maximum reciprocal value statistically analyzed from historical inspection data when the target proportion is too small, resulting in invalid shooting, and is used for normalization constraints. The normalized occupancy ratio value refers to a normalized value that maps the reciprocal of the occupancy ratio to the range of 0 to 1, eliminating the influence of the original numerical magnitude.

[0038] Specifically, first calculate the reciprocal of the target space occupancy ratio, then divide this reciprocal by the upper limit of the reciprocal of the occupancy ratio to obtain the normalized occupancy ratio value. For example, if the target space occupancy ratio is 0.05, its reciprocal = 1 / 0.05 = 20; set the upper limit of the reciprocal of the occupancy ratio to 20; the normalized occupancy ratio value = 20 / 20 = 1.

[0039] Next, the normalized occupancy ratio is multiplied by the second weighting coefficient to obtain the second deviation component. The second deviation component is a shooting quality deviation component composed of the normalized target space occupancy ratio, used to characterize the quality deviation caused by an unreasonable target proportion. The second weighting coefficient is a weighting coefficient determined based on the frequency of shooting failures caused by target proportion imbalance in historical inspections. Multiplying the normalized occupancy ratio by the second weighting coefficient yields the second deviation component. For example, if the frequency of shooting failures due to target proportion imbalance in historical data is 30%, and the second weighting coefficient is set to 0.3, then the second deviation component = 1 × 0.3 = 0.3.

[0040] Furthermore, the maximum contrast value of the illumination zone, pre-calculated from historical inspection data, is obtained as the upper limit of the contrast value. The contrast value of the illumination zone is then divided by this upper limit to obtain the contrast normalization value. The upper limit of the contrast value refers to the maximum critical value of the contrast value of the illumination zone obtained from historical inspections; exceeding this value indicates severe uneven illumination. The contrast normalization value is a value that normalizes the contrast value of the illumination zone to the range of 0 to 1, unifying the units for weighted calculation. Dividing the contrast value of the illumination zone by the upper limit yields the contrast normalization value. For example, if the contrast value of the illumination zone is 0.2, and the upper limit of the historical contrast value is set to 0.5, the contrast normalization value = 0.2 / 0.5 = 0.4.

[0041] Subsequently, the contrast normalization value is calculated by multiplying it by the component importance weight, and then by a third weighting coefficient to obtain the third deviation component. The third deviation component is a deviation component that integrates the contrast of the illumination zones and the component importance, taking into account both the degree of illumination unevenness and the criticality of the component. The third weighting coefficient is a weighting coefficient determined based on the frequency of shooting failures caused by illumination unevenness in historical inspections. First, the contrast normalization value is multiplied by the component importance weight, and then multiplied by the third weighting coefficient to obtain the third deviation component. For example, if S200 obtains a component importance weight of 0.9, the frequency of shooting failures caused by illumination unevenness accounts for 30%, and the third weighting coefficient is set to 0.3; then the third deviation component = 0.4 × 0.9 × 0.3 = 0.108.

[0042] Finally, the first deviation component, the second deviation component, and the third deviation component are added together to obtain the comprehensive shooting quality deviation index. The sum of the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient is 1. Each weighting coefficient is obtained based on the frequency ratio of shooting failures caused by the corresponding parameter in historical inspections. The comprehensive shooting quality deviation index is a comprehensive quality deviation indicator that integrates texture, proportion, lighting, and component importance. The larger the value, the more the shooting quality deviates from the ideal state. The comprehensive shooting quality deviation index is obtained by directly adding the first deviation component, the second deviation component, and the third deviation component; and the sum of the first, second, and third weighting coefficients is strictly 1. For example, the comprehensive shooting quality deviation index = 0.08 + 0.3 + 0.108 = 0.488.

[0043] In this embodiment of the invention, normalization processing eliminates the differences in the dimensions and numerical magnitudes of different perception evaluation parameters, avoiding evaluation imbalance caused by large ranges in parameter values; weights are assigned based on the frequency of shooting failures, which aligns with the actual scenario of power distribution network inspection, while component importance weights are incorporated to achieve differentiated weighting; the final integrated shooting quality deviation index can quantify the degree of deviation in the overall shooting quality of the current image, providing a precise and unified input indicator for the generation of subsequent shooting parameter adjustment strategies.

[0044] S400: Input the comprehensive shooting quality deviation index and the component importance weight into the pre-trained shooting parameter adjustment strategy generation model to obtain the shooting parameter adjustment strategy, wherein the shooting parameter adjustment strategy includes adjustment type and adjustment amplitude, and the adjustment type includes zoom adjustment, focus adjustment and metering adjustment.

[0045] In this embodiment of the invention, the comprehensive shooting quality deviation index and the component importance weights are input into a pre-trained shooting parameter adjustment strategy generation model to obtain a shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment, and metering adjustment. Manually formulated parameter adjustment rules are difficult to adapt to the complex lighting, distance, and component type scenarios in power grid inspection, easily leading to over-adjustment or under-scheduling. Furthermore, the rationality of shooting parameter adjustments needs objective verification. Therefore, a shooting parameter adjustment strategy generation model is constructed. Through adversarial pre-training to learn historical high-quality adjustment logic, the quantified quality deviation and component weights can be directly transformed into precise and executable camera adjustment commands, while ensuring the effectiveness of the adjustment strategy, providing a basis for the automated control of UAV gimbal cameras.

[0046] Step S400 in the method provided in this embodiment of the invention includes: The comprehensive shooting quality deviation index of the current shooting image and the component importance weight are input into the pre-trained shooting parameter adjustment strategy to generate the model; The adjustment type and adjustment amplitude prediction values ​​output by the shooting parameter adjustment strategy generation model are obtained as the shooting parameter adjustment strategy, wherein the shooting parameter adjustment strategy includes adjustment type and adjustment amplitude, and the adjustment type includes zoom adjustment, focus adjustment and metering adjustment.

[0047] The pre-training of the model generated by the shooting parameter adjustment strategy includes: Collect multiple sets of shooting adjustment records from historical inspections. Each set of shooting adjustment records includes the overall shooting quality deviation index of the image before adjustment, the corresponding component importance weight, and the adjustment type and adjustment amplitude that have been manually confirmed as valid, as sample data. A shooting strategy generation model is constructed, which takes the comprehensive shooting quality deviation index and component importance weight as input and the predicted adjustment type and adjustment magnitude as output. An image quality evaluation model is constructed, which takes the adjusted captured image as input and the predicted quality pass probability as output. The shooting strategy generation model and the image quality evaluation model are trained alternately using the sample data. When the shooting parameter adjustment strategy generated by the shooting strategy generation model enables the quality pass probability output by the image quality evaluation model to reach a preset convergence threshold, the pre-trained shooting strategy generation model and image quality evaluation model are obtained, and the shooting parameter adjustment strategy generation model is integrated.

[0048] First, multiple sets of shooting adjustment records from historical inspections were collected. Each set of shooting adjustment records includes the overall shooting quality deviation index of the image before adjustment, the corresponding component importance weights, and the adjustment type and magnitude that have been manually confirmed as valid, serving as sample data. Shooting adjustment records refer to complete shooting adjustment data that have been manually confirmed as valid during historical inspections, and serve as supervised samples for model training.

[0049] Specifically, the dataset is derived from historical data of power distribution network drone field inspections, with 100,000 valid samples selected. Each sample includes the overall shooting quality deviation index before adjustment, component importance weight, effective manual adjustment type, and effective manual adjustment amplitude. The dataset is divided into 80,000 training sets and 20,000 validation sets. For example, a single sample might have the following characteristics: overall shooting quality deviation index before adjustment 0.5, component importance weight 0.8, and effective manual adjustment type and amplitude [zoom × 1.5, focus + 15 steps, metering + 0.3 EV].

[0050] Secondly, a shooting strategy generation model is constructed. This model takes the comprehensive shooting quality deviation index and component importance weights as input, and the predicted adjustment type and adjustment magnitude as output. The shooting parameter adjustment strategy generation model is a general term, including a generation part and a verification part. The shooting strategy generation model is the generation part, and the image quality evaluation model is the verification part. The shooting strategy generation model is responsible for outputting adjustment instructions and employs a multilayer perceptron (MLP) adapted to low-dimensional feature input.

[0051] Specifically, the shooting strategy generation model is constructed using a multilayer perceptron, consisting of an input layer, two hidden layers, and a dual-branch output layer. The input layer has two neurons that receive two input features: the comprehensive shooting quality deviation index and the component importance weights. The first hidden layer has 64 neurons that use the ReLU activation function and are mainly responsible for the initial extraction of input features. The second hidden layer has 32 neurons that also use the ReLU activation function and are used for deep fusion of the extracted features. The output layer is divided into a classification branch and a regression branch. The classification branch has four neurons that use the Softmax activation function and are used to output the adjustment type. The regression branch has three neurons that use the Linear activation function and are used to output the adjustment amplitude. Overall, it realizes a multi-task learning function with the comprehensive shooting quality deviation index and component importance weights as inputs and the adjustment type and adjustment amplitude as outputs.

[0052] The adjustment types are fixed as three categories: zoom adjustment, focus adjustment, and metering adjustment; the adjustment amplitude is the quantitative execution parameter for each type: zoom adjustment amplitude: focal length change factor (×1.0~×3.0); focus adjustment amplitude: focus motor step number + direction (±N steps, positive for telephoto, negative for near-photo); metering adjustment amplitude: exposure compensation value (EV, -2.0~+2.0).

[0053] Next, an image quality assessment model is constructed. This model takes the adjusted captured image as input and outputs the predicted probability of acceptable quality. The image quality assessment model is responsible for evaluating the quality of the adjusted image and employs a convolutional neural network (CNN) adapted to the image input.

[0054] Specifically, the image quality assessment model has the following structure: Input layer: 1920×1080×3 RGB image; Convolutional layer 1: 32 3×3 convolutional kernels, ReLU activation, max pooling; Convolutional layer 2: 64 3×3 convolutional kernels, ReLU activation, max pooling; Convolutional layer 3: 128 3×3 convolutional kernels, ReLU activation, max pooling; Fully connected layer 1: 64 neurons, ReLU activation; Fully connected layer 2: 32 neurons, ReLU activation; Output layer: 1 neuron, Sigmoid activation, outputting a quality pass probability of 0~1.

[0055] Finally, the sample data is used to perform alternating adversarial training on the shooting strategy generation model and the image quality evaluation model. When the shooting parameter adjustment strategy generated by the shooting strategy generation model enables the image quality evaluation model to achieve the quality pass probability output by the image quality evaluation model to reach a preset convergence threshold, the pre-trained shooting strategy generation model and image quality evaluation model are obtained, and the shooting parameter adjustment strategy generation model is integrated. Alternating adversarial training refers to fixing the parameters of one model while training the other model, iterating in a loop to achieve mutual optimization between the generation model and the evaluation model.

[0056] Specifically, an alternating adversarial training method is used to iteratively optimize the two models. The training process is executed in two cyclical phases: In the first phase, the parameters of the shooting strategy generation model are fixed, and the simulated shooting images corresponding to the model's output adjustment strategy are used as negative samples, while the actual shooting images corresponding to the effective manual adjustment strategy are used as positive samples. The binary cross-entropy loss function (BCE) is used to train the image quality evaluation model. In the second phase, the parameters of the image quality evaluation model are fixed, and the optimization objective is to maximize the probability of the model's output quality being acceptable. The parameters of the shooting strategy generation model are updated using a joint loss function consisting of classification cross-entropy loss (CE) and mean squared error loss (MSE).

[0057] The specific conditions for model convergence are as follows: the shooting parameter adjustment strategy generated by the shooting strategy generation model can make the output quality qualification probability of the image quality evaluation model ≥ 95% of the preset convergence threshold, and the decrease of the joint loss function in 10 consecutive iterations is < 1%, and the accuracy fluctuation of the validation set is < 0.5%.

[0058] For example, 100,000 sets of historical shooting and adjustment records of power distribution network inspections collected in the early stage are used as training data. They are divided into 80,000 training sets and 20,000 validation sets in an 8:2 ratio for training. During the iteration process, the changes in the quality pass probability and model loss are monitored in real time. When the quality pass probability output by the image quality evaluation model is consistently stable and not lower than 95%, and the decrease in the joint loss function of 10 consecutive training rounds is less than 1%, the training is stopped. The converged shooting strategy generation model is then integrated with the image quality evaluation model to obtain the trained shooting parameter adjustment strategy generation model.

[0059] Then, the overall shooting quality deviation index and the component importance weight of the current shooting image are input into the pre-trained shooting parameter adjustment strategy generation model. The overall shooting quality deviation index calculated by S300 and the component importance weight obtained by S200 are retrieved, standardized, and then input into the pre-trained shooting parameter adjustment strategy generation model. For example, two sets of data are input: overall shooting quality deviation index = 0.488, component importance weight = 0.9.

[0060] Subsequently, the adjustment type and predicted adjustment amplitude values ​​output by the shooting parameter adjustment strategy generation model are obtained as the shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude, and the adjustment type includes zoom adjustment, focus adjustment, and metering adjustment. The shooting parameter adjustment strategy is a camera execution instruction output by the shooting parameter adjustment strategy generation model, containing adjustment type and adjustment amplitude. The shooting parameter adjustment strategy generation model performs inference calculations on the input features and directly outputs the combination of adjustment types and the corresponding predicted adjustment amplitude values ​​as the final executable adjustment strategy. For example, the output strategy might be: zoom adjustment ×1.5x, focus adjustment +15 steps, metering adjustment +0.3EV.

[0061] In this embodiment of the invention, a pre-trained shooting parameter adjustment strategy generation model directly transforms abstract shooting quality deviation data into three quantifiable and executable camera adjustment strategies: zoom, focus, and metering. This replaces manual experience-based judgment, achieving intelligent and automated shooting control. An alternating adversarial training method is employed, with the generation part responsible for outputting instructions and the verification part responsible for validating the effects, improving the accuracy and effectiveness of the adjustment strategies. The shooting parameter adjustment strategy generation model has a lightweight structure, is adaptable to embedded deployment on drones, and has a fast inference speed, meeting the operational requirements of real-time adjustment during power distribution network inspection.

[0062] S500: Controls the drone gimbal camera to perform adjustment actions according to the shooting parameter adjustment strategy, and re-acquires the shooting image.

[0063] In this embodiment of the invention, the drone gimbal camera is controlled to perform adjustment actions according to the shooting parameter adjustment strategy, and the shooting image is re-acquired. The adjustment logic and execution method of the three parameters—zoom, focus, and metering—are independent of each other. If hardware command conversion and precise execution are not performed in a multi-dimensional manner, the adjustment strategy will fail to be implemented, and the quality of the shooting image cannot be effectively improved. Therefore, it is necessary to convert the various amplitude parameters in the shooting parameter adjustment strategy into executable control actions for camera focal length, focus position, and metering weight, respectively, to complete the real-time adjustment of camera parameters, thereby obtaining optimized shooting images and providing basic data for subsequent image quality evaluation.

[0064] Step S500 in the method provided in this embodiment of the invention includes: Obtain the current focal length value of the drone shooting device, and add the current focal length value to the target focal length change in the shooting parameter adjustment strategy to obtain the target focal length value; Obtain the current focus position of the drone's shooting device, and adjust the focus position according to the focus movement step size and direction in the shooting parameter adjustment strategy; Based on the metering weight offset in the shooting parameter adjustment strategy, the weight distribution of each metering region in the camera's automatic exposure algorithm is adjusted, and the exposure parameters are updated.

[0065] First, the current focal length value of the drone's camera is obtained. This current focal length value is then added to the target focal length change in the shooting parameter adjustment strategy to obtain the target focal length value. The current focal length value refers to the actual optical focal length of the drone's gimbal camera before adjustment. The target focal length change refers to the focal length increase or decrease specified in the shooting parameter adjustment strategy, representing the magnitude of the focal length adjustment. The target focal length value is the target optical focal length that the camera should achieve after adjustment; it is the execution target value of the camera's zoom mechanism.

[0066] Specifically, the current real-time focal length value is read through the camera driver interface. This value is then algebraically added to the target focal length change in the adjustment strategy to calculate the target focal length value. This target focal length value is then sent to the camera zoom actuator to complete the zoom adjustment. For example, if the current focal length of the drone camera is 20mm, the target focal length change in the adjustment strategy is 20 × 1.5 - 20 = +10mm, the target focal length value is 20 + 10 = 30mm, and the camera zoom mechanism moves to the 30mm focal length position.

[0067] Next, the current focus position of the drone's shooting device is obtained, and the focus position is adjusted according to the focus movement step size and direction in the shooting parameter adjustment strategy. The current focus position refers to the initial position of the camera's focusing motor before adjustment, measured by the number of steps from the built-in encoder. The focus movement step size refers to the number of steps the focusing motor moves as specified in the shooting parameter adjustment strategy, representing the magnitude of the focus adjustment. The focus movement direction refers to the direction in which the focusing motor moves; a positive value indicates adjustment towards the telephoto direction, and a negative value indicates adjustment towards the near-focus direction.

[0068] Specifically, the system reads the current focus position encoding value from the camera's focusing module, and drives the focusing motor to perform stepping motion according to the movement step size and direction given by the adjustment strategy, thereby completing the precise adjustment of the focus position. For example, if the camera's current focus position is a baseline of 0 steps, and the focus movement parameter in the adjustment strategy is +15 steps, the focusing motor moves 15 steps towards the telephoto direction to complete the focus adjustment.

[0069] Finally, based on the metering weight offset in the shooting parameter adjustment strategy, the weight distribution of each metering region in the camera's automatic exposure algorithm is adjusted, and the exposure parameters are updated. The metering weight offset refers to the magnitude of the metering region weight adjustment specified in the shooting parameter adjustment strategy, used to change the metering priority of different image regions. The camera's automatic exposure algorithm is a built-in algorithm that calculates exposure parameters based on the brightness of each area of ​​the image. The metering region weight distribution refers to the weight ratio of different image regions on the final exposure result in the automatic exposure algorithm.

[0070] Specifically, based on the metering weight offset, the weight distribution ratio of each metering zone in the camera's automatic exposure algorithm is modified, and exposure parameters such as aperture, shutter speed, and ISO are recalculated and updated to complete the metering adjustment. For example, if the original metering area weight distribution was uniform, and the metering weight offset in the adjustment strategy is +0.3, the metering weight of the target area is increased by 0.3, and the camera automatically updates the exposure parameters to a +0.3EV exposure compensation value. After completing the zoom, focus, and metering adjustments, the drone gimbal camera is controlled to re-acquire the captured image, which will be used as the processing object for subsequent quality evaluation.

[0071] In this embodiment of the invention, the digital shooting parameter adjustment strategy is precisely converted into control actions that can be executed by the camera hardware, and the zoom, focus and metering are independently controlled in different dimensions to ensure the accuracy and real-time performance of parameter adjustment. Through standardized execution at the hardware level, the target ratio, texture clarity and illumination uniformity of the captured image are effectively improved. The re-acquired image is more in line with the imaging requirements of power distribution network component inspection, providing real and effective image data for the subsequent image quality evaluation model.

[0072] S600: Input the re-acquired footage into the pre-trained image quality evaluation model to obtain the probability of passing the quality test.

[0073] In this embodiment of the invention, the re-acquired captured images are input into a pre-trained image quality evaluation model to obtain the probability of quality compliance. Whether the re-acquired images, after parameter adjustment, meet the imaging requirements for power distribution network inspection cannot be quickly and uniformly quantitatively assessed by manual subjective judgment. This can easily lead to inconsistent evaluation standards and low judgment efficiency, and also fails to provide accurate numerical feedback for subsequent model optimization. Therefore, it is necessary to use a pre-trained image quality evaluation model to automatically assess the quality of the adjusted captured images and output a normalized probability of quality compliance, which serves as an objective basis for determining whether the image quality meets the standards.

[0074] The image quality assessment model is constructed using a convolutional neural network, taking images captured by drone inspections as input and outputting the probability of quality compliance in the range of 0 to 1. Training data was collected from historical real-world images captured by power distribution network drones, totaling 100,000 samples, all labeled as qualified by professionals. These samples were divided into an 8:2 ratio: 80,000 training sets and 20,000 validation sets. The image quality assessment model consists of three convolutional layers with 32, 64, and 128 kernels, all using 3×3 kernels and ReLU activation with max pooling. These are followed by two fully connected layers with 64 and 32 neurons, both using ReLU activation. The output layer has one neuron and uses a sigmoid activation. Gradient descent optimization is performed using a binary cross-entropy loss function during training. The convergence condition is that the image quality discrimination accuracy on the validation set is no less than 96%, and the model loss decreases by less than 0.5% over 10 consecutive iterations.

[0075] For example, the insulator string image captured by the S500 after zooming, focusing and metering adjustment is input into the image quality evaluation model. After forward inference calculation by the image quality evaluation model, the final output is that the quality of the current captured image is qualified with a probability of 0.92.

[0076] In this embodiment of the invention, the quality of the adjusted captured images is automatically and quantitatively evaluated, avoiding the subjectivity and lag of manual evaluation. The probability of quality passing can intuitively reflect the degree to which the images meet the inspection requirements. At the same time, it provides accurate quantitative feedback indicators for the adversarial incremental optimization of the subsequent shooting parameter adjustment strategy generation model, and also provides an objective standard for determining whether to terminate the shooting parameter adjustment process, effectively improving the intelligent level of quality control of power distribution network drone inspection images.

[0077] S700: Based on the quality pass probability and the comprehensive shooting quality deviation index, perform adversarial incremental optimization on the shooting parameter adjustment strategy generation model until the quality pass probability meets the preset conditions.

[0078] In this embodiment of the invention, based on the quality pass probability and the comprehensive shooting quality deviation index, the shooting parameter adjustment strategy generation model is subjected to adversarial incremental optimization until the quality pass probability meets the preset conditions. The UAV power distribution network inspection scenario involves complex situations such as sudden changes in lighting, component occlusion, and distance variations. Relying solely on the initially pre-trained shooting parameter adjustment strategy generation model is insufficient to adapt to diverse inspection scenarios in the long term, and the accuracy of the adjustment strategy will gradually decline. Simultaneously, the quality pass probability and the comprehensive shooting quality deviation index directly reflect the actual effect of the current shooting parameter adjustment. If the shooting parameter adjustment strategy generation model is not dynamically optimized in conjunction with real-time inspection data, it will be unable to continuously learn the adjustment patterns of new scenarios, and the shooting quality cannot be guaranteed to consistently meet the preset standards. Therefore, it is necessary to continuously iterate and optimize the shooting parameter adjustment strategy generation model parameters by calculating quantitative indicators related to the adjustment effect, constructing incremental samples, and performing adversarial incremental training until the quality pass probability meets the requirements for power distribution network inspection shooting.

[0079] Step S700 in the method provided in this embodiment of the invention includes: Calculate the difference between the quality pass probability and the preset target pass probability to obtain the pass probability deviation; Calculate the difference between the overall shooting quality deviation index before adjustment and the overall shooting quality deviation index after adjustment to obtain the quality improvement magnitude; Calculate the ratio of the pass probability deviation to the quality improvement magnitude, and use it as the incremental optimization intensity coefficient; The comprehensive shooting quality deviation index, component importance weight, actual adjustment type and adjustment magnitude, and quality pass probability before the adjustment in the current shooting process are taken as a set of incremental sample data and added to the incremental training sample set. When the number of samples in the incremental training sample set reaches the preset incremental threshold, the incremental optimization intensity coefficient is used as the learning rate adjustment factor, and the shooting parameter adjustment strategy generation model is subjected to alternating adversarial training using the incremental training sample set to obtain the incrementally optimized model.

[0080] First, the difference between the quality pass probability and the preset target pass probability is calculated to obtain the pass probability deviation. The pass probability deviation is a numerical value representing the difference between the actual quality pass probability and the preset ideal pass probability, used to measure the degree of difference between the current captured image quality and the target quality. The preset target pass probability is a quality pass probability threshold preset according to the power distribution network inspection imaging standard, and is the target value for achieving the shooting quality standard. The difference between the quality pass probability obtained by S600 and the preset target pass probability is the pass probability deviation. For example, if the preset target pass probability is 0.95 and the current quality pass probability is 0.92, the pass probability deviation = 0.92 - 0.95 = -0.03.

[0081] Secondly, the difference between the overall shooting quality deviation index before and after adjustment is calculated to obtain the quality improvement magnitude. The quality improvement magnitude represents the reduction in the overall shooting quality deviation index before and after the shooting parameter adjustment; the larger the value, the more significant the improvement in shooting quality. The adjusted overall shooting quality deviation index refers to the overall shooting quality deviation index recalculated after optimizing the camera parameters according to the adjustment strategy.

[0082] Specifically, the difference between the overall shooting quality deviation index before and after adjustment is the quality improvement margin. For example, if the overall shooting quality deviation index before adjustment is 0.488 and after adjustment is 0.158, the quality improvement margin is 0.488 - 0.158 = 0.33.

[0083] Next, the ratio of the pass probability deviation to the quality improvement magnitude is calculated as the incremental optimization intensity coefficient. The incremental optimization intensity coefficient is used to dynamically adjust the model's incremental training learning rate. It is obtained from the ratio of the adjustment effect feedback index and can adaptively control the training intensity based on the actual optimization effect. The incremental optimization intensity coefficient is the ratio obtained by quotienting the pass probability deviation and the quality improvement magnitude. For example, if the pass probability deviation is -0.03 and the quality improvement magnitude is 0.33, the incremental optimization intensity coefficient... .

[0084] Furthermore, the comprehensive shooting quality deviation index, component importance weights, actual adjustment type and magnitude, and quality pass probability before adjustment during the current shooting process are collected as incremental sample data and added to the incremental training sample set. The incremental sample data is training samples generated based on this real-time inspection shooting process, including the comprehensive shooting quality deviation index, component importance weights, actual execution instructions, and adjustment effect feedback, used for incremental learning of the shooting parameter adjustment strategy generation model. The incremental training sample set is a dataset storing multiple sets of real-time incremental samples, used for adversarial incremental training of the shooting parameter adjustment strategy generation model.

[0085] Specifically, the overall shooting quality deviation index, component importance weights, actual adjustment types and magnitudes, and the final quality pass probability before adjustments during this inspection are integrated into a complete incremental sample and added to the incremental training sample set. For example, an incremental sample is generated with the following parameters: overall shooting quality deviation index 0.488, component importance weight 0.9, adjustment strategy (zoom × 1.5, focus +15 steps, metering +0.3EV), and quality pass probability 0.92. This sample is then added to the incremental training sample set.

[0086] Then, when the number of samples in the incremental training sample set reaches the preset incremental threshold, the incremental optimization intensity coefficient is used as the learning rate adjustment factor, and the shooting parameter adjustment strategy generation model is subjected to alternating adversarial training using the incremental training sample set to obtain the incrementally optimized model. The preset incremental threshold is a pre-set minimum number of samples in the incremental training sample set; incremental training of the model can only be started when this number is reached. The total number of samples in the incremental training sample set is counted in real time to determine whether it is greater than or equal to the preset incremental threshold. If it is reached, alternating adversarial training is started; if it is not reached, the next round of inspection incremental samples is collected. For example, if the preset incremental threshold is 200 sets, and the current incremental training sample set has accumulated 215 sets of samples, the incremental training condition is met.

[0087] The alternating adversarial training includes: By fixing the parameters of the shooting strategy generation model, the pre-adjustment comprehensive shooting quality deviation index from the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy. The simulated shooting images corresponding to the generated adjustment strategy are used as negative samples, and the shooting images corresponding to the manually confirmed effective adjustment strategies in the sample data are used as positive samples to train the image quality evaluation model. By fixing the parameters of the image quality evaluation model, the pre-adjustment comprehensive shooting quality deviation index in the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy; The simulated shooting footage corresponding to the generation and adjustment strategy is input into the image quality evaluation model to obtain the quality pass probability, and the parameters of the shooting strategy generation model are updated with the goal of maximizing the quality pass probability. Repeated adversarial training continues until the shooting parameter adjustment strategy generated by the shooting strategy generation model makes the difference between the quality pass probability output by the image quality evaluation model and the quality pass probability of manually confirmed valid samples less than a preset difference threshold.

[0088] First, the parameters of the shooting strategy generation model are fixed, and the pre-adjustment comprehensive shooting quality deviation index from the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy. The pre-adjustment comprehensive shooting quality deviation index refers to a quantitative indicator calculated based on perceptual evaluation parameters and component importance weights from the sample data before any shooting parameter adjustments are performed; it is used to characterize the degree of deviation in initial shooting quality. The generated adjustment strategy refers to the shooting parameter adjustment instructions that the generator autonomously predicts and outputs based on the input pre-adjustment comprehensive shooting quality deviation index, including three adjustment types: zoom, focus, and metering, and their corresponding amplitudes.

[0089] Specifically, firstly, all network parameters of the shooting strategy generation model are fixed, including the weights of hidden layer neurons and activation function parameters, ensuring that they are not updated in this step. Then, a single set of pre-adjustment comprehensive shooting quality deviation indices is extracted from the sample data and input into the fixed-parameter generation model. Through forward inference, a set of generation adjustment strategies is obtained. For example, with all generator parameters fixed, the pre-adjustment comprehensive shooting quality deviation index extracted from the sample data is 0.488. This index is input into the generator, and after inference, the generation adjustment strategy is obtained: zoom ×1.5x, focus +15 steps, metering +0.3EV.

[0090] Secondly, the simulated shooting images corresponding to the generated adjustment strategy are used as negative samples, and the shooting images corresponding to the manually confirmed effective adjustment strategies in the sample data are used as positive samples to train the image quality evaluation model. Using the simulated shooting images corresponding to the generated adjustment strategies as negative samples and the actual shooting images corresponding to the manually confirmed effective adjustment strategies in the sample data as positive samples, the two types of samples are input into the image quality evaluation model in a 1:1 ratio. A binary cross-entropy loss function (BCE) is used, and the network parameters of the evaluator are optimized through backpropagation algorithm to train the evaluator to accurately distinguish between positive and negative samples. Essentially, this involves using unqualified images generated by the generator and qualified images confirmed by humans to calibrate the evaluator's discrimination criteria and improve its evaluation accuracy.

[0091] The simulated shooting image is based on the original shooting image before adjustment, and is obtained by simulating the execution of the shooting parameter adjustment strategy using image simulation transformation technology. The specific generation method is as follows: First, obtain the original shooting image before adjustment. Then, according to the adjustment type and adjustment amplitude in the shooting parameter adjustment strategy, call a preset image processing algorithm library to process the original image. If the adjustment type is zoom adjustment, the original image is scaled proportionally and cropped at the center according to the zoom amplitude to simulate changes in the field of view; if the adjustment type is focus adjustment, the image is convolved according to the focus offset step size mapped to the size of the Gaussian blur kernel to simulate a defocusing effect; if the adjustment type is metering adjustment, the luminance component V in the HSV color space of the image is linearly transformed according to the exposure compensation amplitude to simulate luminance changes. Finally, the above processing results are superimposed to generate the final simulated shooting image.

[0092] For example, assuming the original captured image before adjustment is a frame of a distribution network insulator, and the adjustment strategy output by the shooting parameter adjustment strategy generation model is zoom adjustment × 1.5x, focus adjustment + 15 steps, and metering adjustment + 0.3EV, the generation process of the simulated captured image is as follows: Zoom adjustment: Based on the zoom amplitude of 1.5x, the original image is magnified by 1.5x, and then an area of ​​the same size as the original image is cropped from the center of the magnified image to simulate the change in field of view after the drone camera zooms in; Focus adjustment: Based on the focus offset step size + 15 steps, according to the preset step size-blur kernel size mapping relationship, a Gaussian blur kernel radius of 7.5 pixels is calculated, a Gaussian kernel of the corresponding size is generated, and a convolution operation is performed on the original image to simulate the out-of-focus blur effect caused by focus shift; Metering adjustment: Based on the exposure compensation amplitude of + 0.3EV, the original image is converted to the HSV color space, and the luminance component V is multiplied by a coefficient of 2. 0.3 Then, the image is converted back to RGB space to simulate the brightness change effect after increasing exposure compensation. The three images, which have undergone zoom, focus, and metering processing, are then superimposed in sequence to generate the final image, which is the simulated shooting image corresponding to this adjustment strategy.

[0093] Subsequently, the simulated insulator string images corresponding to the generation strategy are used as negative samples, and the clear insulator string images taken after manual adjustment are used as positive samples. These are input into the evaluator for training, and the discrimination parameters of the evaluator are optimized so that it can accurately identify the simulated images as unqualified (output qualified probability ≤ 0.3) and the real images as qualified (output qualified probability ≥ 0.95).

[0094] Next, with the parameters of the image quality assessment model fixed, the pre-adjustment comprehensive shooting quality deviation index from the sample data is input into the shooting strategy generation model to obtain a generated adjustment strategy. After the evaluator training is completed, all network parameters of the image quality assessment model are fixed; the same pre-adjustment comprehensive shooting quality deviation index obtained above is extracted from the sample data again and input into the shooting strategy generation model to obtain a new set of generated adjustment strategies.

[0095] For example, with the evaluator parameters fixed after training, the overall shooting quality deviation index of 0.488 before adjustment is input into the generator again to obtain the same generation adjustment strategy as above: zoom × 1.5x, focus +15 steps, metering +0.3EV.

[0096] Then, the simulated shooting footage corresponding to the generation and adjustment strategy is input into the image quality evaluation model to obtain the quality pass probability. The parameters of the shooting strategy generation model are updated with the goal of maximizing this quality pass probability. The simulated shooting footage corresponding to the generation and adjustment strategy is input into the image quality evaluation model with fixed parameters, and the evaluator outputs the quality pass probability of the simulated footage. With maximizing this quality pass probability as the optimization objective, a joint loss function consisting of classification cross-entropy loss (CE) and mean squared error loss (MSE) is used. Through the backpropagation algorithm, the network parameters of the shooting strategy generation model are updated. Essentially, the quality pass probability output by the evaluator is used to judge the quality of the generator's generated strategy, thereby optimizing the generator's output accuracy and making its generated strategy closer to an effective human strategy.

[0097] For example, the simulated image corresponding to the generation strategy is input into an evaluator with fixed parameters, and the quality pass probability is 0.82. With the goal of maximizing this probability, the generator parameters are updated in reverse through the joint loss function (CE+MSE). After the update, the deviation index of 0.488 is input again, and the generator outputs a new adjustment strategy: zoom ×1.6x, focus +16 steps, metering +0.4EV, which increases the pass probability output by the evaluator to 0.88.

[0098] Finally, the alternating adversarial training is repeated until the shooting parameter adjustment strategy generated by the shooting strategy generation model makes the difference between the quality pass probability output by the image quality evaluation model and the quality pass probability of manually confirmed valid samples less than a preset difference threshold. The preset difference threshold is a pre-set probability error threshold used to determine whether the alternating adversarial training is complete, with a value range of 0~0.05, for example, a preset difference threshold set to 0.02. The quality pass probability of manually confirmed valid samples refers to the quality pass probability obtained after inputting the actual shot corresponding to the manually confirmed valid adjustment strategy into the image quality evaluation model, serving as a standard value for judging the generator's generation effect; for example, the pass probability of manually confirmed valid samples is ≥0.95. Convergence condition: The absolute difference between the quality pass probability corresponding to the generation adjustment strategy output by the shooting strategy generation model and the quality pass probability of manually confirmed valid samples is less than the preset difference threshold of 0.02, and this difference remains below 0.02 for 5 consecutive iterations, indicating that the shooting parameter adjustment strategy generation model training has converged.

[0099] Specifically, the above alternating training process is repeated cyclically. Each round of training follows the logic of fixing the generator, training the evaluator, fixing the evaluator, and updating the generator, continuously optimizing the parameters of the two models. After each round of training, the absolute difference between the quality pass probability corresponding to the generation strategy and the pass probability of the human effective sample is calculated to determine whether the convergence condition is met. If not, the next round of alternating training continues. If it is met, the alternating adversarial training stops. At this point, both the generator and the evaluator have reached the training convergence state and can be used for subsequent shooting parameter adjustment and quality evaluation.

[0100] For example, the preset difference threshold is 0.02, and the probability of a manually confirmed valid sample being of acceptable quality is 0.95. After 80 rounds of alternating training, the probability of the generator output adjustment strategy being of acceptable quality is 0.94, and the difference between it and the probability of a manually confirmed valid sample is 0.01. This difference remains at 0.01 for 5 consecutive rounds, which is less than 0.02, thus satisfying the convergence condition and stopping the alternating adversarial training.

[0101] In this embodiment of the invention, by combining the adjustment effect data of real-time inspection, the optimization strength of the model is adaptively determined by calculating quantitative indicators, and incremental samples are constructed to complete adversarial incremental training, realizing the online dynamic iteration of the shooting parameter adjustment strategy generation model. This effectively improves the adaptability of the shooting parameter adjustment strategy generation model to complex power distribution network inspection scenarios, allowing the adjustment strategy to be continuously optimized as the scenario changes. At the same time, the cyclic judgment mechanism ensures that the probability of quality qualification eventually meets the preset conditions, ensuring the stability and accuracy of the drone inspection shooting quality and extending the effective use period of the shooting parameter adjustment strategy generation model.

[0102] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides an intelligent drone shooting control method and system for power distribution network inspection. By introducing target space occupancy ratio, texture clarity attenuation, and illumination zone contrast as perception evaluation parameters, and combining them with component importance weights to calculate a comprehensive shooting quality deviation index, it can accurately quantify the gap between the current shooting quality and the ideal state. Utilizing a pre-trained shooting parameter adjustment strategy generation model, it automatically outputs multiple adjustment strategies such as zoom, focus, and metering, achieving adaptive control of drone shooting parameters. Simultaneously, an image quality evaluation model provides feedback on the probability of quality compliance, and an adversarial incremental optimization mechanism continuously iterates the model until the probability of compliance meets preset conditions. This invention effectively solves the problems of neglecting component importance and multi-dimensional perception quality, the inability to adaptively adjust shooting parameters, and the lack of online optimization capabilities in existing technologies, thus improving the shooting quality and detection reliability of power distribution network inspection images.

[0103] Example 2, as Figure 2 As shown, this invention provides an intelligent drone shooting control system for power distribution network inspection, the system comprising: The perception parameter acquisition module 11 is used to acquire the perception evaluation parameters of the current shooting scene, wherein the perception evaluation parameters include the target space occupancy ratio, texture clarity attenuation degree and illumination zone contrast value. The component weight acquisition module 12 is used to acquire the importance weight of the components corresponding to the current shooting target; The deviation index calculation module 13 is used to analyze and obtain the comprehensive shooting quality deviation index based on the perception evaluation parameters and the component importance weights. The adjustment strategy generation module 14 is used to input the comprehensive shooting quality deviation index and the component importance weight into the pre-trained shooting parameter adjustment strategy generation model to obtain the shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment and metering adjustment. The camera adjustment framing module 15 is used to control the UAV gimbal camera to perform adjustment actions according to the shooting parameter adjustment strategy, and to reacquire the shooting image; The quality probability assessment module 16 is used to input the re-acquired shooting images into the pre-trained image quality assessment model to obtain the quality pass probability; The adversarial incremental optimization module 17 is used to perform adversarial incremental optimization on the shooting parameter adjustment strategy generation model based on the quality pass probability and the comprehensive shooting quality deviation index, until the quality pass probability meets the preset conditions.

[0104] In one embodiment, the sensing parameter acquisition module 11 is further configured to: Obtain the pixel area value of the target detection box in the current shooting frame as the detection box area, and obtain the total pixel value of the current shooting frame as the total area of ​​the frame; The target space occupancy ratio is obtained by dividing the area of ​​the detection frame by the total area of ​​the screen. Edge detection is performed on the image region within the target detection box to obtain the average edge intensity as the current edge intensity; the reference edge intensity is obtained according to the preset target type-distance-reference edge intensity three-dimensional mapping table, and the texture sharpness attenuation is calculated, where texture sharpness attenuation = 1 - (current edge intensity / reference edge intensity); The current captured image is divided into a preset number of grid regions, and the average brightness value of each grid region is obtained; the standard deviation of the average brightness values ​​of all grid regions is calculated and recorded as the brightness standard deviation, and the arithmetic mean of the average brightness values ​​of all grid regions is calculated as the average brightness of the entire image. The brightness standard deviation is calculated and divided by the average brightness of the entire screen to obtain the contrast value of the illumination zone.

[0105] In one embodiment, the component weight acquisition module 12 is further configured to: The component type corresponding to the current shooting point is obtained based on the current coordinates of the drone, wherein the component type includes insulator string, wire clamp, pin, bolt and crossarm; Based on the component type, the corresponding importance weight is retrieved from a pre-established component type-importance weight mapping table and used as the component importance weight.

[0106] In one embodiment, the deviation index calculation module 13 is further configured to: Multiply the texture clarity attenuation by the first weighting coefficient to obtain the first deviation component; Calculate the reciprocal of the target space occupancy ratio, and divide the reciprocal of the target space occupancy ratio by the upper limit of the reciprocal of the occupancy ratio to obtain the normalized value of the occupancy ratio; Multiply the occupancy ratio normalization value by the second weighting coefficient to obtain the second deviation component; The maximum value of the contrast value of the illumination zone is obtained from the historical inspection data in advance, and is used as the upper limit of the contrast value. The contrast value of the illumination zone is divided by the upper limit of the contrast value to obtain the contrast normalization value. The third deviation component is obtained by multiplying the contrast normalization value by the component importance weight, and then multiplying it by the third weighting coefficient. The first deviation component, the second deviation component, and the third deviation component are added together to obtain the comprehensive shooting quality deviation index, wherein the sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is 1, and each weight coefficient is obtained according to the frequency ratio of shooting failure caused by the corresponding parameter in the historical inspection.

[0107] In one embodiment, the adjustment strategy generation module 14 is further configured to: The comprehensive shooting quality deviation index of the current shooting image and the component importance weight are input into the pre-trained shooting parameter adjustment strategy to generate the model; The adjustment type and adjustment amplitude prediction values ​​output by the shooting parameter adjustment strategy generation model are obtained as the shooting parameter adjustment strategy, wherein the shooting parameter adjustment strategy includes adjustment type and adjustment amplitude, and the adjustment type includes zoom adjustment, focus adjustment and metering adjustment.

[0108] The pre-training of the model generated by the shooting parameter adjustment strategy includes: Collect multiple sets of shooting adjustment records from historical inspections. Each set of shooting adjustment records includes the overall shooting quality deviation index of the image before adjustment, the corresponding component importance weight, and the adjustment type and adjustment amplitude that have been manually confirmed as valid, as sample data. A shooting strategy generation model is constructed, which takes the comprehensive shooting quality deviation index and component importance weight as input and the predicted adjustment type and adjustment magnitude as output. An image quality evaluation model is constructed, which takes the adjusted captured image as input and the predicted quality pass probability as output. The shooting strategy generation model and the image quality evaluation model are trained alternately using the sample data. When the shooting parameter adjustment strategy generated by the shooting strategy generation model enables the quality pass probability output by the image quality evaluation model to reach a preset convergence threshold, the pre-trained shooting strategy generation model and image quality evaluation model are obtained, and the shooting parameter adjustment strategy generation model is integrated.

[0109] In one embodiment, the camera framing module 15 is further configured to: Obtain the current focal length value of the drone shooting device, and add the current focal length value to the target focal length change in the shooting parameter adjustment strategy to obtain the target focal length value; Obtain the current focus position of the drone's shooting device, and adjust the focus position according to the focus movement step size and direction in the shooting parameter adjustment strategy; Based on the metering weight offset in the shooting parameter adjustment strategy, the weight distribution of each metering region in the camera's automatic exposure algorithm is adjusted, and the exposure parameters are updated.

[0110] In one embodiment, the counter-incremental optimization module 17 is further configured to: Calculate the difference between the quality pass probability and the preset target pass probability to obtain the pass probability deviation; Calculate the difference between the overall shooting quality deviation index before adjustment and the overall shooting quality deviation index after adjustment to obtain the quality improvement magnitude; Calculate the ratio of the pass probability deviation to the quality improvement magnitude, and use it as the incremental optimization intensity coefficient; The comprehensive shooting quality deviation index, component importance weight, actual adjustment type and adjustment magnitude, and quality pass probability before the adjustment in the current shooting process are taken as a set of incremental sample data and added to the incremental training sample set. When the number of samples in the incremental training sample set reaches the preset incremental threshold, the incremental optimization intensity coefficient is used as the learning rate adjustment factor, and the shooting parameter adjustment strategy generation model is subjected to alternating adversarial training using the incremental training sample set to obtain the incrementally optimized model.

[0111] The alternating adversarial training includes: By fixing the parameters of the shooting strategy generation model, the pre-adjustment comprehensive shooting quality deviation index from the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy. The simulated shooting images corresponding to the generated adjustment strategy are used as negative samples, and the shooting images corresponding to the manually confirmed effective adjustment strategies in the sample data are used as positive samples to train the image quality evaluation model. By fixing the parameters of the image quality evaluation model, the pre-adjustment comprehensive shooting quality deviation index in the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy; The simulated shooting footage corresponding to the generation and adjustment strategy is input into the image quality evaluation model to obtain the quality pass probability, and the parameters of the shooting strategy generation model are updated with the goal of maximizing the quality pass probability. Repeated adversarial training continues until the shooting parameter adjustment strategy generated by the shooting strategy generation model makes the difference between the quality pass probability output by the image quality evaluation model and the quality pass probability of manually confirmed valid samples less than a preset difference threshold.

[0112] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent shooting and control of unmanned aerial vehicles (UAVs) for power distribution network inspection, characterized in that, include: Obtain the perception evaluation parameters of the currently captured image, wherein the perception evaluation parameters include the target space occupancy ratio, texture clarity attenuation degree, and illumination zone contrast value; Obtain the importance weights of the components corresponding to the current shooting target; Based on the perceived evaluation parameters and the importance weights of the components, an overall shooting quality deviation index is obtained through analysis, including: Multiply the texture clarity attenuation by the first weighting coefficient to obtain the first deviation component; Calculate the reciprocal of the target space occupancy ratio, and divide the reciprocal of the target space occupancy ratio by the upper limit of the reciprocal of the occupancy ratio to obtain the normalized value of the occupancy ratio; Multiply the occupancy ratio normalization value by the second weighting coefficient to obtain the second deviation component; The maximum value of the contrast value of the illumination zone is obtained from the historical inspection data in advance, and is used as the upper limit of the contrast value. The contrast value of the illumination zone is divided by the upper limit of the contrast value to obtain the contrast normalization value. The third deviation component is obtained by multiplying the contrast normalization value by the component importance weight, and then multiplying it by the third weighting coefficient. The first deviation component, the second deviation component, and the third deviation component are added together to obtain the comprehensive shooting quality deviation index, wherein the sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is 1, and each weight coefficient is obtained according to the frequency ratio of shooting failure caused by the corresponding parameter in the historical inspection. The comprehensive shooting quality deviation index and the component importance weights are input into a pre-trained shooting parameter adjustment strategy generation model to obtain a shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment and metering adjustment. The drone gimbal camera is controlled to perform adjustment actions according to the shooting parameter adjustment strategy, and the shooting image is re-acquired; The re-acquired footage is input into a pre-trained image quality assessment model to obtain the probability of passing quality assessment. Based on the quality pass probability and the comprehensive shooting quality deviation index, the shooting parameter adjustment strategy generation model is subjected to adversarial incremental optimization until the quality pass probability meets the preset conditions.

2. The intelligent shooting and control method for UAVs for power distribution network inspection according to claim 1, characterized in that, The acquisition of perception evaluation parameters of the currently captured image includes target space occupancy ratio, texture sharpness attenuation, and illumination zone contrast value, among others: Obtain the pixel area value of the target detection box in the current shooting frame as the detection box area, and obtain the total pixel value of the current shooting frame as the total area of ​​the frame; The target space occupancy ratio is obtained by dividing the area of ​​the detection frame by the total area of ​​the screen. Edge detection is performed on the image region within the target detection box to obtain the average edge intensity as the current edge intensity; the reference edge intensity is obtained according to the preset target type-distance-reference edge intensity three-dimensional mapping table, and the texture sharpness attenuation is calculated, where texture sharpness attenuation = 1 - (current edge intensity / reference edge intensity); The current captured image is divided into a preset number of grid regions, and the average brightness value of each grid region is obtained; the standard deviation of the average brightness values ​​of all grid regions is calculated and recorded as the brightness standard deviation, and the arithmetic mean of the average brightness values ​​of all grid regions is calculated as the average brightness of the entire image. The brightness standard deviation is calculated and divided by the average brightness of the entire screen to obtain the contrast value of the illumination zone.

3. The intelligent shooting and control method for UAVs for power distribution network inspection according to claim 1, characterized in that, The step of obtaining the importance weights of the components corresponding to the current shooting target includes: The component type corresponding to the current shooting point is obtained based on the current coordinates of the drone, wherein the component type includes insulator string, wire clamp, pin, bolt and crossarm; Based on the component type, the corresponding importance weight is retrieved from a pre-established component type-importance weight mapping table and used as the component importance weight.

4. The intelligent shooting and control method for UAVs for power distribution network inspection according to claim 1, characterized in that, The process involves inputting the comprehensive shooting quality deviation index and the component importance weights into a pre-trained shooting parameter adjustment strategy generation model to obtain a shooting parameter adjustment strategy. This shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment, and metering adjustment, including: The comprehensive shooting quality deviation index of the current shooting image and the component importance weight are input into the pre-trained shooting parameter adjustment strategy to generate the model; The adjustment type and adjustment amplitude prediction values ​​output by the shooting parameter adjustment strategy generation model are obtained as the shooting parameter adjustment strategy, wherein the shooting parameter adjustment strategy includes adjustment type and adjustment amplitude, and the adjustment type includes zoom adjustment, focus adjustment and metering adjustment.

5. The intelligent shooting control method for UAVs for power distribution network inspection according to claim 4, characterized in that, The pre-training of the model generated by the shooting parameter adjustment strategy includes: Collect multiple sets of shooting adjustment records from historical inspections. Each set of shooting adjustment records includes the overall shooting quality deviation index of the image before adjustment, the corresponding component importance weight, and the adjustment type and adjustment amplitude that have been manually confirmed as valid, as sample data. A shooting strategy generation model is constructed, which takes the comprehensive shooting quality deviation index and component importance weight as input and the predicted adjustment type and adjustment magnitude as output. An image quality evaluation model is constructed, which takes the adjusted captured image as input and the predicted quality pass probability as output. The shooting strategy generation model and the image quality evaluation model are trained alternately using the sample data. When the shooting parameter adjustment strategy generated by the shooting strategy generation model enables the quality pass probability output by the image quality evaluation model to reach a preset convergence threshold, the pre-trained shooting strategy generation model and image quality evaluation model are obtained, and the shooting parameter adjustment strategy generation model is integrated.

6. The intelligent shooting control method for UAVs for power distribution network inspection according to claim 4, characterized in that, The step of controlling the drone gimbal camera to perform adjustment actions according to the shooting parameter adjustment strategy and re-acquiring the shooting image includes: Obtain the current focal length value of the drone shooting device, and add the current focal length value to the target focal length change in the shooting parameter adjustment strategy to obtain the target focal length value; Obtain the current focus position of the drone's shooting device, and adjust the focus position according to the focus movement step size and direction in the shooting parameter adjustment strategy; Based on the metering weight offset in the shooting parameter adjustment strategy, the weight distribution of each metering region in the camera's automatic exposure algorithm is adjusted, and the exposure parameters are updated.

7. The intelligent shooting control method for UAVs for power distribution network inspection according to claim 1, characterized in that, The adversarial incremental optimization of the shooting parameter adjustment strategy generation model based on the quality pass probability and the comprehensive shooting quality deviation index, until the quality pass probability meets a preset condition, includes: Calculate the difference between the quality pass probability and the preset target pass probability to obtain the pass probability deviation; Calculate the difference between the overall shooting quality deviation index before adjustment and the overall shooting quality deviation index after adjustment to obtain the quality improvement magnitude; Calculate the ratio of the pass probability deviation to the quality improvement magnitude, and use it as the incremental optimization intensity coefficient; The comprehensive shooting quality deviation index, component importance weight, actual adjustment type and adjustment magnitude, and quality pass probability before the adjustment in the current shooting process are taken as a set of incremental sample data and added to the incremental training sample set. When the number of samples in the incremental training sample set reaches the preset incremental threshold, the incremental optimization intensity coefficient is used as the learning rate adjustment factor, and the shooting parameter adjustment strategy generation model is subjected to alternating adversarial training using the incremental training sample set to obtain the incrementally optimized model.

8. The intelligent shooting control method for UAVs for power distribution network inspection according to claim 5, characterized in that, The alternating adversarial training includes: By fixing the parameters of the shooting strategy generation model, the pre-adjustment comprehensive shooting quality deviation index from the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy. The simulated shooting images corresponding to the generated adjustment strategy are used as negative samples, and the shooting images corresponding to the manually confirmed effective adjustment strategies in the sample data are used as positive samples to train the image quality evaluation model. By fixing the parameters of the image quality evaluation model, the pre-adjustment comprehensive shooting quality deviation index in the sample data is input into the shooting strategy generation model to obtain the generated adjustment strategy; The simulated shooting footage corresponding to the generation and adjustment strategy is input into the image quality evaluation model to obtain the quality pass probability, and the parameters of the shooting strategy generation model are updated with the goal of maximizing the quality pass probability. Repeated adversarial training continues until the shooting parameter adjustment strategy generated by the shooting strategy generation model makes the difference between the quality pass probability output by the image quality evaluation model and the quality pass probability of manually confirmed valid samples less than a preset difference threshold.

9. A drone intelligent shooting control system for power distribution network inspection, characterized in that, The intelligent drone shooting control method for implementing any one of claims 1-8 for power distribution network inspection includes: The perception parameter acquisition module is used to acquire the perception evaluation parameters of the current shooting image, wherein the perception evaluation parameters include the target space occupancy ratio, texture clarity attenuation degree, and illumination zone contrast value. The component weighting acquisition module is used to acquire the importance weights of the components corresponding to the current shooting target; The deviation index calculation module is used to analyze and obtain the comprehensive shooting quality deviation index based on the perception evaluation parameters and the importance weights of the components. The adjustment strategy generation module is used to input the comprehensive shooting quality deviation index and the component importance weight into the pre-trained shooting parameter adjustment strategy generation model to obtain the shooting parameter adjustment strategy. The shooting parameter adjustment strategy includes adjustment type and adjustment amplitude. The adjustment type includes zoom adjustment, focus adjustment and metering adjustment. The camera adjustment framing module is used to control the UAV gimbal camera to perform adjustment actions according to the shooting parameter adjustment strategy, and to reacquire the shooting image; The quality probability assessment module is used to input the re-acquired captured images into the pre-trained image quality assessment model to obtain the probability of quality passing. The adversarial incremental optimization module is used to perform adversarial incremental optimization on the shooting parameter adjustment strategy generation model based on the quality pass probability and the comprehensive shooting quality deviation index, until the quality pass probability meets the preset conditions.

Images