Dam leakage detection method and system based on unmanned aerial vehicle vision

Abstract

The application provides a kind of dam leakage detection method and system based on unmanned aerial vehicle vision, including collecting dam image, using the hybrid convolution kernel containing direction sensitivity, irrelevant sub-core, combined with gradient covariance matrix to generate coherence map and local structure main direction angle, directional response feature map is obtained by sub-core direction re-alignment and coherence map gate weighted fusion;Positioning potential leakage center, extract neighborhood features and establish local coordinate system, determine the main axis of leakage diffusion according to the covariance principal component vector, and after rotating alignment, map to polar coordinates to obtain radial angle feature tensor composed of feature channel, radial distance and polar angle;Build a convolution module, input the leakage center feature to generate radial, polar angle one-dimensional convolution kernel shared by each feature channel, extract the leakage radial diffusion intensity and angle uniformity feature to form the discrimination feature, and output the dam leakage area position.

Description

Technical Field

[0001] This application belongs to the field of UAV vision, and in particular relates to a method and system for detecting dam seepage based on UAV vision. Background Technology

[0002] Dikes are crucial water conservancy infrastructure for flood control, disaster reduction, water resource allocation, and industrial and agricultural production. Affected by geological conditions, climate change, and aging from long-term service, dikes are prone to various defects during operation, among which seepage is characterized by its high frequency, strong concealment, and high risk. If not detected and addressed in time, seepage may further develop into piping or even dam failure, causing serious loss of life and property. Traditional dike seepage inspection mainly relies on manual inspection, which suffers from low efficiency, strong subjectivity, limited coverage, and difficulty in adapting to rugged and complex terrain. In recent years, drone technology has developed rapidly. Using drones equipped with visual sensors for dike inspection has become an important, safe, efficient, and low-cost technical means. Drones can flexibly acquire high-resolution images of large areas of the dike surface, providing effective data support for intelligent operation and maintenance of water conservancy projects.

[0003] However, during drone flight data collection, the seepage areas in the images often exhibit significant scale variations and complex directional distributions due to the influence of the shooting angle, flight attitude, and the undulating morphology of the dam surface. Because water flow is affected by gravity, dam material, and local micro-topography, seepage defects often have obvious spatial spread textures and show a morphological pattern of diffusion from the seepage center outwards. Existing conventional image feature extraction methods, when dealing with such natural targets with direction dependence and radial diffusion patterns, typically lack targeted modeling of rotational changes and diffusion structures, making it difficult to accurately characterize the local dominant direction of the seepage area. They also often remain at the level of Cartesian coordinates for local area analysis, failing to fully explore the intensity variations and angular uniformity of radial diffusion, resulting in extraction results easily affected by changes in the drone's viewing angle and shooting orientation. This inadequate feature representation and spatial misalignment limit the model's ability to identify irregular seepage areas in complex backgrounds, affecting detection accuracy and stability. Therefore, there is an urgent need for a detection method that can integrate local image directional information with seepage diffusion characteristics to improve the reliability of dam seepage identification under varying drone perspectives. Summary of the Invention

[0004] According to one aspect of the present invention, a method for detecting dam seepage based on UAV vision is proposed to solve the identification problem under the changing perspective of UAVs, comprising the following steps: Images of the dam collected by a drone are acquired, and a hybrid convolutional kernel consisting of a direction-sensitive kernel and a direction-independent kernel is constructed. A coherence map and a principal orientation angle of the local structure are generated based on the local gradient covariance matrix. Image features are extracted using the hybrid convolutional kernel. The outputs of the direction-sensitive kernel and the direction-independent kernel are weighted and fused using the coherence map as a gating signal. The principal orientation angle of the local structure is used to remanipulate the direction-sensitive kernel so that it is aligned with the extension direction of the local structure, thereby generating a directional response feature map. The potential leakage center is located on the directional response feature map. The neighborhood feature map is extracted as the neighborhood feature and a local coordinate system is established. The covariance matrix of the feature distribution within the neighborhood feature is calculated and the direction of the principal feature vector is determined as the leakage diffusion principal axis. The local coordinate system principal axis is aligned with the leakage diffusion principal axis by rotation alignment. The aligned neighborhood feature is transformed to the polar coordinate system based on the polar coordinate mapping relationship to obtain the radial angle feature tensor composed of feature channel, radial distance and polar angle. A convolution module is constructed, which takes the feature vector of the potential leakage center as the input of the potential leakage center feature. It generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor. One-dimensional convolution is performed on the radial angle feature tensor along the radial axis and polar axis respectively to identify the diffusion intensity feature of the leakage along the radial direction and the uniformity feature of the angular distribution of the leakage as the leakage diffusion characteristics, generating leakage discrimination features. Based on the leakage discrimination features, the location of the leakage area is output.

[0005] According to another aspect of the present invention, a dam seepage detection system based on UAV vision is proposed, comprising the following modules: The module is used to acquire images of the dam collected by UAVs, construct a hybrid convolutional kernel composed of orientation-sensitive and orientation-independent kernels, generate a coherence map and local structure principal orientation angle based on the local gradient covariance matrix, extract image features using the hybrid convolutional kernel, perform weighted fusion of the outputs of the orientation-sensitive and orientation-independent kernels using the coherence map as a gating signal, and remanufacture the orientation-sensitive kernel using the local structure principal orientation angle to align the orientation-sensitive kernel with the extension direction of the local structure, thereby generating an orientation response feature map. The calculation module is used to locate potential leakage centers on the directional response feature map, extract neighborhood feature maps as neighborhood features and establish a local coordinate system, calculate the covariance matrix of feature distribution within the neighborhood features and determine the direction of the principal feature vector as the leakage diffusion principal axis, align the principal axis of the local coordinate system with the leakage diffusion principal axis through rotation and alignment, and transform the aligned neighborhood features to the polar coordinate system based on the polar coordinate mapping relationship to obtain the radial angle feature tensor composed of feature channels, radial distance and polar angle; The output module is used to construct the convolution module. It takes the feature vector of the potential leakage center as the input of the potential leakage center feature and generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor. One-dimensional convolution is performed on the radial angle feature tensor along the radial axis and polar axis respectively to identify the diffusion intensity feature of the leakage along the radial direction and the uniformity feature of the angular distribution of the leakage as the leakage diffusion characteristics, generate leakage discrimination features, and output the location of the leakage area based on the leakage discrimination features.

[0006] This invention constructs a hybrid convolutional kernel composed of a direction-sensitive sub-kernel and a direction-independent sub-kernel. It uses a coherence graph to perform gated weighted fusion of the two types of sub-kernel outputs and re-manipulates and aligns the direction-sensitive sub-kernel based on the principal orientation angle of the local structure. This enhances the feature extraction capability of directional textures in dam images and reduces interference from irrelevant background textures. A local coordinate system is established for potential seepage areas, and the seepage diffusion principal axis is determined. After rotation and alignment, the neighborhood features are mapped to a polar coordinate system, forming a radial angle feature tensor composed of feature channels, radial distance, and polar angle. This makes the feature expression more consistent with the spatial law of seepage diffusion from the center outwards and reduces feature deviation caused by UAV perspective rotation. Furthermore, based on the potential seepage center features, a one-dimensional radial convolutional kernel and a one-dimensional polar angle convolutional kernel shared by all feature channels are generated. Radial diffusion intensity features and angular distribution uniformity features are extracted in the polar coordinate domain, respectively, achieving decoupled expression of seepage diffusion characteristics and improving the characterization accuracy of seepage discrimination features and the stability of dam seepage area location detection. Attached Figure Description

[0007] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram showing the variation curves of interpolation coefficients of the manipulable basis filter in the orientation-sensitive subnucleus under different discrete alignment angles; Figure 3 This is a schematic diagram of the decay curve of the characteristic amplitude along the radial distance in polar coordinates; Figure 4 This is a schematic diagram of the leakage characteristic response density distribution curve in the polar angular dimension. Detailed Implementation

[0008] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0009] In the first embodiment, the present invention proposes a method for detecting dam seepage based on UAV vision, such as... Figure 1 It includes the following steps: S1. Acquire images of the dam collected by the UAV, construct a hybrid convolutional kernel composed of a direction-sensitive kernel and a direction-independent kernel, generate a coherence map and a local structure principal orientation angle based on the local gradient covariance matrix, extract image features using the hybrid convolutional kernel, use the coherence map as a gating signal to perform weighted fusion of the outputs of the direction-sensitive kernel and the direction-independent kernel, and use the local structure principal orientation angle to remanipulate the direction-sensitive kernel so that it is aligned with the extension direction of the local structure, thereby generating a directional response feature map.

[0010] Real-time video stream data is acquired using an unmanned aerial vehicle (UAV) system, and the `cv2.VideoCapture` function is used to read the current frame as the captured embankment image. The `cv2.resize` function is used to normalize the resolution of the captured embankment image to 512 x 512 pixels. This image is then input into a PyTorch deep learning architecture and converted into floating-point tensor data. A network model with two parallel branches is constructed: a direction-sensitive branch uses a manipulable basis filter composed of Gaussian derivative functions to generate a direction-sensitive kernel, and a direction-independent branch uses a standard two-dimensional learnable convolutional layer to generate a direction-independent kernel. The `cv2.Sobel` function is used to calculate the pixel gradients of the embankment image in the horizontal and vertical directions. For each pixel coordinate, the sum of the squared gradients in the five-by-five neighborhood of that coordinate is extracted and combined with the product of the gradients to form a two-row, two-column local gradient covariance matrix. Two real eigenvalues ​​of each local gradient covariance matrix are calculated. The quotient of the difference between the largest and smallest eigenvalues ​​is calculated and divided by the sum of the two eigenvalues. The quotients for all coordinate positions are combined to generate a coherence map ranging from 0 to 1. In the local gradient covariance matrix, the eigenvector corresponding to the largest eigenvalue represents the gradient normal where the grayscale change is most drastic, and the eigenvector corresponding to the smallest eigenvalue represents the local edge or texture extension direction orthogonal to the gradient normal. Therefore, the numpy.linalg.eig function is called to obtain the two-dimensional eigenvector corresponding to the smallest eigenvalue, and the arctangent angle value is calculated based on the eigenvector to determine the local structural principal direction angle of each pixel position.

[0011] During feature extraction, the orientation-sensitive sub-kernels are first discretized and remanufactured based on the principal orientation angle of the local structure, resulting in several global orientation-sensitive sub-kernels aligned with different discrete orientations. Then, each orientation-sensitive sub-kernel undergoes global convolution to pre-generate a multi-directional response feature map. The feature response of the corresponding orientation is selected according to the discrete orientation angle index of each pixel, yielding the orientation-sensitive feature output. Orientation-independent sub-kernels perform regular convolution on the same input feature, resulting in orientation-independent feature output. The coherence map is no longer compressed using the Sigmoid algorithm, but is instead pruned or maintained in the 0-1 range using a Min-Max linear mapping, serving as a gating weight matrix. This gating weight matrix is ​​element-wise multiplied with the orientation-sensitive feature output. The complementary weights, obtained by subtracting the gating weight matrix from a scalar of 1, are then element-wise multiplied with the orientation-independent feature output. The two weighted results are then summed and activated using the ReLU function to generate an orientation response feature map. This process avoids a second rotation of the convolution kernel after fusion, ensuring that orientation remanufacturing, orientation response selection, and gating fusion are implemented in a closed loop along a single forward computation link.

[0012] In an optional embodiment, the step of acquiring dam images collected by a drone and constructing a hybrid convolutional kernel composed of orientation-sensitive and orientation-independent sub-kernels includes: The Gaussian derivative function is used as a steerable basis filter with a direction-sensitive subkernel, and learnable weight parameters are assigned to steerable basis filters of each order. A two-dimensional orientation-sensitive subnucleus is generated by linearly weighting and summing the manipulated basis filters based on the weight parameters. Initialize a two-dimensional matrix network layer containing learnable parameters as a direction-independent sub-kernel; Along the output channel dimension, the convolution weights corresponding to the direction-sensitive sub-kernel and the convolution weights corresponding to the direction-independent sub-kernel are concatenated and fused to construct a hybrid convolution kernel weight tensor containing both direction-sensitive and direction-independent output channels.

[0013] When constructing the hybrid convolutional kernel, Gaussian derivative functions of orders 1 to 3 are selected as manipulable basis filters for the direction-sensitive sub-kernel, with the filter space size set to 7×7 or 9×9 pixels. Independent learnable floating-point weight parameters are assigned to manipulable basis filters of different orders and directions. These weight parameters are generated using the He normal distribution method during network initialization. The weight parameters are multiplied element-wise with the corresponding manipulable basis filters in the spatial domain and then linearly weighted and summed to generate a two-dimensional direction-sensitive sub-kernel. The number of output channels of this direction-sensitive sub-kernel is preset to half the total number of feature channels, such as 32 channels. When the direction-sensitive sub-kernel is synthesized from Gaussian derivative basis filters, the direction alignment means that the structural extension direction of the synthesized sub-kernel is consistent with the local structural extension direction; its gray-level derivative normal can be expressed by the orthogonal response direction inside the Gaussian derivative basis function, and the gradient normal is no longer mistakenly defined as the main direction of the local structure. A two-dimensional matrix network layer with a spatial size of 7×7 and containing standard learnable parameters is randomly initialized in the system. This two-dimensional matrix network layer is used as an orientation-independent sub-kernel to identify background textures without orientation preference. The number of output channels is set to 32. The two-dimensional matrix network layer is a linear structure with single-layer learnable weights. The input is the bottom-level input features, and the output is 32 channels of orientation-independent features. When constructing the hybrid convolutional kernel, the spatially aligned 32-channel orientation-sensitive convolutional weights and 32-channel orientation-independent convolutional weights are tensor-concatenated and fused along the output channel dimension to construct a hybrid convolutional kernel weight tensor with 64 output channels. These 64 output channels are used to represent the parallel output set of the two types of sub-kernels in the convolutional layer, where 32 output channels correspond to orientation-sensitive feature outputs and the other 32 output channels correspond to orientation-independent feature outputs. After gating and weighting fusion, a 32-channel directional response feature map is formed.

[0014] In an optional embodiment, generating the coherence map and local structural principal orientation angles based on the local gradient covariance matrix includes: The Sobel operator is used to calculate the horizontal and vertical gradients of each pixel in the dam image. The local gradient covariance matrix of each pixel is constructed using the squared terms and product terms of the horizontal and vertical gradients, and the matrix is ​​then subjected to Gaussian smoothing filtering. Calculate the eigenvalues ​​and eigenvectors of the smoothed local gradient covariance matrix, and use the normalized difference between the maximum and minimum eigenvalues ​​as the coherence map values ​​of the pixel. The direction angle of the feature vector corresponding to the smallest feature value is taken as the local structural principal direction angle of the pixel.

[0015] When generating the coherence map and determining the local dominant orientation angle, the 3×3 standard-size Sobel edge operator is used to perform spatial convolution on the dam image with an input resolution of 512×512, and the horizontal gradient at each pixel is calculated. and vertical gradient The squared term is calculated using the obtained horizontal and vertical gradients. , and product terms A 2×2 local gradient covariance matrix is ​​constructed for each pixel. To reduce high-frequency noise interference in the image, a two-dimensional Gaussian smoothing filter with a kernel size of 5×5 and a standard deviation σ of 1.0 is applied to locally smooth all elements within this covariance matrix. The two real eigenvalues ​​of the smoothed covariance matrix are then solved using singular value decomposition. and ,in And find the corresponding orthogonal eigenvectors; find the largest eigenvalue. with minimum eigenvalue The difference is used as the numerator, and the sum of the two is added to a value set as . Using a small constant to prevent division by zero as the denominator, the division between the two yields a normalized difference in the interval [0,1]. This difference is then assigned as the value of the pixel in the coherence map, representing the coherence intensity. The minimum feature value is extracted. The corresponding two-dimensional feature vector , By calculating the arctangent function Obtain the angle value within the half-circle angle range of [-90°, 90°), and set this angle value as the local structural main direction angle of the pixel, so that the subsequent direction-sensitive sub-kernel can be used for orientation alignment along the extension direction of the local structure.

[0016] In an optional embodiment, the weighted fusion of the outputs of the direction-sensitive sub-core and the direction-independent sub-core using the coherence map as a gating signal includes: The pixel values ​​of the coherence map are normalized to a preset range and used as the gate weight coefficients. The image features extracted using hybrid convolution kernels are decoupled along the channel dimension into orientation-sensitive feature outputs and orientation-independent feature outputs with the same number of channels. The orientation-sensitive feature output is weighted by pixel-by-pixel multiplication using gating weight coefficients; The difference between the value 1 and the gate weight coefficient is used as a complementary coefficient to perform pixel-by-pixel multiplication and weighting on the direction-independent feature output; The weighted direction-sensitive feature output and the direction-independent feature output are added element-wise to generate a fused feature map that covers the entire array.

[0017] During the gated weighted fusion operation, data in the coherence map falling within the floating-point range is normalized to a preset probability range of [0,1] using the Min-Max linear mapping algorithm, constructing a two-dimensional gated weight coefficient matrix with a resolution of 512×512. If the coherence map has already been calculated from the normalized difference and is within the [0,1] range, the coherence map is directly cropped to obtain the gated weight coefficient matrix without applying an additional Sigmoid function to avoid boosting the weights of low-coherence background regions above 0.5. The feature tensor extracted from the 64-channel hybrid convolution kernel output in the previous step is uniformly cut and stripped along the depth dimension of the feature tensor, decoupling it into 32-channel direction-sensitive feature output and 32-channel direction-independent feature output. In the computation graph, the above two-dimensional gated weight coefficient matrix is ​​extended to 32-layer dimensions along the channel axis and subjected to a per-spatial-pixel, per-channel Hadamard multiplication operation with the decoupled direction-sensitive feature output to amplify the response activation values ​​of strongly directional texture pixels. A complementary coefficient matrix representing isotropic probabilities is generated by subtracting the gated weight coefficient matrix pointwise using a scalar constant of 1.0. This matrix is ​​then dimensionally expanded and multiplied pixel-wise with the orientation-independent feature output. This process preserves necessary background context in low-coherence regions and reduces interference from irrelevant background textures in high-coherence regions. Element-wise tensor addition is applied to the orientation-sensitive and orientation-independent feature tensors, which have undergone independent weighting, at their spatial locations. This results in the generation of a global full-array coverage fusion feature map with a channel dimension restored to 32.

[0018] In an optional embodiment, the step of remanipulating the orientation-sensitive sub-core using the principal orientation angle of the local structure to align the orientation-sensitive sub-core with the extension direction of the local structure and generate an orientation response feature map includes: Obtain the local structural principal orientation angles corresponding to the extraction of spatial location features, and quantize the orientation angles into a preset number of discrete orientation angles; Each discrete direction angle is used as an analytical parameter, and a set of interpolation coefficients corresponding to each angle are calculated. The linear combination weights of the manipulable basis filters in the direction-sensitive sub-kernel are updated using each set of interpolation coefficients to generate a set of global direction-sensitive sub-kernels aligned with each discrete direction angle. A set of global orientation-sensitive sub-kernels are used to perform global convolution calculations to pre-generate a set of multi-directional response feature maps; At each spatial location, based on the corresponding discrete orientation angle, feature responses of the corresponding angle are selected from a set of multi-directional response feature maps to generate directional response feature maps.

[0019] During the remanipulation of the direction-sensitive subkernel, the continuous local structural principal orientation angles recorded at each spatial coordinate x, y in the feature map are extracted. Dividing the angles into steps of 22.5°, all angle data are mapped to eight preset fixed discrete orientation angle indices within the half-circle angle interval [-90°, 90°], such as -90°, -67.5°, -45°, etc. These eight discrete orientation angles are then substituted into the orientation expansion relation of the Gaussian derivative-manipulated basis filter, using the corresponding sine and cosine function values ​​as analytical input parameters. Through matrix operations, eight sets of interpolation coefficient vectors corresponding to each discrete angle are obtained. For example... Figure 2 As shown, this illustrates the variation of interpolation weights of each manipulable basis filter in the direction-sensitive subkernel with the main orientation angle of the local structure, supporting the calculation logic of weighted coefficients for rotational remanipulation and angle alignment of the direction-sensitive subkernel. Using these eight sets of interpolation coefficient vectors, the static weights of each order of manipulable basis filters obtained from previous training are replaced by vector multiplication and addition combinations, generating a set of eight global direction-sensitive subkernel tensors of size 7×7, aligned with eight preset discrete orientations, at the network parameter level. In the inference computation flow, these eight global direction-sensitive subkernels are used in parallel to apply standard two-dimensional sliding convolutions to the input image or low-level input features, pre-generating multi-directional response feature maps in memory, each with 32 channels, reflecting eight different orientation response emphases. By traversing the two-dimensional spatial coordinate grid, based on the discrete orientation angle integer index mask map pre-established at each coordinate, the Gather index selection operator is called to extract the feature pixel values ​​at the corresponding angle level from the eight pre-generated feature maps, forming a 32-channel direction-sensitive feature output. The direction-sensitive feature output and the 32-channel direction-independent feature output are then gated and weighted and fused according to the coherence map to generate a directional response feature map that has the ability to respond to any curve boundary. The Gather index selection operator is a tensor selection operation that extracts corresponding elements from multiple sets of candidate feature maps or tensors according to their positions and recombines them into the target output feature based on a pre-given index mask.

[0020] S2. Locate the potential leakage center on the directional response feature map, extract the neighborhood feature map as the neighborhood feature and establish a local coordinate system, calculate the covariance matrix of the feature distribution within the neighborhood feature and determine the direction of the principal feature vector as the leakage diffusion principal axis, align the local coordinate system principal axis with the leakage diffusion principal axis by rotation and alignment, transform the aligned neighborhood feature to the polar coordinate system based on the polar coordinate mapping relationship, and obtain the radial angle feature tensor composed of feature channel, radial distance and polar angle.

[0021] The directional response feature map is subjected to absolute value summation, maximum value projection, or 1×1 convolution mapping along the channel dimension to generate a single-channel spatial activation amplitude map, which is then normalized. Threshold segmentation and connected component analysis are performed on the normalized spatial activation amplitude map to filter out isolated regions with an area smaller than a preset threshold. The centroid coordinates of the remaining connected regions are used as potential leakage centers. A planar Cartesian local coordinate system is constructed with the pixel coordinates of each potential leakage center as the spatial origin. Using tensor index slicing, a spatial square window with a length and width of 65 pixels centered at the origin of the local coordinate system is extracted from the directional response feature map, and the corresponding neighborhood feature map is extracted as the neighborhood feature. The row and column coordinates of all pixels within the extracted neighborhood features are considered as data points distributed in a two-dimensional space. The absolute values ​​of the response features of each coordinate point across all channel dimensions are summed and set as the data distribution quality weight corresponding to that coordinate point. Based on the quality weight, a 2×2 covariance matrix of the weighted two-dimensional spatial feature distribution is calculated. This covariance matrix is ​​obtained using the singular value decomposition algorithm, and the left singular vector corresponding to the largest singular value is extracted as the principal feature vector. The direction of the line containing this left singular vector is defined and determined as the leakage diffusion principal axis. The angular deviation of the leakage diffusion principal axis relative to the horizontal axis of the currently established local coordinate system is calculated, and a corresponding two-dimensional affine rotation transformation matrix is ​​constructed based on this. Batch matrix multiplication is performed to complete the two-dimensional spatial transformation of the local coordinate system to achieve rotation alignment, so that the positive horizontal axis of the aligned local coordinate system is consistent with the direction of the leakage diffusion principal axis. For each new pixel that falls within the neighborhood features after the rotation and alignment operation, the radial distance scalar from each point to the local origin is calculated using the distance formula of the Cartesian coordinate system. The `torch.atan2` operation is called in parallel to obtain the polar angle scalar of each point relative to the local origin based on its x and y coordinates, establishing a polar coordinate mapping relationship. The radial angle feature tensor is uniformly defined as follows: the channel direction represents the feature channel C, the row direction represents the radial distance r, and the column direction represents the polar angle θ. The radial distance r is sampled according to an integer radius of 0 to 32 pixels, resulting in 33 radial sampling points. The polar angle θ is divided into 64 angle sampling points according to the interval [0, 2π). Therefore, a radial angle feature tensor with a shape of C×33×64 is generated, where each feature channel corresponds to a 33-row, 64-column radial angle feature plane. The torch.nn.functional.grid_sample resampling function is used to perform the geometric mapping transformation from Cartesian coordinate space features to polar coordinate space features, generating a radial angle feature tensor composed of feature channels, radial distances, and polar angles, where each feature channel corresponds to a two-dimensional matrix form of radial angle feature plane.

[0022] In an optional embodiment, the steps of locating the potential leakage center on the directional response feature map, extracting the neighborhood feature map as the neighborhood feature and establishing a local coordinate system, calculating the covariance matrix of the feature distribution within the neighborhood feature and determining the direction of the principal feature vector as the leakage diffusion principal axis, aligning the local coordinate system principal axis with the leakage diffusion principal axis through rotation, and transforming the aligned neighborhood feature to the polar coordinate system based on the polar coordinate mapping relationship to obtain a radial angle feature tensor composed of feature channels, radial distance, and polar angle, including: Threshold segmentation and connected component analysis are performed on the directional response feature map to extract the location coordinates of potential leakage centers; A fixed-size neighborhood feature map is extracted with the potential leakage center as the origin as the neighborhood feature and a local coordinate system is established. The covariance matrix of the feature distribution within the neighborhood feature is calculated, and the direction of the main feature vector is determined as the main axis of leakage diffusion. The principal axis of the local coordinate system is aligned with the principal axis of the leakage diffusion by rotation and alignment. Based on the polar coordinate mapping relationship, the aligned neighborhood features are resampled to the polar coordinate system through bilinear interpolation to generate a radial angle feature tensor consisting of feature channels, radial distance, and polar angle, where each feature channel corresponds to a two-dimensional matrix form of radial angle feature plane.

[0023] When generating the radial angle feature tensor in polar coordinates, a global fixed threshold of 0.6 is set to perform hard thresholding on the extracted spatial activation amplitude map of the directional response feature map to obtain a binarized image. A connected component labeling algorithm is applied to filter isolated artifact connected regions with an area less than fifty pixels. The remaining connected regions are used as the x, y positions of the potential leakage center, calculated using the first-order moment operator and the pixel centroid geometric coordinates. Using this centroid coordinate as the geometric origin, a local square neighborhood feature block with a resolution of 65×65 pixels is extracted from the directional response feature map, and a local Cartesian relative coordinate system is established. The multi-channel activation intensity feature measures at each spatial pixel coordinate point within this neighborhood are statistically analyzed to construct the corresponding feature distribution spatial covariance matrix. Eigenvalue decomposition is performed to extract the eigenvector direction angle corresponding to the largest eigenvalue of the matrix. This direction angle is defined as the main seepage diffusion axis during the dispersion of the leakage source. Based on the angle between the principal axis of leakage diffusion and the X-axis of the local coordinate system, a two-dimensional affine rotation matrix with a rotation angle equal to the inverse of this angle is constructed. An inverse geometric affine transformation is then performed on the aforementioned 65×65 neighborhood features around the center point, rotating the entire local coordinate system's X-axis until it overlaps and aligns with the detected principal axis of leakage diffusion in that direction. Forward and inverse mapping equations between rectangular and polar coordinates are then established. , The sampling space for the radial distance r along the polar axis is set to a radius of 0, 1, 2 up to 32 pixels, with a total of 33 sampling points. The sampling range of the polar angle θ spans the radian interval [0, 2π) and is divided into 64 equal sampling segments with a step size of approximately 5.625°. Using a bilinear interpolation algorithm based on the pixel intensity of four neighboring known points, the aligned feature values ​​within the Cartesian system are reprojected and filled into a new grid space, generating a radial angle feature tensor of shape C×33×64, where C represents the number of channels of the neighborhood features. Each channel forms a radial angle feature plane in the form of a two-dimensional real matrix with 33 rows corresponding to the radial distance axis r and 64 columns corresponding to the polar angle axis θ.

[0024] S3. Construct a convolution module. Take the feature vector of the potential leakage center as the input of the potential leakage center feature. Generate a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor. Perform one-dimensional convolution on the radial angle feature tensor along the radial axis and polar axis respectively to identify the diffusion intensity feature of the leakage along the radial direction and the uniformity feature of the angular distribution of the leakage as the leakage diffusion characteristics. Generate leakage discrimination features. Output the location of the leakage area based on the leakage discrimination features.

[0025] A deep network layer with an attention mechanism is instantiated to construct a convolutional module. The 32-dimensional channel feature vectors corresponding to the locations of each potential leakage center on the directional response feature map are treated as potential leakage center features and input into a multilayer perceptron. The linear output layer at the end of the multilayer perceptron network generates convolutional kernel weights. Tensor deformation and segmentation operations are used to transform the output weights into independent parameter sets, generating a one-dimensional radial convolutional kernel of length 5 shared by all feature channels and a one-dimensional polar angle convolutional kernel of length 7 shared by all feature channels. The input radial angle feature tensor is received and represented as a tensor of shape C×R×Θ, where C represents the number of feature channels, R represents the number of radial sampling points, and Θ represents the number of polar angle sampling points. In this embodiment, R is 33 and Θ is 64. When performing convolution along the radial axis, the tensor is rearranged into a one-dimensional input form with different polar angle columns as intra-batch sequences. This transforms C×R×Θ into Θ sets of C×R sequences. Then, `torch.nn.functional.conv1d` is called to apply a one-dimensional radial convolution kernel of length 5 along the R dimension. Reflection padding, copy padding, or zero padding are used at both ends of the radial dimension to maintain R sampling points in the radial convolution output. This kernel is applied to each feature channel through grouped convolution or channel-wise shared convolution. When performing convolution along the polar angle axis, the tensor is rearranged into a one-dimensional input form with different radial lines as intra-batch sequences. This transforms C×R×Θ into R sets of C×Θ sequences. After annular padding at both ends of the θ dimension, `torch.nn.functional.conv1d` is called to apply a one-dimensional polar angle convolution kernel of length 7 along the θ dimension. This kernel is applied to each feature channel through grouped convolution or channel-wise shared convolution. This allows the extraction of the feature value decay pattern from the inside out, identifying the radial diffusion intensity characteristics of leakage, such as... Figure 3 As shown, the leakage characteristic response gradually decreases from the leakage center outwards with increasing pixel distance. Based on this, the symmetry of the characteristic response in each diffusion direction is further extracted, the angular distribution uniformity of the leakage is identified, and the radial diffusion intensity feature and the angular distribution uniformity feature are combined as the leakage diffusion characteristics. Figure 4 As shown, leakage characteristics may exhibit multi-peak aggregation or non-uniform distribution in the annular angular domain, which illustrates that a one-dimensional polar angle convolution kernel can extract response differences in different polar angle directions and identify the uniformity characteristics of leakage angle distribution.

[0026] The outputs of the two one-dimensional convolution operations of different dimensions are concatenated and fused along the channel dimension axis of the feature tensor to obtain a fused feature tensor. This fused feature tensor is then flattened to obtain a joint representation vector. This joint representation vector is then passed through a fully connected network module, a LeakyReLU activation function, and a preset limiting function to extract the classification boundary representation vector, generating leakage discrimination features for classification and localization regression. The classification branch is used to calculate the binary classification Softmax confidence score of the leakage discrimination features, while the regression branch outputs the relative coordinate offset, width and height dimensions, or diffusion envelope polygon parameters of the potential leakage center. When the classification confidence score exceeds a preset judgment threshold, the boundary coordinates of the leakage area are decoded based on the coordinates of the potential leakage center and the localization parameters output by the regression branch. Then, the cv2.rectangle or polygon drawing function is called to draw a bright colored frame or envelope line in the initial input original dam drone-collected image, and the location of the leakage area in the current image is output to the detection end.

[0027] In an optional embodiment, the convolutional module, using the feature vector of the potential leakage center as input, generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor, including: The feature vectors of potential leakage centers are input into the first multilayer perceptron network and the second multilayer perceptron network, respectively. The first weight vector is output using the first multilayer perceptron network, and the first weight vector is reshaped into a one-dimensional radial convolution kernel shared by each feature channel; The second weight vector is output using the second multilayer perceptron network, and the second weight vector is reshaped into a one-dimensional polar angle convolution kernel shared by each feature channel; The weights of the one-dimensional radial convolution kernel and the one-dimensional polar convolution kernel are adaptively generated based on the eigenvector input of different potential leakage centers.

[0028] When generating a one-dimensional target-aware convolutional kernel, within the network layer corresponding to the original directional response feature map, all 32-dimensional deep contextual feature channel response values ​​of the previously located potential leakage center location at coordinates x, y are extracted as the center input feature vector. This 32-dimensional one-dimensional feature tensor is independently copied into two copies and fed in parallel into a feedforward structure consisting of two independent fully connected branches: a first multilayer perceptron network and a second multilayer perceptron network containing dimensionality-reduced hidden layers and ReLU activation functions. The first and second multilayer perceptron networks have identical structures, each containing a fully connected input layer and a layer connected via a formula. The hidden layer is calculated by the feature representation and a formula. The fully connected output layer yields the result, where x is the input feature, h is the hidden layer feature, and y is the output vector. and This is the weight matrix. and The bias vector is used as the input. Both the first and second multilayer perceptron networks receive the aforementioned 32-dimensional center input feature vector. Hidden layers map the input to 32-dimensional features. Tensor forward propagation is performed using the last linear layer of the first multilayer perceptron network (without nonlinear activation), outputting a 5-length floating-point 1D first weight vector. This vector is then reshaped to a 5-length 1D radial convolution kernel shared by all feature channels. In parallel operation, the output of the second multilayer perceptron network calculates a 7-length second weight vector array. The same reshaping command is used to generate a 7-length 1D polar convolution kernel shared by all feature channels. These 1D radial and polar convolution kernels are not fixed global static kernels but are generated in real-time by the trained multilayer perceptron network based on the feature vectors of different potential leakage centers. The parameters of the multilayer perceptron network are updated through backpropagation during training, and during inference, the current target-specific convolution weight values ​​are output based on the center features.

[0029] In an optional embodiment, the step of performing one-dimensional convolution on the radial angle feature tensor along the radial axis and polar axis respectively to identify the radial diffusion intensity feature and the angular distribution uniformity feature of the leakage as leakage diffusion characteristics, generating leakage discrimination features, including: Using the generated one-dimensional radial convolution kernel, a sliding convolution is performed on the radial angle feature tensor along the radial axis, and boundary padding is used in the radial dimension to keep the convolution output the same radial sampling size as the input, thereby extracting the diffusion intensity features of leakage along the radial direction. Using the generated one-dimensional polar angle convolution kernel, a one-dimensional circular sliding convolution is performed along the polar angle axis on the radial angle feature tensor, and a circular filling is used in the polar angle dimension to eliminate the influence of the boundary break at 0 degrees and 360 degrees, thereby extracting the angular distribution uniformity feature of leakage. The extracted diffusion intensity features and angular distribution uniformity features are concatenated in the channel dimension, and feature aggregation and compression are performed through a fully connected layer to generate leakage discrimination features for outputting the location and extent of leakage areas.

[0030] When generating comprehensive leakage discrimination features to determine the range boundary, a one-dimensional radial convolution kernel generated based on the target features in the previous process is retrieved. The kernel size is 5, and the kernel has a preset step size of one pixel. In the C×33×64 radial angle feature tensor unique to the detection area, convolution scanning is performed along the 33×64 radial angle feature plane corresponding to each feature channel. A one-dimensional linear convolution sliding scan is applied along the row direction representing the radius distance amplification, i.e., the radial axis r. The pixel intensity continuous decay gradient and jump information when the leakage response amplitude diffuses from the center origin to the surrounding edge are extracted and calculated, which represents the diffusion intensity feature matrix that expands outward. Instead, another extracted one-dimensional polar convolution kernel, with a size of 7, is used to perform a one-dimensional sliding scan along the polar axis θ of the radial angle feature plane corresponding to each feature channel in the same radial angle feature tensor. In order to remove the cutting effect error of the polar coordinate 0° and 360° breakpoints, a ring-shaped cyclic boundary filling strategy is adopted on both sides of the polar convolution scan dimension to collect the phase dispersion and directional smoothness of the diffusion concentration of the leakage medium on a specific radius circle, and to extract the angular distribution uniformity feature matrix.

[0031] The C×33×64 diffusion intensity feature tensor output by radial convolution and the C×33×64 angular uniformity feature tensor output by polar convolution are concatenated along the feature channel dimension to form a fused feature tensor with 2C channels. This 3D tensor is flattened to transform it into a one-dimensional continuous vector sequence. This sequence is then serially input into a multilayer fully connected classification and regression network to perform high-order nonlinear aggregation mapping and dimensionality reduction data compression mechanisms. Simultaneously, at the output node, the leakage authenticity classification result containing Softmax classification confidence is output, along with the normalized two-dimensional center pixel relative coordinates corresponding to the leakage surface and the regression offset data location parameter set of the associated horizontal length, width, or diffusion envelope polygon. The structure of the multilayer fully connected classification and regression network includes a fully connected input layer, a first hidden layer with 512 nodes, a second hidden layer with 256 nodes, and two parallel fully connected output branches for classification and regression, respectively. The input to the classification and regression network is the aforementioned stacked and flattened one-dimensional continuous vector sequence; the output of the classification and regression network includes a target authenticity leakage attribute category discrimination feature output by a classification branch, and a regression offset data location parameter set containing normalized two-dimensional center pixel relative coordinates and size position information output by a regression branch. The probability prediction result of the classification branch is expressed by the formula... The value is calculated, where z is the inactive logical value of the classification output layer. Valid candidate regions are selected based on the confidence level of the leakage category output by the classification branch, and the leakage region boundary is decoded based on the location parameter set output by the regression branch, thus forming a complete leakage region location output process.

[0032] During the training phase, classification and location labels are generated using the bounding boxes or diffusion envelope polygons of the labeled seepage areas. The classification branch uses the cross-entropy loss function to calculate the error between true and false categories of seepage, while the regression branch uses smoothed L1 loss, IoU loss, or a weighted combination of both to calculate the location error of the region. The classification loss and regression loss are weighted and summed to obtain the joint training loss, and the parameters of the orientation-sensitive kernel, orientation-independent kernel, multilayer perceptron network, and classification and regression branches are updated through backpropagation, enabling the model to perform end-to-end detection of seepage areas in UAV dam images after training.

[0033] In the second embodiment, the present invention also proposes a dam seepage detection system based on UAV vision, comprising the following modules: The module is used to acquire images of the dam collected by UAVs, construct a hybrid convolutional kernel composed of orientation-sensitive and orientation-independent kernels, generate a coherence map and local structure principal orientation angle based on the local gradient covariance matrix, extract image features using the hybrid convolutional kernel, perform weighted fusion of the outputs of the orientation-sensitive and orientation-independent kernels using the coherence map as a gating signal, and remanufacture the orientation-sensitive kernel using the local structure principal orientation angle to align the orientation-sensitive kernel with the extension direction of the local structure, thereby generating an orientation response feature map. The calculation module is used to locate potential leakage centers on the directional response feature map, extract neighborhood feature maps as neighborhood features and establish a local coordinate system, calculate the covariance matrix of feature distribution within the neighborhood features and determine the direction of the principal feature vector as the leakage diffusion principal axis, align the principal axis of the local coordinate system with the leakage diffusion principal axis through rotation and alignment, and transform the aligned neighborhood features to the polar coordinate system based on the polar coordinate mapping relationship to obtain the radial angle feature tensor composed of feature channels, radial distance and polar angle; The output module is used to construct the convolution module. It takes the feature vector of the potential leakage center as the input of the potential leakage center feature and generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor. One-dimensional convolution is performed on the radial angle feature tensor along the radial axis and polar axis respectively to identify the diffusion intensity feature of the leakage along the radial direction and the uniformity feature of the angular distribution of the leakage as the leakage diffusion characteristics, generate leakage discrimination features, and output the location of the leakage area based on the leakage discrimination features.

[0034] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0035] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for detecting dam seepage based on UAV vision, characterized in that, Includes the following steps: Images of the dam collected by a drone are acquired, and a hybrid convolutional kernel consisting of a direction-sensitive kernel and a direction-independent kernel is constructed. A coherence map and a principal orientation angle of the local structure are generated based on the local gradient covariance matrix. Image features are extracted using the hybrid convolutional kernel. The outputs of the direction-sensitive kernel and the direction-independent kernel are weighted and fused using the coherence map as a gating signal. The principal orientation angle of the local structure is used to remanipulate the direction-sensitive kernel so that it is aligned with the extension direction of the local structure, thereby generating a directional response feature map. The potential leakage center is located on the directional response feature map. The neighborhood feature map is extracted as the neighborhood feature and a local coordinate system is established. The covariance matrix of the feature distribution within the neighborhood feature is calculated and the direction of the principal feature vector is determined as the leakage diffusion principal axis. The local coordinate system principal axis is aligned with the leakage diffusion principal axis by rotation alignment. The aligned neighborhood feature is transformed to the polar coordinate system based on the polar coordinate mapping relationship to obtain the radial angle feature tensor composed of feature channel, radial distance and polar angle. A convolution module is constructed, which takes the feature vector of the potential leakage center as the input of the potential leakage center feature. It generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor. One-dimensional convolution is performed on the radial angle feature tensor along the radial axis and polar axis respectively to identify the diffusion intensity feature of the leakage along the radial direction and the uniformity feature of the angular distribution of the leakage as the leakage diffusion characteristics, generating leakage discrimination features. Based on the leakage discrimination features, the location of the leakage area is output.

2. The method according to claim 1, characterized in that, The process of acquiring images of the dam collected by the drone and constructing a hybrid convolution kernel composed of orientation-sensitive and orientation-independent sub-kernels includes: The Gaussian derivative function is used as a steerable basis filter with a direction-sensitive subkernel, and learnable weight parameters are assigned to steerable basis filters of each order. A two-dimensional orientation-sensitive subnucleus is generated by linearly weighting and summing the manipulated basis filters based on the weight parameters. Initialize a two-dimensional matrix network layer containing learnable parameters as a direction-independent sub-kernel; Along the output channel dimension, the convolution weights corresponding to the direction-sensitive sub-kernel and the convolution weights corresponding to the direction-independent sub-kernel are concatenated and fused to construct a hybrid convolution kernel weight tensor containing both direction-sensitive and direction-independent output channels.

3. The method according to claim 2, characterized in that, The generation of coherence maps and local structural principal orientation angles based on the local gradient covariance matrix includes: The Sobel operator is used to calculate the horizontal and vertical gradients of each pixel in the dam image. The local gradient covariance matrix of each pixel is constructed using the squared terms and product terms of the horizontal and vertical gradients, and the matrix is ​​then subjected to Gaussian smoothing filtering. Calculate the eigenvalues ​​and eigenvectors of the smoothed local gradient covariance matrix, and use the normalized difference between the maximum and minimum eigenvalues ​​as the coherence map values ​​of the pixel. The orientation angle of the feature vector corresponding to the minimum feature value is taken as the local structural principal orientation angle of the pixel.

4. The method according to claim 1 or 2, characterized in that, The weighted fusion of the outputs of the direction-sensitive sub-core and the direction-independent sub-core, using the coherence graph as a gating signal, includes: The pixel values ​​of the coherence map are normalized to a preset range and used as the gate weight coefficients. The image features extracted using hybrid convolution kernels are decoupled along the channel dimension into orientation-sensitive feature outputs and orientation-independent feature outputs with the same number of channels. The orientation-sensitive feature output is weighted by pixel-by-pixel multiplication using gating weight coefficients; The difference between the value 1 and the gate weight coefficient is used as a complementary coefficient to perform pixel-by-pixel multiplication and weighting on the direction-independent feature output; The weighted direction-sensitive feature output and the direction-independent feature output are added element-wise to generate a fused feature map that covers the entire array.

5. The method according to claim 1, characterized in that, The method of remanipulating the orientation-sensitive sub-core using the principal orientation angle of the local structure to align the orientation-sensitive sub-core with the extension direction of the local structure and generate an orientation response feature map includes: Obtain the local structural main orientation angle corresponding to the extraction of spatial location features, and quantize the orientation angle into a preset number of discrete orientation angles; Each discrete direction angle is used as an analytical parameter, and a set of interpolation coefficients corresponding to each angle are calculated. The linear combination weights of the manipulable basis filters in the direction-sensitive sub-kernel are updated using each set of interpolation coefficients to generate a set of global direction-sensitive sub-kernels aligned with each discrete direction angle. A set of global orientation-sensitive sub-kernels are used to perform global convolution calculations to pre-generate a set of multi-directional response feature maps; At each spatial location, based on the corresponding discrete orientation angle, feature responses of the corresponding angle are selected from a set of multi-directional response feature maps to generate directional response feature maps.

6. The method according to claim 1, characterized in that, The process involves locating potential leakage centers on the directional response feature map, extracting neighborhood feature maps as neighborhood features and establishing a local coordinate system, calculating the covariance matrix of the feature distribution within the neighborhood features, determining the direction of the principal feature vector as the leakage diffusion principal axis, aligning the local coordinate system principal axis with the leakage diffusion principal axis through rotation, and transforming the aligned neighborhood features to a polar coordinate system based on a polar coordinate mapping relationship to obtain a radial angle feature tensor composed of feature channels, radial distances, and polar angles, including: Threshold segmentation and connected component analysis are performed on the directional response feature map to extract the location coordinates of potential leakage centers; A fixed-size neighborhood feature map is extracted with the potential leakage center as the origin as the neighborhood feature and a local coordinate system is established. The covariance matrix of the feature distribution within the neighborhood feature is calculated, and the direction of the main feature vector is determined as the main axis of leakage diffusion. The principal axis of the local coordinate system is aligned with the principal axis of the leakage diffusion by rotation and alignment. Based on the polar coordinate mapping relationship, the aligned neighborhood features are resampled to the polar coordinate system through bilinear interpolation to generate a radial angle feature tensor consisting of feature channels, radial distance, and polar angle, where each feature channel corresponds to a two-dimensional matrix form of radial angle feature plane.

7. The method according to claim 1, characterized in that, The convolutional module, using the feature vector of the potential leakage center as input, generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor, including: The feature vectors of potential leakage centers are input into the first multilayer perceptron network and the second multilayer perceptron network, respectively. The first weight vector is output using the first multilayer perceptron network, and the first weight vector is reshaped into a one-dimensional radial convolution kernel shared by each feature channel; The second weight vector is output using the second multilayer perceptron network, and the second weight vector is reshaped into a one-dimensional polar angle convolution kernel shared by each feature channel; The weights of the one-dimensional radial convolution kernel and the one-dimensional polar convolution kernel are adaptively generated according to the feature vector input of different potential leakage centers.

8. The method according to claim 1, characterized in that, The method involves performing one-dimensional convolutions along the radial axis and polar axis on the radial angle feature tensor to identify the radial diffusion intensity characteristics and the angular distribution uniformity characteristics of the leakage as leakage diffusion characteristics, generating leakage discrimination features, including: Using the generated one-dimensional radial convolution kernel, a sliding convolution is performed on the radial angle feature tensor along the radial axis, and boundary padding is used in the radial dimension to keep the convolution output the same radial sampling size as the input, thereby extracting the diffusion intensity features of leakage along the radial direction. Using the generated one-dimensional polar angle convolution kernel, a one-dimensional circular sliding convolution is performed along the polar angle axis on the radial angle feature tensor, and a circular filling is used in the polar angle dimension to eliminate the influence of the boundary break at 0 degrees and 360 degrees, thereby extracting the angular distribution uniformity features of leakage. The extracted diffusion intensity features and angular distribution uniformity features are concatenated in the channel dimension, and feature aggregation and compression are performed through a fully connected layer to generate leakage discrimination features for outputting the location and extent of leakage areas.

9. A dam seepage detection system based on UAV vision, characterized in that, Includes the following modules: The module is used to acquire images of the dam collected by UAVs, construct a hybrid convolutional kernel composed of orientation-sensitive and orientation-independent kernels, generate a coherence map and local structure principal orientation angle based on the local gradient covariance matrix, extract image features using the hybrid convolutional kernel, perform weighted fusion of the outputs of the orientation-sensitive and orientation-independent kernels using the coherence map as a gating signal, and remanufacture the orientation-sensitive kernel using the local structure principal orientation angle to align the orientation-sensitive kernel with the extension direction of the local structure, thereby generating an orientation response feature map. The calculation module is used to locate potential leakage centers on the directional response feature map, extract neighborhood feature maps as neighborhood features and establish a local coordinate system, calculate the covariance matrix of feature distribution within the neighborhood features and determine the direction of the principal feature vector as the leakage diffusion principal axis, align the principal axis of the local coordinate system with the leakage diffusion principal axis through rotation and alignment, and transform the aligned neighborhood features to the polar coordinate system based on the polar coordinate mapping relationship to obtain the radial angle feature tensor composed of feature channels, radial distance and polar angle; The output module is used to construct the convolution module. It takes the feature vector of the potential leakage center as the input of the potential leakage center feature and generates a one-dimensional radial convolution kernel and a one-dimensional polar convolution kernel shared by each feature channel of the radial angle feature tensor. One-dimensional convolution is performed on the radial angle feature tensor along the radial axis and polar axis respectively to identify the diffusion intensity feature of the leakage along the radial direction and the uniformity feature of the angular distribution of the leakage as the leakage diffusion characteristics, generate leakage discrimination features, and output the location of the leakage area based on the leakage discrimination features.

10. The system according to claim 9, characterized in that, The process of acquiring images of the dam collected by the drone and constructing a hybrid convolution kernel composed of orientation-sensitive and orientation-independent sub-kernels includes: The Gaussian derivative function is used as a steerable basis filter with a direction-sensitive subkernel, and learnable weight parameters are assigned to steerable basis filters of each order. A two-dimensional orientation-sensitive subnucleus is generated by linearly weighting and summing the manipulated basis filters based on the weight parameters. Initialize a two-dimensional matrix network layer containing learnable parameters as a direction-independent sub-kernel; Along the output channel dimension, the convolution weights corresponding to the direction-sensitive sub-kernel and the convolution weights corresponding to the direction-independent sub-kernel are concatenated and fused to construct a hybrid convolution kernel weight tensor containing both direction-sensitive and direction-independent output channels.

Images