Unmanned aerial vehicle multi-scene inspection model construction method fusing transfer learning and physical state guidance

Abstract

The invention relates to the technical field of artificial intelligence, unmanned aerial vehicle remote sensing and cross-domain adaptive learning, and discloses an unmanned aerial vehicle multi-scene inspection model construction method fusing transfer learning and physical state guidance. The method is used for solving the problems that the distribution difference of inspection data of an unmanned aerial vehicle is large and the generalization ability of a model is weak under the conditions of zooming, variable attitudes and complex environment illumination. The method comprises the following steps: analyzing the inspection data to construct an acquisition state fingerprint vector containing geometric and environment components; searching a source domain anchor point set by utilizing a geometric manifold index database, and resolving a geometric and environment difference vector; performing feature map space transformation and weight nonlinear fusion by using difference vectors through the mapping network and the super network, simulating a physical view angle of a target domain and adapting to environment features; the physical equivalent geometric parameters are inversely calculated in combination with the ground sampling distance, and the consistency loss value of the standard specification data is calculated to drive parameter updating.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, UAV remote sensing and cross-domain adaptive learning technology, specifically a method for constructing a multi-scenario inspection model for UAVs that integrates transfer learning and physical state guidance. Background Technology

[0002] With the widespread application of drone technology in power, transportation, and security inspections, deep learning-based image recognition technology has become a core means of processing massive amounts of inspection data. In practical applications, drones often need to perform tasks under different routes, seasons, and lighting conditions, resulting in variations in the distribution of the collected image data. To maintain model performance in new scenarios without requiring extensive labeling of new data, transfer learning and domain adaptation techniques are widely adopted, aiming to transfer existing source domain knowledge to the target domain.

[0003] However, existing transfer learning methods have certain limitations when processing UAV inspection data. Most current domain adaptation algorithms are based on the statistical distribution alignment of image features, focusing on reducing the distance between the source and target domains within the feature space, but often neglecting the underlying physical acquisition conditions of the images. The differences in UAV inspection images mainly stem from two dimensions: first, geometric perspective differences caused by variations in flight attitude and shooting distance; and second, environmental texture differences caused by changes in weather and lighting. Existing technologies typically mix these two types of differences together for overall feature alignment, failing to decouple them specifically. For example, when the source domain is primarily a top-down view while the target domain is a level view, forced feature alignment often leads to the loss or distortion of spatial structural information in the image, causing negative transfer and making it difficult to obtain a high-quality detection model.

[0004] Furthermore, traditional visual inspection models primarily rely on pixel-level texture and shape features for inference, lacking the ability to perceive physical scale. In scenarios involving drone zoom or significant changes in flight altitude, the pixel size of the same type of component in the image can vary drastically, while background objects (such as debris on the ground) may exhibit similar texture and pixel size to electrical components at a specific focal length. Because existing models do not incorporate physical parameters such as ground sampling distance for constraint, it is difficult to distinguish the true size of objects in the physical world based solely on visual features, easily leading to logically illogical false detections and reducing the practicality and reliability of the inspection system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for constructing a multi-scenario inspection model for UAVs that integrates transfer learning and physical state guidance. This method solves the problems of large differences in the distribution of inspection data and weak model generalization ability under conditions of zoom, attitude change, and complex lighting.

[0006] To achieve the above objectives, this invention employs the following technical solution: First, the raw inspection data is parsed and vectorized to construct a standardized acquisition state fingerprint vector. The raw inspection data encompasses both source and target domain data. The acquisition state fingerprint vector is explicitly decoupled in dimension into geometric state components and environmental state components, respectively representing the spatial positional relationship and optical environment characteristics during imaging.

[0007] Secondly, the geometric state components of the source domain data are received to construct a geometric manifold index. Simultaneously, a baseline geometric state vector for the target domain data is calculated and used as a query vector to retrieve the source domain anchor set from the geometric manifold index. By calculating the state difference vector between the target domain vector and the source domain anchor set vector, this difference is further decomposed into a geometric difference vector and an environmental difference vector, thereby quantifying the specific deviations between the target domain and the source domain in physical space and environmental conditions.

[0008] Next, the decoupled difference vectors are used to drive the spatial alignment and style adaptation of the feature maps. On one hand, the geometric difference vector is received and input into the mapping network to generate an affine transformation parameter matrix. This matrix is ​​then used to perform grid sampling and interpolation operations on the feature maps extracted from the source domain image, so that the feature maps simulate the physical perspective of the target domain in spatial distribution. On the other hand, the environmental difference vector is received and input into the hypernetwork structure to generate a gating vector. This gating vector is then used to perform nonlinear weighted fusion of the weight parameters of the pre-trained model, dynamically reorganizing the model weights to adapt to the lighting and texture features of the target domain.

[0009] Finally, physical prior constraints are introduced during the backpropagation phase of model training. The target detection bounding box coordinates output by the model are received, and the physical equivalent geometric parameters of the predicted targets are calculated. By calculating the consistency loss value between the physical equivalent geometric parameters and pre-stored standard specification data, parameter updates are driven, ultimately resulting in a physically consistent inspection data analysis model.

[0010] In a preferred embodiment, the process of constructing a standardized acquisition state fingerprint vector includes: extracting telemetry data and sensor metadata associated with each frame of the original inspection data. When constructing the geometric state component, relative height, shooting distance, gimbal pitch angle, and relative yaw angle are selected to characterize the spatial geometric relationship during imaging; when constructing the environmental state component, ground sampling distance, exposure value, and image texture entropy are selected to characterize the ambient lighting and texture sharpness conditions during imaging. Subsequently, using the statistical distribution parameters of historical source domain data, the geometric state component and the environmental state component are Z-score standardized to eliminate dimensional differences.

[0011] In a preferred embodiment, a spatial partitioning algorithm is used to construct the geometric manifold index. Specifically, a K-dimensional tree algorithm is used as the spatial partitioning structure. During construction, the variance of the geometric state components of the source domain data in each dimension is calculated. The dimension with the largest variance is selected as the splitting dimension of the current node, and the median is selected as the splitting node in this splitting dimension. The left and right subtree spaces are recursively partitioned until the number of samples contained in the leaf nodes is less than a preset capacity threshold. During retrieval, weighted Euclidean distance is used as a metric to measure the similarity of geometric state vectors, and the weight of the shooting distance dimension is set greater than the weight of the angle dimension to prioritize scale consistency.

[0012] In a preferred embodiment, to obtain a stable target domain representation, when calculating the baseline geometric state vector of the target domain data, a target domain calibration dataset containing several pre-collected typical samples is acquired. The samples in this dataset are standardized using the global mean and global standard deviation of the source domain data, and the arithmetic mean of all processed sample vectors is calculated to obtain the baseline geometric state vector.

[0013] In a preferred embodiment, to accurately extract migration features, when decomposing the state difference vector into a geometric difference vector and an environmental difference vector, the initial difference vector between each anchor sample in the anchor set and the baseline geometric state vector is calculated. Using a Softmax function with a temperature coefficient, an aggregation weight is calculated based on the geometrical distance between each anchor sample and the baseline geometric state vector. This aggregation weight is then used to weight and sum the initial difference vectors of all anchor samples, thereby obtaining the aggregated geometric difference vector and environmental difference vector.

[0014] In a preferred embodiment, the specific implementation of performing grid sampling and interpolation operations on the feature map extracted from the source domain image is as follows: a normalized sampling grid is generated using an affine transformation parameter matrix; for non-integer coordinates in the sampling grid, a bilinear interpolation algorithm is used to sample the pixel values ​​at the corresponding positions in the source domain feature map to generate a geometrically corrected feature map.

[0015] In a preferred embodiment, to prevent excessive geometric deformation from damaging the feature structure, a confidence gating step is introduced before performing mesh sampling and interpolation operations. The Euclidean modulus of the geometric difference vector is calculated, and it is determined whether it exceeds a preset safe transformation threshold. If it exceeds the threshold, the affine transformation parameter matrix is ​​reset to the identity transformation matrix, or the mesh sampling operation is stopped directly, and the original source domain feature map is output.

[0016] In a preferred embodiment, the process of generating a gated vector from the input hypernetwork structure includes: establishing a mapping relationship between environmental difference features and convolutional channel importance using the hypernetwork, and outputting a feature channel importance descriptor; processing the descriptor using an activation function to generate a gated vector mapped to a normalized numerical range. The nonlinear weighted fusion specifically manifests as using this gated vector to perform a channel multiplication-based recalibration operation on the convolutional kernel weights of the convolutional layer, thereby achieving parameter-level domain adaptation.

[0017] In a preferred embodiment, the parameter inverse calculation step in the physical prior constraints includes: receiving the target detection bounding box output by the model and obtaining the pixel dimension value of the bounding box; multiplying the pixel dimension value by the ground sampling distance value recorded in the environment state component to obtain the physical equivalent geometric parameters of the predicted target. Here, the ground sampling distance value represents the actual physical distance represented by a single pixel in the image.

[0018] In a preferred embodiment, when calculating the consistency loss value between the physical equivalent geometric parameters and pre-stored standard specification data, the statistical mean and standard deviation of the physical equivalent geometric parameters of the predicted target's category are retrieved from the pre-stored standard component specification data. A Gaussian kernel function is used to calculate the physical consistency probability that the predicted target's physical equivalent geometric parameters belong to that category. A loss function is constructed based on this physical consistency probability, or this probability is used as a weighting factor to calibrate the confidence level of the predicted target, thereby suppressing spurious targets with unreasonable physical equivalent geometric parameters.

[0019] This invention provides a method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance. It has the following beneficial effects: 1. This invention solves the problem of blind source domain sample selection in traditional transfer learning by constructing a collection state fingerprint vector containing geometric and environmental components and retrieving the set of source domain anchor points with the closest spatial distance from a geometric manifold index library. This physical state-based optimization mechanism can accurately select samples from massive historical data that are closest to the current target domain in terms of shooting perspective and distance as transfer benchmarks, reducing the distribution difference between the source and target domains and thus improving the model's initial adaptation ability to new scenes.

[0020] 2. This invention decouples geometric difference vectors and environmental difference vectors, driving the mapping network for spatial alignment of feature maps and the supernetwork for channel-level weight adjustment, respectively. It utilizes affine transformations to simulate the physical perspective of the target domain, while simultaneously using gating vectors to adjust model parameters to adapt to changes in lighting and texture. This strategy of independently handling spatial deformation and environmental style avoids aliasing interference of multidimensional features during the transfer process, ensuring the model's feature extraction accuracy under complex UAV flight attitudes and variable weather conditions. 3. This invention introduces physical prior constraints and combines ground sampling distance parameters to back-calculate the physical equivalent geometric parameters of the predicted target, and calculates the consistency loss between the predicted target and pre-stored standard specification data. This mechanism adds logical verification of physical scale to conventional visual feature detection, which can effectively identify and eliminate false targets with similar texture features but seriously inconsistent physical equivalent geometric parameters (such as misdetecting ground background as power components), reducing the false alarm rate in UAV zoom inspection scenarios and improving the reliability of detection results. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of a multi-scenario UAV inspection data analysis model construction system that integrates geometric anchor point optimization and physical scale constraints, according to an embodiment of the present invention. Figure 2 This is a flowchart of a method for constructing a multi-scenario UAV inspection data analysis model that integrates geometric anchor point optimization and physical scale constraints according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of geometric manifold space and anchor point retrieval according to an embodiment of the present invention; Figure 4 This is a schematic diagram of feature mesh sampling driven by geometric differences according to an embodiment of the present invention; Figure 5 This is a schematic diagram of environment-difference-driven channel attention recalibration according to an embodiment of the present invention; Figure 6 This is a schematic diagram of physical scale back calculation and consistency verification according to an embodiment of the present invention.

[0023] Among them, 100 is the UAV data acquisition platform; 110 is the imaging device; 120 is the positioning module; 130 is the flight attitude monitoring module; 200 is the data analysis and processing terminal; 210 is the acquisition state space construction module; 220 is the source domain index establishment module; 230 is the dynamic anchor point optimization module; 240 is the feature manifold remapping module; 250 is the environment adaptive weighting module; and 260 is the physical scale constraint module. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See attached document Figure 1This embodiment of the system includes a UAV data acquisition platform 100 and a data analysis and processing terminal 200. The UAV data acquisition platform 100 is equipped with an imaging device 110, a positioning module 120, and a flight attitude monitoring module 130. The imaging device 110 is used to acquire high-resolution image data containing the target object and record the optical parameters synchronized with it. The optical parameters include at least focal length, aperture value, and exposure time. The positioning module 120 obtains the three-dimensional geographic coordinates of the sensor center through real-time dynamic differential technology. The flight attitude monitoring module 130 records the gimbal pitch angle, yaw angle, and fuselage attitude data at the moment of imaging. The image data, optical parameters, coordinate data, and attitude data together constitute the original inspection dataset. The data analysis and processing terminal 200 is logically divided into a state space construction module 210, a source domain index establishment module 220, a dynamic anchor point optimization module 230, a feature manifold remapping module 240, an environment adaptive weighting module 250, and a physical scale constraint module 260.

[0026] See attached document Figure 2 This invention provides a method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance, comprising the following steps: S10, the raw inspection data is parsed and vectorized by the acquisition state space construction module 210, and the telemetry data and sensor metadata associated with each frame of image are extracted to construct a standardized acquisition state fingerprint vector; the acquisition state fingerprint vector is decoupled into geometric state components and environmental state components in terms of dimension; the geometric state components include relative height, shooting distance, pitch angle and relative yaw angle, which are used to characterize the spatial geometric relationship during imaging; the environmental state components include ground sampling distance, exposure value and image texture entropy, which are used to characterize the ambient lighting and texture sharpness conditions during imaging; S20, the source domain index building module 220 receives the geometric state components of the source domain data and uses a spatial partitioning algorithm to build a geometric manifold index library; the geometric manifold index library supports fast nearest neighbor retrieval based on Euclidean distance or Mahalanobis distance to achieve fast localization of similar physical perspectives; S30, the reference geometric state vector of the target domain data is calculated by the dynamic anchor point optimization module 230, and it is used as a query vector to be input into the geometric manifold index library. The source domain samples that are closest in geometric space are retrieved as the anchor point set. The state difference vector between the target domain vector and the anchor point set vector is calculated, and the state difference vector is further decomposed into geometric difference vector and environmental difference vector. S40: The geometric difference vector is received through the feature manifold remapping module 240, and an affine transformation parameter matrix is ​​generated through the mapping network. Based on the matrix, grid sampling and interpolation operations are performed on the feature map extracted from the source domain image, so that the feature map of the source domain simulates the physical perspective of the target domain in spatial distribution. When the magnitude of the geometric difference vector exceeds a preset threshold, the geometric transformation operation is automatically stopped to maintain the topological integrity of the feature map. S50 receives the environmental difference vector through the environment adaptive weight module 250 and generates a gated vector through the hypernetwork structure; the gated vector is used to perform nonlinear weighted fusion on the weight parameters of the pre-trained model to generate special weight parameters that are adapted to the lighting and texture features of the target domain. S60 introduces physical prior constraints during the backpropagation stage of model training through the physical scale constraint module 260; receives the target detection bounding box coordinates output by the model, combines them with the corresponding ground sampling distance parameters, and back-calculates the physical equivalent geometric parameters of the predicted target; retrieves pre-stored standard component specification data, calculates the physical consistency loss value between the predicted size and the actual size; combines this loss value with the task classification loss value to form the total loss function, driving the parameter update of the feature manifold remapping module 240 and the environment adaptive weight module 250.

[0027] In step S10, the raw inspection data is parsed and vectorized by the state space construction module 210. The core logic of this step is to transform unstructured image data into a quantifiable physical state description, thereby extracting the acquisition condition features beyond the image content. By decoupling the state into rigid geometric components and flexible environmental components, independent mathematical modeling can be performed for spatial perspective transformation and image style transfer, respectively. Specifically, this includes the following sub-steps: S101, Data Source Parsing and Spatiotemporal Alignment. The acquisition state space construction module 210 receives the raw data packets transmitted back by the UAV, which contain image files and their associated auxiliary data files. The module first reads the Exif metadata or XMP extended data from the image files to extract the absolute timestamp of the imaging time; simultaneously, it parses the flight control log file and RTK positioning data stream, using the timestamp index to achieve millisecond-level synchronization matching between image frames and telemetry data. For the image metadata, specifically, the focal length is extracted. (Unit: mm), Aperture value Exposure time (Unit: seconds) and ISO sensitivity; for telemetry data, specifically extract the UAV's three-dimensional coordinates in the world coordinate system, the Euler angle of the gimbal relative to the fuselage, and the absolute heading angle of the fuselage relative to true north (already corrected in conjunction with local magnetic declination data).

[0028] S102, Construct the geometric state components. The geometric state components are used to quantitatively describe the relative spatial positional relationship between the UAV and the target at the moment of imaging. This component is defined as a vector. Among them, relative height The elevation is obtained by calculating the difference between the drone's RTK elevation and the target tower base elevation. The target tower base elevation is obtained from a pre-established 3D power line database by matching the tower number index of the currently captured image with the nearest neighbor distance. (Gimbal pitch angle) Pitch axis data is directly read from the gimbal's IMU sensor to characterize the vertical tilt angle of the line of sight. Relative yaw angle. This represents the angle between the main optical axis of the image and the direction of the transmission line. The calculation process is as follows: First, retrieve the coordinates of the current tower and the next adjacent tower from the database. Then, use the arctangent function atan2 to calculate the geographical azimuth angle of the line direction. Subsequently, the absolute heading angle of the drone gimbal was combined with the data. Calculate the difference. To eliminate the influence of angular periodicity, the following formula for calculating the minimum angular difference is used: ; This correction ensures calculation continuity near true north (0° / 360°), avoiding numerical jumps. This parameter effectively distinguishes between flight views along the flight path, lateral side-view views, and large-angle circumferential views. Imaging distance. Preferably, data is obtained directly using an airborne laser rangefinder; in the absence of laser rangefinder data, the Euclidean distance between the UAV's three-dimensional coordinates and the center coordinates of the target component in the ledger is used for alternative calculations.

[0029] S103, Construct the environment state component. The environment state component is used to quantitatively describe the optical imaging quality and environmental texture features during imaging. This component is defined as a vector. Ground sampling distance This parameter represents the actual physical distance represented by a single pixel in an image, measured in millimeters per pixel. It reflects the image's ability to resolve physical details and is calculated using the following formula: ; In the formula, The physical width (in millimeters) of the imaging sensor target surface is an inherent hardware attribute of the imaging device 110, which is obtained by reading the device model and indexing in a preset sensor specification lookup table. The horizontal pixel width (in pixels) of the image; Focal length (mm); The imaging distance (in meters) is as calculated above; 1000 in the formula is a unit conversion factor. Exposure value Used to characterize ambient light intensity and image brightness levels, it is calculated based on the additive logarithmic aperture equation (APEX) system, which can transform the combination of aperture and shutter speed into a linear brightness perception index: ; In the formula, This refers to the aperture F-number. Exposure time. Image texture entropy. This metric is used to characterize the complexity and sharpness of image content. Physically, it quantifies the amount of information in an image; the entropy value decreases when fog, lens smudges, or overexposure cause loss of detail. During calculation, the original image is converted to grayscale, and the probability distribution of the grayscale histogram is statistically analyzed. Calculate using Shannon's entropy formula: ; S104, Generation and Standardization of Acquisition State Fingerprint Vector. The geometric state components and environmental state components calculated above are concatenated to generate the original acquisition state vector. Given the inconsistent physical dimensions and vastly different numerical ranges of the components (e.g., distances of hundreds of meters, while entropy values ​​are typically less than 10), the original vector is Z-score standardized using the statistical distribution parameters of historical source domain data to generate the final acquisition state fingerprint vector. The standardization formula is: ; In the formula, For the first element in the original vector The values ​​of each dimension and The full historical data of the source domain are respectively in the th The mean and standard deviation are pre-calculated across each dimension. The final output is... This serves as the mathematical basis for similarity retrieval and feature correction in subsequent steps.

[0030] In step S20, the source domain index building module 220 receives the geometric state components of the source domain data and constructs a geometric manifold index library using a spatial partitioning algorithm. The physical principle of this step lies in mapping discrete UAV inspection sampling points to a continuous four-dimensional Euclidean space. Within this space, the distance between data points no longer represents geographical proximity but rather the similarity of the shooting perspective. By constructing a spatial index, the system can transform the content-based image retrieval problem into a vector retrieval problem based on physical parameters. Specifically, it includes the following sub-steps: S201, Index Feature Subset Extraction and Data Cleaning. The source domain index building module 220 traverses all samples in the source domain dataset. For any sample... The standardized geometric state components are separated from the complete acquired state fingerprint vector. This component contains only four dimensions: relative altitude, pitch angle, relative yaw angle, and shooting distance. Since these four dimensions are Z-score normalized, they collectively constitute an isotropic unit hypercube feature space. During this process, the module performs outlier removal. If any dimension value exceeds a preset standard deviation threshold (preferably 3 times the standard deviation in this embodiment), then the value is considered to be within the threshold range. The samples are marked as outliers and not included in the index in order to remove noisy data caused by sensor failure or extreme aerobatic flight maneuvers, thus ensuring the smoothness of the manifold surface.

[0031] S202, KD-Tree construction based on variance maximization strategy. The source domain index building module 220 uses the KD-Tree algorithm to construct the geometric manifold index library. Given that the geometric state components are only four-dimensional, belonging to a low-dimensional data space, KD-Tree provides mathematically rigorous and accurate nearest neighbor results compared to approximate nearest neighbor algorithms (such as HNSW), and has lower memory consumption. The construction process adopts a recursive binary search strategy: First, calculate the variance of all geometric state vectors contained in the current node in each dimension, and select the dimension with the largest variance as the split dimension. Choosing the dimension with the largest variance as the splitting criterion ensures that the splitting plane is perpendicular to the direction of the most discrete data distribution, thereby minimizing the tree depth and optimizing retrieval efficiency. Second, sort all data of the current node along the split dimension, select the data point corresponding to the median as the splitting node, and divide the remaining data into the left subtree space and the right subtree space according to the value of this dimension. Finally, recursively execute the above splitting steps on the left and right subtrees until the number of samples contained in the child node is less than the preset leaf node capacity threshold. The optimal value for the leaf node capacity threshold is between 20 and 50. If the threshold is too small, it will lead to an excessively deep tree, increasing backtracking overhead; if it is too large, it will degenerate into a linear scan. Setting it within this range achieves a balance between storage overhead and search speed. The resulting tree structure is the geometric manifold index.

[0032] S203, Configuration of the retrieval metric space. To support subsequent nearest neighbor search, the source domain index building module 220 configures weighted Euclidean distance as a metric function to measure the similarity between two geometric state vectors. For any two geometric state vectors... and Their geometric difference distance The calculation formula is defined as follows: ; In the formula, The vector represents the first time. Numerical values ​​in each geometric dimension (i.e., altitude, pitch, yaw, and distance); The weighting coefficients for this dimension are determined in advance through sensitivity analysis on the validation set: while keeping other parameters constant, the weighting coefficients are individually perturbated to the first dimension. These parameters are used to observe the decrease in model detection accuracy (mAP). The greater the decrease, the more accurate the corresponding... The larger the value, the greater the weighting. In the default embodiment, given that changes in shooting distance have the most direct impact on target scale, the weighting value for the shooting distance dimension is set greater than the weighting value for the angle dimension. This metric configuration ensures that search results preferentially match the physical parameters that have the greatest impact on imaging features.

[0033] In step S30, the average geometric state vector of the target domain data is calculated by the dynamic anchor point selection module 230, and this vector is input into the geometric manifold index library as a query vector to retrieve and solve the difference vector. This step aims to solve the problem of blind selection of source domain samples in traditional transfer learning. Guided by physical parameters, the system can select the subset of samples with the highest geometric consistency with the current task's acquisition perspective from the source domain database. This mechanism simplifies the complex nonlinear domain transfer problem into a problem of small perturbation correction centered on physical anchor points, thereby reducing the difficulty of model adaptation. Specifically, it includes the following sub-steps: S301, Target Domain Reference State Vector Calculation. The dynamic anchor point optimization module 230 receives the target domain calibration dataset for the inspection task to be performed. This dataset consists of a small number of typical samples (e.g., 3 to 5 representative inspection images) collected by the UAV during the pre-flight phase of the target route, along with their accompanying telemetry data. The module performs an arithmetic mean of the acquisition state fingerprint vectors of all samples in the calibration dataset to obtain the target domain reference geometric vector. With the target domain reference environment vector In this calculation process, the principle of metric space uniformity must be followed: the Z-score standardization of the target domain data must strictly use the global mean and global standard deviation obtained from the source domain in step S104. Using local statistical parameters of the target domain itself will cause the coordinate systems of the two domain vectors to not coincide in the feature space, thus leading to the failure of distance calculation. The calculation formula is as follows: ; In the formula, To calibrate the sample size of the dataset for the target domain, For the first The original geometric data of each sample and These are the statistical parameters of the source domain.

[0034] S302, K-Nearest Neighbor Geometric Anchor Point Retrieval. As the query vector, a K-nearest neighbor search is performed in the geometric manifold index database constructed in step S20. The system retrieves the nearest neighbor in Euclidean distance. Each source domain sample constitutes a source domain anchor set. .parameter As a hyperparameter, its value is preferably set to a range of 5 to 10. This range is determined based on the following: if... Anchor point sets are easily affected by noise interference from individual outlier samples in the source domain; if This may introduce edge samples with significant geometric differences, leading to blurred average features in the subsequent generation. The output anchor set is formally defined as: ; In the formula, This represents the original image data corresponding to the anchor point.

[0035] S303, Decoupling Calculation of State Difference Vector. For each anchor point sample in the anchor point set, the state difference between it and the target domain reference vector is calculated. The system rigorously decouples the total difference into a geometric difference vector. Vector of difference with environment The formula for calculating the geometric difference vector is: ; Each component of this difference vector has a clear physical guiding meaning, directly corresponding to the operation type of subsequent image transformation: the height difference and distance difference are mainly related to the scaling transformation of the image; the pitch difference and yaw difference are mainly related to the perspective projection transformation and rotation transformation of the image.

[0036] The formula for calculating the environmental difference vector is: ; The environmental difference vector represents the numerical deviations in illumination gain, contrast, and sharpness, and is independent of geometric deformation. This orthogonal decoupling ensures that subsequent modules can independently handle both deformation and discoloration issues, avoiding feature aliasing caused by coupling interference in traditional methods.

[0037] S304, weighted aggregation of difference vectors. To reduce computational overhead and extract the most representative transfer features, the system is based on geometric distance pairs. Anchor points The difference vectors are weighted and aggregated. Weight coefficients Calculated using the Softmax function with a temperature coefficient: ; In the formula, This is a temperature coefficient used to adjust the smoothness of the weight distribution; in this embodiment, it is set to 1.0. The setting of this parameter is based on balancing the local specificity of the feature space with the smoothness of the manifold. Specifically, since the input geometric vector has been Z-score normalized, its Euclidean distance... The values ​​are usually distributed between 0 and 2.

[0038] like If the value is set too small (e.g., <0.5), the Softmax function will approach the Argmax operation, causing the aggregation result to be determined only by the nearest anchor point. This loses the linear interpolation advantage of the manifold surface and causes abrupt changes in feature variations when switching between different anchor points. like Setting the value too high (e.g., >5.0) will cause the weights of all anchor points to tend to be averaged, making the aggregated difference vector lose its specificity to the current viewpoint, resulting in a blurry feature map.

[0039] Setting it to 1.0 is based on experimental results from the validation set gradient. At this setting, the weight distribution can appropriately reflect the geometric neighborhood structure within the anchor set, retaining the dominant role of the nearest neighbor while introducing a smooth correction from the second nearest neighbor.

[0040] when When the distance is reduced, the weights will be more concentrated on the nearest anchor point. The final output is the aggregated geometric difference vector. This vector will be used as the sole geometric control signal input to the subsequent feature manifold remapping module 240.

[0041] In step S40, the aggregated geometric difference vector is received through the feature manifold remapping module 240, and the spatial alignment of the convolutional feature maps is driven by physical priors. The technical principle of this step lies in establishing a differentiable functional relationship between physical spatial differences and image feature space transformation. Unlike traditional spatial transformation networks (STNs) that rely on implicit learning to infer transformation parameters, this scheme directly utilizes explicit physical sensor differences (such as height differences and angle differences) to regress affine transformation parameters, thereby achieving geometric correction of features in an unsupervised manner. Specifically, it includes the following sub-steps: S401, Construction and initialization of the physical parameter mapping network. The feature manifold remapping module 240 integrates a lightweight mapping network using a multilayer perceptron (MLP) structure. This network specifically includes an input layer (4 dimensions), hidden layers (e.g., 64 dimensions, equipped with ReLU activation function), and an output layer (6 dimensions). The module aggregates the geometric difference vector output from step S30. The input signal is fed into the mapping network. To ensure that the feature map does not undergo unexpected and drastic deformation during the initial training phase, the output layer adopts an identity mapping initialization strategy: the initial values ​​of the output layer weights are set to all zeros, and the initial values ​​of the output layer biases are set to physical parameters that can produce an identity transformation (i.e., [1,1,0,0,0,0], corresponding to...). After forward computation and processing by a hyperbolic tangent (Tanh) activation layer at the end, the four-dimensional physical difference vector is nonlinearly mapped to a six-dimensional affine transformation parameter set. .in, The scaling factor of the corresponding feature map in the horizontal and vertical directions. Corresponding rotation factor, Corresponding shear factor, Corresponding translation factor.

[0042] S402, Generation of the affine transformation matrix. Parameter set obtained based on mapping network regression. Construct a 2x3 affine transformation matrix for feature map space sampling. This matrix defines the mapping relationship between the feature coordinates of the source domain and the feature coordinates of the target domain: ; Each element of the matrix corresponds to a transformation component of the physical camera's imaging model: the diagonal elements mainly respond to changes in shooting distance and focal length, while the off-diagonal elements mainly respond to changes in gimbal pitch and yaw angles.

[0043] S403, Mesh sampling and spatial transformation of the feature map. The feature manifold remapping module 240 receives the original feature map of the source domain extracted by the backbone network. ,in These represent the number of channels, height, and width, respectively. The generated affine transformation matrix is ​​then used. This generates a normalized sampling grid. The coordinate space range of the feature map is defined here as follows: Where (-1, -1) corresponds to the top left corner of the feature map, and (1, 1) corresponds to the bottom right corner. For any normalized pixel coordinate on the output feature map... The corresponding sampling coordinates on the input feature map are calculated according to the following formula. : ; Due to the calculation Typically, the coordinates are non-integer coordinates, and the module uses a bilinear interpolation algorithm for sampling. Generate a geometrically corrected feature map from the pixel values ​​at the corresponding positions. .

[0044] S404, confidence-gated calculation based on geometric modulus. To prevent forced geometric transformations from causing loss of effective information or severe stretching and distortion of the feature map when the physical viewpoints of the source and target domains differ too much (e.g., the source domain is viewed from above, and the target domain is viewed at eye level), the system introduces a confidence-gated mechanism. It calculates the Euclidean modulus of the aggregated geometric difference vector. and compare it with the preset security transformation threshold. Compare and generate confidence coefficients : ; In the formula, This is a safety transformation threshold. Since the input geometric difference vector is Z-score normalized, this threshold is statistically significant and typically ranges from 3.0 to 5.0. The threshold setting is based on statistical principles. Rule (Three Sigma Criterion). Reviewing step S104, the collected state fingerprint vector is Z-score normalized based on the mean and standard deviation of historical data from the source domain. This means that the feature distribution of the source domain data mathematically approximates a standard normal distribution. According to the properties of normal distribution, approximately 99.7% of the sample data points should fall within 3 standard deviations (i.e., a value of 3.0) of the mean.

[0045] When the calculated geometric difference vector magnitude When the value exceeds 3.0 (or is relaxed to 5.0 to tolerate a slight long-tailed distribution), it is statistically determined that the current target domain sample has fallen outside the effective coverage of the source domain data.

[0046] In a physical sense, this means that the current shooting angles (such as extreme low-angle shots or ultra-long distance shots) have never had similar reference samples in the training set.

[0047] If affine transformation parameters are forcibly generated through extrapolation via a neural network at this point, it is highly likely to produce erroneous stretching or distortion (i.e., negative transfer). Therefore, utilizing... The statistical boundary, acting as a truncation threshold, mathematically guarantees the safety and robustness of feature remapping operations. This means the maximum allowable difference range is 3 to 5 standard deviations of the source domain's statistical distribution. When the difference magnitude exceeds this range, the physical difference is considered too large to be aligned using a simple planar affine transformation. Setting it to 0 will automatically block the path transformation.

[0048] S405, soft-weighted residual fusion. Utilizing confidence coefficients. For the original feature map Compared with the corrected feature map Dynamic weighted fusion is performed to obtain the final output remapped feature map. : ; This fusion step constitutes a physically controlled residual connection structure. In scenarios with small-angle fine-tuning, the network primarily utilizes corrected features to achieve more accurate spatial alignment; in scenarios with large-angle drastic changes, the network automatically degenerates to use the original features, ensuring the model's robustness under extreme conditions. Finally... It is then transmitted to subsequent detection heads or classification layers for task prediction.

[0049] In step S50, the environment difference vector is received through the environment adaptive weight module 250, and channel attention gating vectors are generated through the hypernetwork structure to achieve dynamic reorganization of model weights. The core technical concept of this step lies in parameter-level domain adaptation, that is, without retraining the model, the channel response weights of the convolutional neural network feature extraction layer are dynamically adjusted according to the physical differences between the current task environment (such as light intensity and atmospheric transparency) and the source domain environment. This mechanism allows the same set of model parameters to adapt to different physical imaging environments through soft adjustment. Specifically, it includes the following sub-steps: S501, Weighted aggregation of environmental difference vectors. Similar to the geometric difference aggregation in step S304, this step first... The environmental difference vectors corresponding to each anchor point are aggregated to extract the overall environmental offset feature of the target domain relative to the source domain. The weighting coefficients based on geometric distance calculated in step S304 are then used. Calculate the aggregated environment difference vector : ; Geometric weights are reused here. The basis for this is that sampling points that are physically close (such as continuous shooting of the same tower) usually have a high degree of spatiotemporal correlation in their ambient lighting conditions and background texture features.

[0050] S502, Hypernetwork Construction and Feature Mapping. The environment-adaptive weight module 250 embeds a lightweight hypernetwork to establish the mapping relationship between physical environment parameters and deep learning feature channels. This hypernetwork adopts a multilayer perceptron (MLP) structure containing input, hidden, and output layers. The input layer dimension is 3 (corresponding to ground sampling distance differences, exposure value differences, and texture entropy differences), and the output layer dimension is... ,in The number of channels in the convolutional layer to be adjusted. The hidden layer dimension is preferably set to [value missing]. To achieve feature compression. The hypernetwork receives and aggregates the environmental difference vector. The system extracts latent environmental features through nonlinear transformations. This process converts low-dimensional physical environment descriptors into high-dimensional feature channel importance descriptors. For example, when exposure differences indicate a darker target area, the hypernetwork automatically enhances channel features that are more sensitive to edge extraction and suppresses channel features that are more dependent on color and texture.

[0051] S503, Generation of Channel Attention Gated Vectors. To transform the hypernetwork's output into normalized coefficients that can be used to adjust the weights, a Sigmoid activation function is used to process the hypernetwork's output, generating channel attention gated vectors. The calculation formula is: ; In the formula, and All of these are learnable parameters of the hypernetwork; The sigmoid function ensures that the output value is confined to the (0,1) interval. To improve the model's stability under minor environmental differences, this embodiment improves the above formula by employing a residual gating mechanism, defining the final gating vector as... This improvement makes the gating vector approach 1 when the environmental difference is 0, that is, the original weights remain unchanged.

[0052] S504, nonlinear weighted fusion of convolutional kernel weights (feature recalibration). Utilizing the generated gate vectors... Channel-level recalibration is performed on the convolutional kernel weights of specific convolutional layers in the pre-trained model. Let the original convolutional kernel weight tensor be... ,in Number of output channels Input the number of channels. This represents the kernel size. The adjusted dynamic weights. The calculation method is as follows: ; In the formula, The first in the gated vector scalar coefficients corresponding to each channel; Indicates the first All parameters of each convolutional kernel. Through this step, the model has adaptively adjusted the parameters according to environmental differences before performing forward inference computation. This adjustment method involves a transformation of the parameter space, unlike traditional attention mechanisms based on input feature maps (such as SE-Net). It directly changes the physical response characteristics of the filters, thereby achieving low-level adaptation to different lighting and sharpness conditions. Adjusted weights It is loaded into the convolutional layer for subsequent feature extraction operations.

[0053] In step S60, the original detection results output by the neural network are post-processed and verified by the physical scale constraint module 260. The core technical principle of this step lies in establishing a closed-loop verification mechanism between visual features and physical rules. Given that deep learning-based target detection models are essentially probabilistic inferences based on pixel statistical regularities, they are prone to visual deception during cross-domain transfer (i.e., detecting false targets with similar textures but sizes that do not conform to physical common sense). This step uses the precisely calculated ground sampling distance (GSD) in step S10 as a physical benchmark to restore the pixel-scale detection boxes to the true size of the physical world, and combines statistical prior distribution to calibrate the detection confidence, thereby significantly reducing the false alarm rate. Specifically, it includes the following sub-steps: S601, Restoration of the physical equivalent geometric parameters of the candidate target. In this embodiment, the physical equivalent geometric parameters are specifically represented as physical length (for insulators) or physical diameter (for dampers). The physical scale constraint module 260 receives the original candidate box set output by the detection head. For any one of the candidate boxes... Each candidate target has original information including the pixel width of the predicted bounding box. Pixel height and the original confidence level of the category. The module calls the ground sampling distance of the current image calculated in step S103. (Unit: mm / pixel) Calculate the physical equivalent geometric parameters of the candidate target. To improve robustness to rotation and viewpoint changes, this embodiment preferably uses the longer side of the bounding rectangle of the candidate box as the scaling metric, rather than simply calculating the area, to avoid area estimation errors caused by occlusion. The calculation formula is: ; In the formula, the calculation result The physical unit is the millimeter.

[0054] S602, Construction of Category-Level Physical Prior Distributions. The system pre-loads the statistical distributions of the physical equivalent geometric parameters of various components in the source domain dataset. For the first... For objects of this type (such as vibration dampers and insulator strings), their physical equivalent geometric parameters typically follow a Gaussian distribution. The module reads the statistical mean of this type of object. with statistical standard deviation The statistical parameters are obtained as follows: From the historical data of the source domain, the physically equivalent geometric parameters of all labeled samples are inversely calculated using their corresponding historical GSDs. The intrinsic size distribution parameters of each category in the physical world are then obtained through maximum likelihood estimation. These parameters are embedded as prior knowledge in the model configuration file.

[0055] S603, Physical Consistency Probability Calculation. This involves using the calculated physical equivalent geometric parameters. Calculate the physical consistency probability of the candidate target using the prior distribution parameters of the corresponding category. This probability value characterizes the likelihood that the currently detected object belongs to this category in terms of physically equivalent geometric parameters. To prevent missed detections due to single-dimensional size truncation, a Gaussian kernel function is used for soft constraint calculation: ; In the formula, is the relaxation coefficient, used to control the tolerance for dimensional anomalies. In this embodiment, slight measurement errors that may exist in the target domain are taken into account. The preferred setting is 2.0 to 3.0. The physical meaning of this formula is that the further the physical equivalent geometric parameters of the detected target deviate from the mean, the more exponentially the probability of its physical consistency decreases.

[0056] S604, Confidence Recalibration and Non-Maximum Suppression. This involves adjusting the physical consistency probability. As a weighting factor, the confidence level of the original detection Calibration is performed to obtain the final closed-loop confidence level. : ; Through this multiplicative fusion operation, artifacts with similar visual textures but extremely unreasonable physical equivalent geometric parameters (such as misidentifying white litter on the ground as a string of insulators several meters long, or misidentifying distant buildings as tiny bolts) can be effectively fused. It will approach 0, thus greatly suppressing Subsequently, the calibrated confidence level was used. Non-maximum suppression is performed to remove overlapping boxes and output the final detection results. This process ensures that the final inspection report is not only visually credible but also has rigorous consistency in physical logic.

[0057] See attached document Figure 3-6 Specific application example: Defect inspection of insulator strings for 500kV transmission lines Scene background: Source domain (training data): historically accumulated standard inspection datasets with clear skies, orthophoto perspective, and a distance of 20 meters.

[0058] Target domain (current mission): A drone performing a mission is encountering atypical conditions such as cloudy weather (low light), large side-view angle (yaw angle 30 degrees), and long distance (35 meters).

[0059] Test object: Composite insulator string (standard physical length is 2400mm).

[0060] Detailed Explanation of Implementation Steps Step S1: Data parsing and state fingerprint construction. The drone acquired the image of frame 1024.

[0061] Geometric data extraction: Laser ranging for distance display Meters, gimbal yaw angle Environmental data extraction: Due to the long exposure time, the exposure value was calculated. (Darker than usual); Calculate ground sampling distance mm / pixel. Vectorization: After Z-score normalization, the target domain reference vector is obtained. .

[0062] Steps S2-S3: The manifold indexing and dynamic anchor point optimization system will... Input the KD-Tree index.

[0063] Search results: Five geometrically closest images (anchor point set) were found in the historical database. Although the historical database is mostly orthophotos, the search algorithm found a small number of side-view samples taken at previous turns.

[0064] Difference resolution: Geometric difference vector: Calculated to show that the target is 15 meters further away from the anchor point and the angle is 5 degrees off.

[0065] Environmental difference vector: Calculates that the target is 2 stops darker than the anchor point.

[0066] Step S4: Feature manifold remapping (geometric alignment) Affine transformation: The mapping network generates a scaling factor based on a distance of 15 meters. (That is, the feature map needs to be scaled down to simulate the distant view), and a rotation factor is generated based on the 5-degree deviation.

[0067] Mesh sampling: Bilinear interpolation is performed on the source domain feature map. Effect: This makes the model believe it is seeing the object at standard distances and angles, thus activating the correct feature extractors.

[0068] Step S5: Environment-Adaptive Weights (Style Alignment) Hypernetwork computing: Input dark 2-level environmental differences.

[0069] Gated generation: The supernetwork outputs a weight vector. It automatically increases the weights of convolutional channels related to contour extraction (because color is unreliable in low light, and contours are more important), and decreases the weights of channels related to texture details.

[0070] Effect: The model parameters were dynamically adjusted, turning it into a model specifically for low-light conditions.

[0071] Step S6: Physical Scale Consistency Closed Loop Model prediction: The neural network outlines a target in the image with a pixel width of 460 pixels.

[0072] Physical inverse calculation: =2392mm.

[0073] Consistency verification: According to the table, the standard length of this type of insulator is 2400±50mm. Calculation shows that 2392mm falls within the high probability range, with extremely high confidence, and is therefore determined to be a valid target.

[0074] Counterexample comparison: If the model incorrectly identifies a white, long strip of trash (actual physical length 500mm) in the background as an insulator, the probability of physical consistency after reverse calculation is extremely low, and the system will automatically filter out the false alarm.

[0075] Experimental verification: To verify the effectiveness and advancement of the proposed method for constructing a UAV multi-scenario inspection model that integrates transfer learning and physical state guidance, the applicant conducted detailed comparative tests and verifications in a real 500kV ultra-high voltage transmission line inspection scenario.

[0076] 1. Experimental Dataset Construction: The experimental data comes from the historical inspection database provided by a provincial power grid company and recent special test flights.

[0077] Source Domain dataset: 5,000 standard, refined inspection images accumulated over time. Data characteristics include: sufficient lighting (sunny morning 10:00-14:00), shooting angle of orthogonal or slight tilt (pitch angle -90°±5°), and shooting distance stable within 15-20 meters.

[0078] Target Domain Dataset: This dataset contains 1,200 test images recently collected for complex operating conditions. The data features include: coverage of cloudy and low-light environments at dusk; large-angle side shots (yaw angle difference > 25°); and long-distance aerial shots (distance > 30 meters). These data simulate actual conditions during non-standard operations or emergency inspections.

[0079] 2. Evaluation Metrics and Benchmark Comparison The main evaluation metrics used in the benchmark experiment are the mean accuracy (mAP@0.5) to assess detection accuracy and the false alarm rate (FAR) to assess the model's robustness against non-target backgrounds (such as litter on the ground, tree branches, etc.). The comparison algorithms include: Baseline: A standard Faster R-CNN model trained using only source domain data, tested directly in the target domain.

[0080] Method A (DANN): A method that introduces the classic Domain-Adversarial Neural Network for feature alignment.

[0081] Method B (CycleGAN): A method for pixel-level enhancement using an image style transfer network.

[0082] Ours: This invention proposes a method that integrates geometric manifold remapping and physical scale constraints.

[0083] The experimental results and analysis show the performance of each method on the target domain test set in the following table: The symbols “×” and “√” are used to indicate whether the corresponding technical module has been used.

[0084] Results analysis: Overall performance improvement: Experimental data shows that our invention (Ours) achieved an mAP of 74.7% in the target domain, which is 12.3% higher than the baseline model (62.4%) that directly transfers data. This indicates that by constructing the state space and remapping the feature manifold, the model effectively overcomes the differences in feature distribution caused by changes in shooting angle and distance.

[0085] The false positive rate was significantly reduced: After introducing the physical scale constraint module 260, the false positive rate (FAR) dropped dramatically from 28.5% at the baseline to 9.8%, a reduction of 18.7%. Comparing Ours (without physical constraints) and Ours (complete model), it can be seen that while relying solely on visual feature alignment can improve the recall rate, it cannot completely filter out visually similar artifacts; however, with the addition of physical scale constraints, a large number of false targets that conform to visual features but do not conform to physically equivalent geometric parameters (such as being too small or too large) were successfully eliminated.

[0086] Computational efficiency: Although this invention introduces a state vector retrieval and parameter generation network, the use of a lightweight super network and KD-Tree index increases the inference time per frame by only about 13ms, which still meets the real-time or near-real-time requirements of UAV inspection data processing.

[0087] In a typical scenario qualitative analysis, during a side-look long-distance inspection test of a 500kV insulator string, the baseline model failed to identify the insulator (missed detection) due to blurred target texture, and misidentified a section of white agricultural film in the background as an insulator (false detection). The model of this invention, by reading RTK data and the pan-tilt angle, automatically activated geometric feature remapping, successfully enhancing the insulator's feature response and achieving accurate detection. Simultaneously, by calculating the ground sampling distance of the agricultural film and inferring its physical length to be only 0.6 meters (far smaller than the standard insulator length of 2.4 meters), it used physical consistency probability to classify it as a negative sample, thus greatly suppressing false alarms while maintaining a high detection rate.

[0088] In summary, by introducing explicit physical prior knowledge and a geometric manifold indexing mechanism, this invention solves the problem of poor generalization ability of traditional data-driven algorithms in situations with few samples and across different scenarios, and has extremely high engineering application value.

Claims

1. A method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance, characterized in that, Includes the following steps: The raw inspection data is parsed and vectorized to construct a standardized collection state fingerprint vector, wherein the raw inspection data includes source domain data and target domain data; the collection state fingerprint vector is decoupled in dimension into geometric state components and environmental state components. Receive the geometric state components of the source domain data and construct a geometric manifold index library; Calculate the baseline geometric state vector of the target domain data, use it as a query vector to retrieve the source domain anchor set in the geometric manifold index, calculate the state difference vector between the target domain vector and the source domain anchor set vector, and decompose the state difference vector into a geometric difference vector and an environmental difference vector. The geometric difference vector is received, and the mapping network is used to generate an affine transformation parameter matrix. Mesh sampling and interpolation operations are then performed on the feature map extracted from the source domain image. The environmental difference vector is received, and a gating vector is generated by inputting it into the hypernetwork structure. The weight parameters of the pre-trained model are then nonlinearly weighted and fused. Physical prior constraints are introduced during the backpropagation stage of model training. The target detection bounding box coordinates output by the model are received and the physical equivalent geometric parameters of the predicted target are calculated. The consistency loss value between the physical equivalent geometric parameters and the pre-stored standard specification data is calculated to drive parameter updates and obtain the inspection data analysis model.

2. The method for constructing a UAV multi-scenario inspection model integrating transfer learning and physical state guidance according to claim 1, characterized in that, The steps for constructing a standardized acquisition state fingerprint vector include: Extract the telemetry data and sensor metadata associated with each frame of the original inspection data; The geometric state components are constructed, including relative height, shooting distance, gimbal pitch angle and relative yaw angle, to characterize the spatial geometric relationship during imaging; The environmental state component is constructed, including ground sampling distance, exposure value and image texture entropy, to characterize the ambient lighting and texture sharpness conditions during imaging; Using the statistical distribution parameters of historical data from the source domain, the geometric state component and the environmental state component are Z-score standardized to generate the acquisition state fingerprint vector.

3. The method for constructing a UAV multi-scenario inspection model integrating transfer learning and physical state guidance according to claim 2, characterized in that, The construction of the geometric manifold index library employs a spatial partitioning algorithm, specifically including: The K-dimensional tree algorithm is used as the spatial partitioning structure; Calculate the variance of the geometric state components of the source domain data in each dimension, and select the dimension with the largest variance as the split dimension of the current node; The median is selected as the splitting node along the splitting dimension, and the left subtree space and the right subtree space are recursively divided until the number of samples contained in the leaf node is less than the preset capacity threshold. We use weighted Euclidean distance as a metric to measure the similarity of geometric state vectors, where the weight of the shooting distance dimension is set greater than the weight of the angle dimension.

4. The method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance according to claim 1, characterized in that, The steps for calculating the baseline geometric state vector of the target domain data include: Obtain the target domain calibration dataset, which contains several pre-collected typical samples; The samples in the target domain calibration dataset are standardized using the global mean and global standard deviation of the source domain data. The reference geometric state vector is obtained by arithmetically averaging all sample vectors in the processed target domain calibration dataset.

5. The method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance according to claim 1, characterized in that, The specific steps of decomposing the state difference vector into a geometric difference vector and an environmental difference vector include: Calculate the initial difference vector between each anchor point sample in the anchor point set and the reference geometric state vector; Using the Softmax function with a temperature coefficient, the aggregation weight is calculated based on the geometric spatial distance between each anchor point sample and the baseline geometric state vector. The initial difference vectors of all anchor point samples are weighted and summed using the aggregation weights to obtain the aggregated geometric difference vector and the environmental difference vector, respectively.

6. The method for constructing a UAV multi-scenario inspection model integrating transfer learning and physical state guidance according to claim 1, characterized in that, The specific steps for performing grid sampling and interpolation on the feature maps extracted from the source domain image include: A normalized sampling grid is generated using the affine transformation parameter matrix; For non-integer coordinates in the sampling grid, a bilinear interpolation algorithm is used to sample the pixel values ​​at the corresponding positions in the source domain feature map to generate a geometrically corrected feature map.

7. The method for constructing a UAV multi-scenario inspection model integrating transfer learning and physical state guidance according to claim 1, characterized in that, Before performing grid sampling and interpolation operations on the feature maps extracted from the source domain image, a confidence gating step is also included: Calculate the Euclidean modulus of the geometric difference vector; Determine whether the Euclidean modulus exceeds a preset safety transformation threshold; If the safety transformation threshold is exceeded, the affine transformation parameter matrix is ​​reset to the identity transformation matrix, or the grid sampling operation is stopped and the original source domain feature map is directly output.

8. The method for constructing a UAV multi-scenario inspection model integrating transfer learning and physical state guidance according to claim 1, characterized in that, The steps for generating gated vectors from the input hypernetwork structure include: A hypernetwork is used to establish a mapping relationship between environmental difference features and the importance of convolutional channels, and a feature channel importance descriptor is output. The feature channel importance descriptors are processed using an activation function to generate the gated vector mapped to a normalized numerical range; The nonlinear weighted fusion is specifically manifested in the use of the gating vector to perform a channel multiplication-based recalibration operation on the convolutional kernel weights of the convolutional layer.

9. The method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance according to claim 2, characterized in that, The steps of receiving the target detection bounding box coordinates output by the model and back-calculating the predicted target's physical equivalent geometric parameters include: Receive the target detection bounding box output by the model and obtain the pixel dimension value of the bounding box; The pixel dimension value is multiplied by the ground sampling distance value recorded in the environmental state component to obtain the physical equivalent geometric parameters of the predicted target. The ground sampling distance value represents the actual physical distance represented by a single pixel in the image.

10. The method for constructing a multi-scenario inspection model for unmanned aerial vehicles (UAVs) that integrates transfer learning and physical state guidance according to claim 1, characterized in that, The steps for calculating the consistency loss value between the physical equivalent geometric parameters and the pre-stored standard specification data specifically include: Retrieve the statistical mean and standard deviation of the physical equivalent geometric parameters of the predicted target category from the pre-stored standard component specification data; The physical consistency probability of the predicted target's physical equivalent geometric parameters belonging to the category to which the predicted target belongs is calculated using the Gaussian kernel function; A loss function can be constructed based on the physical consistency probability, or the physical consistency probability can be used as a weighting factor to calibrate the confidence of the predicted target, thereby suppressing spurious targets with unreasonable physical equivalent geometric parameters.

Images