Method and system for registering electric wires among aerial images of unmanned aerial vehicle based on deep learning

Abstract

The invention discloses a method and a system for registering electric wires among aerial images of an unmanned aerial vehicle based on deep learning. The method comprises the following steps: carrying out adaptive image enhancement and dynamic range compression on an original aerial image; constructing a nonlinear multi-scale feature pyramid and extracting enhanced wire features; position perception features are obtained through spatial perception self-attention processing; obtaining consistency enhancement features by using a multi-layer mutual attention network and bidirectional consistency check; generating and optimizing a matching matrix based on multiple constraints; and finally, obtaining a final matching result through reliability scoring and threshold screening. According to the invention, through multi-stage adaptive feature extraction and matching optimization, the problems of accuracy and robustness of wire registration in aerial images are effectively solved, and the method is especially suitable for intelligent routing inspection of power lines in complex scenes.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and artificial intelligence, and in particular to a method and system for wire registration between drone aerial images based on deep learning. Background Art

[0002] In recent years, with the rapid economic development and the acceleration of urbanization, the demand for electricity has increased dramatically, which has put forward higher requirements for the construction and maintenance of power grids. In the construction of power grids, electric wires are the hub connecting the power industry and users, and the safety of power lines is of vital importance. However, power lines span a vast area and often pass through complex terrain and environment, which brings huge challenges to the construction, monitoring and maintenance of lines. The traditional manual inspection method is not only time-consuming and laborious, but also has safety hazards in bad weather or remote areas, and can no longer meet the needs of modern power grid management. The safe and stable operation of the power system is directly related to the development of the national economy and the stability of social life, and the intelligent inspection of transmission lines is an important means to ensure the safe operation of the power grid.

[0003] Against this backdrop, drone technology has emerged, showing great potential in the power grid industry with its flexibility and efficiency. Drones can quickly fly over complex terrain, take high-precision aerial photos of power lines, and collect a large amount of image data, providing an unprecedented perspective for the risk assessment of potential threats posed by elements around the power lines to the safe operation of power lines. This approach not only greatly improves inspection efficiency and reduces labor costs, but also enables more comprehensive and clear line status information to be obtained. However, since drones shoot the same transmission line section at different times and angles, how to accurately achieve automatic alignment of wires between different images, and then achieve dynamic monitoring and change analysis of wire status, has become a key scientific problem that needs to be solved urgently. Accurate wire alignment is not only the basis for realizing intelligent inspection of transmission lines, but also an important prerequisite for subsequent fault diagnosis such as insulator shedding, hardware corrosion, and line icing.

[0004] At present, the research on wire registration is mainly focused on methods based on traditional image processing and machine learning. Traditional methods often use feature operators such as SIFT and SURF to extract local feature points, combined with algorithms such as RANSAC for feature matching and geometric verification. Some researchers have also proposed using Hough transform to detect straight line features, or using local binary patterns (LBP) to describe wire texture features. In terms of machine learning, some methods use support vector machines (SVM) or random forests to classify the extracted features to improve the accuracy of matching. At the same time, some researchers have tried to use graph matching-based methods to model the wires as graph structures and achieve registration through graph matching algorithms. These methods can achieve certain results under ideal conditions, but there are still many limitations in actual complex scenes.

[0005] Specifically, the existing methods have the following technical problems: First, traditional feature extraction methods are difficult to cope with the complex lighting changes in drone aerial images, especially under conditions such as backlight and strong light, and are prone to insufficient or mis-extraction of feature points; Second, the existing feature description methods fail to fully consider the slender structural characteristics of wires, resulting in a lack of effective description of the geometric shape of wires in the extracted features, especially when the wires are twisted or deformed, the distinguishability of feature descriptions is significantly reduced; Third, the information fusion and analysis of wires between multiple images, especially when the wires perform differently at different angles and lighting conditions, how to accurately identify and match wires has become a key problem to be solved. Finally, traditional image processing methods, such as matching algorithms based on feature points, when dealing with slender and continuous targets such as wires, most of them focus on matching single key points, which often have poor results. These methods ignore the importance of the coherence and spatial layout of the wire as a whole. At present, the matching of wires between images is essentially the matching of a group of points, which pays more attention to the similarities and differences between different groups, and is more difficult than the matching of a single point. These technical problems seriously restrict the application effect of wire registration algorithms in practical engineering. Summary of the invention

[0006] The purpose of the invention is to propose a method and system for wire registration between drone aerial images based on deep learning to solve the above-mentioned problems existing in the prior art.

[0007] The technical solution is a method for wire registration between UAV aerial images based on deep learning, which includes the following steps:

[0008] S1. Obtain at least two original drone aerial images containing overlapping areas, and perform adaptive dynamic range compression processing on them to obtain preprocessed images; based on the preprocessed images, construct a nonlinear multi-scale feature pyramid, extract the wire features of each scale layer, and obtain an enhanced wire feature set; based on the enhanced wire feature set, construct a fusion feature matrix through dynamic feature fusion;

[0009] S2. Based on the fused feature matrix, a position-aware feature matrix is ​​obtained through space-aware self-attention processing; based on the position-aware feature matrix, an enhanced self-attention score is calculated to obtain a fused feature vector between different wires;

[0010] S3. Based on the fused feature vector, a multi-layer mutual attention network is constructed to obtain a multi-layer attention weight set; based on the multi-layer attention weight set, feature aggregation is performed to obtain cross-image fusion features; based on the cross-image fusion features, a consistency enhancement feature is obtained through a bidirectional consistency check;

[0011] S4. Based on the consistency enhancement feature, a multi-constraint matching matrix is ​​constructed to obtain an initial matching matrix; the initial matching matrix is ​​optimized to obtain an optimized matching matrix;

[0012] S5. Based on the optimized matching matrix, the matching reliability score is calculated to obtain a reliability scoring matrix; based on the reliability scoring matrix and a preset threshold, the final matching result is screened.

[0013] The deep learning-based wire registration system between drone aerial images includes:

[0014] at least one processor; and,

[0015] a memory communicatively connected to at least one of the processors; wherein,

[0016] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the deep learning-based wire registration method between drone aerial images.

[0017] Beneficial effects: The present invention can effectively handle various complex situations in aerial photography scenes, such as lighting changes, perspective differences, local deformation, etc., while also improving the accuracy of wire detection and alignment, reducing computational complexity, and improving processing efficiency, thereby achieving high-precision automatic alignment of wires between drone aerial images. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the method of the present invention.

[0019] Figure 2 This is a flow chart of step S1 of the present invention.

[0020] Figure 3 This is a flow chart of step S2 of the present invention.

[0021] Figure 4 This is a flow chart of step S3 of the present invention.

[0022] Figure 5 This is a flow chart of step S4 of the present invention.

[0023] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0024] like Figure 1 As shown, this application proposes a method for wire registration between drone aerial images based on deep learning, comprising the following steps:

[0025] S1. Obtain at least two original drone aerial images containing overlapping areas, and perform adaptive dynamic range compression processing on them to obtain preprocessed images; based on the preprocessed images, construct a nonlinear multi-scale feature pyramid, extract the wire features of each scale layer, and obtain an enhanced wire feature set; based on the enhanced wire feature set, construct a fusion feature matrix through dynamic feature fusion;

[0026] S2. Based on the fused feature matrix, a position-aware feature matrix is ​​obtained through space-aware self-attention processing; based on the position-aware feature matrix, an enhanced self-attention score is calculated to obtain a fused feature vector between different wires;

[0027] S3. Based on the fused feature vector, a multi-layer mutual attention network is constructed to obtain a multi-layer attention weight set; based on the multi-layer attention weight set, feature aggregation is performed to obtain cross-image fusion features; based on the cross-image fusion features, a consistency enhancement feature is obtained through a bidirectional consistency check;

[0028] S4. Based on the consistency enhancement feature, a multi-constraint matching matrix is ​​constructed to obtain an initial matching matrix; the initial matching matrix is ​​optimized to obtain an optimized matching matrix;

[0029] S5. Based on the optimized matching matrix, the matching reliability score is calculated to obtain a reliability scoring matrix; based on the reliability scoring matrix and a preset threshold, the final matching result is screened.

[0030] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0031] S11, obtaining at least two original drone aerial images containing overlapping areas; evaluating the image quality of the original drone aerial images to obtain an image quality feature vector; performing adaptive enhancement based on the image quality feature vector to obtain an enhanced image; performing illumination balancing and noise suppression based on the enhanced image to obtain a balanced image; performing edge feature enhancement based on the balanced image to obtain an edge enhanced image; performing adaptive dynamic range compression based on the edge enhanced image to obtain a preprocessed image;

[0032] S12. Based on the preprocessed image, construct an improved Laplacian pyramid; based on the Laplacian pyramid, perform compression operation on each layer to obtain a compressed feature map; based on the compressed feature map, perform pyramid decomposition operation to obtain a Laplacian feature map; based on the Laplacian feature map, generate a multi-scale feature pyramid;

[0033] S13. Based on the multi-scale feature pyramid, extract local structural features to obtain a structural feature matrix; based on the structural feature matrix, calculate the directional consistency features to obtain a directional feature set; based on the directional feature set, extract and fuse context information to obtain context enhancement features; based on the context enhancement features, integrate to obtain a multi-dimensional feature tensor, and finally generate an enhanced wire feature set;

[0034] S14. Calculate the dynamic fusion weight based on the enhanced wire feature set; perform a weighted sum operation based on the dynamic fusion weight to obtain a fusion feature matrix.

[0035] In one embodiment of the present application, two original drone aerial images IA and IB containing overlapping areas are acquired and enhanced to obtain edge-enhanced images, and an adaptive dynamic range compression operation is performed on each pixel value x: y = α·log(1+βx) / log(1+β), where α is the compression coefficient, β is the dynamic adjustment factor, and y is the pixel value after the adaptive dynamic range compression operation, to obtain preprocessed images IA' and IB'.

[0036] Get the preprocessed images IA' and IB', construct an improved Laplacian pyramid, perform compression operation on each layer k to obtain Gk+1=x·tanh(γx), where γ is a learnable parameter and Gk+1 is the feature map after compression operation on the k+1th layer of the Laplacian pyramid; perform pyramid decomposition operation to obtain Lk = Gk - Expand(Gk+1), where Expand is an upsampling expansion operation, and finally obtain multi-scale feature pyramids PA and PB.

[0037] Get multi-scale feature pyramids PA and PB, extract wire features at each scale layer, and for the M wires LA in image IA and the N wires LB in image IB, represent each wire as an n ordered point set Li = {pi1, pi2, ..., pin}, where each point contains spatial coordinates (x, y), scale response s, and directional feature o, to obtain enhanced wire feature sets FA and FB.

[0038] Obtain the enhanced wire feature sets FA and FB, calculate the dynamic fusion weight wi = softmax(νi·tanh(μFi)), where νi and μ are learnable parameters, and Fi represents the feature vector extracted from the enhanced wire feature set; perform the weighted sum operation Z = Σ(wi·Fi) to obtain the fusion feature matrix Z with a dimension of M×N×D.

[0039] This embodiment achieves high-quality extraction and enhancement of wire features in drone aerial images through multi-stage adaptive image preprocessing and nonlinear multi-scale feature extraction. Specifically, the adaptive adjustment mechanism of the compression coefficient and dynamic adjustment factor in the adaptive dynamic range compression operation enables the image to obtain the best dynamic range display under different lighting and contrast conditions; the improved Laplacian pyramid enhances the feature retention ability of slender targets such as wires by introducing compression operations with learnable parameters, avoiding the problem of easy loss of details in traditional pyramid decomposition; in the wire feature extraction process, by representing each wire as an ordered point set containing spatial coordinates, scale response and directional features, the geometric features of the wire are retained and its local morphological information is captured; finally, through the design of dynamic fusion weights, adaptive integration of different features is achieved. This embodiment can accurately extract and enhance wire features in complex aerial photography scenes, lay a solid foundation for subsequent feature matching, and improve the accuracy and robustness of subsequent registration.

[0040] According to one aspect of the present application, step S11 is further:

[0041] S111, obtaining at least two original drone aerial images containing overlapping areas, calculating image clarity features, illumination uniformity features, and signal-to-noise ratio features; combining the image clarity features, illumination uniformity features, and signal-to-noise ratio features to obtain an image quality feature vector;

[0042] S112, constructing an adaptive enhancement function based on the original drone aerial image and its corresponding image quality feature vector; performing an enhancement operation on each pixel based on the adaptive enhancement function to obtain an enhanced image;

[0043] S113, constructing an adaptive weight matrix based on the enhanced image, performing weighted smoothing operation to obtain a smoothed image; performing local contrast correction based on the smoothed image to obtain a balanced image;

[0044] S114, based on the balanced image, calculating multi-directional gradient responses and constructing a directional enhancement kernel; based on the directional enhancement kernel, performing a directional selective convolution operation to obtain an edge enhanced image;

[0045] S115. Based on the edge-enhanced image, calculate local statistical features and construct an adaptive compression coefficient; based on the adaptive compression coefficient, calculate a dynamic adjustment factor; based on the edge-enhanced image and the dynamic adjustment factor, perform a pixel-level adaptive compression operation to obtain a preprocessed image.

[0046] In one embodiment of the present application, two original drone aerial images IA and IB containing overlapping areas are obtained, and the image clarity feature fc = Σ|▽I(x, y)| / N is calculated, where ▽I(x, y) is the image gradient and N is the total number of pixels; the illumination uniformity feature fl = std(I) / mean(I) is calculated, where std is the standard deviation operator and mean is the mean operator; the signal-to-noise ratio feature fn = 10·log10(P_signal / P_noise) is calculated, where P_signal is the signal power and P_noise is the noise power; the three features are combined to obtain an image quality feature vector Q = [fc, fl, fn].

[0047] The original UAV aerial images IA and IB and their corresponding image quality feature vectors Q are obtained, and an adaptive enhancement function E(x) = x +λ·(1-Q)·sinh(γx) is constructed, where x is the input pixel value, λ is the enhancement intensity coefficient, which is dynamically adjusted by the clarity feature fc, and γ is the nonlinear coefficient, which is adjusted by the illumination uniformity feature fl; the enhancement operation is performed on each pixel to obtain the enhanced images IE_A and IE_B.

[0048] Get enhanced images IE_A and IE_B, and construct the adaptive weight matrix W(x, y) = exp(-||I(x, y)-μ||2 / 2σ 2 ), where I(x, y) is the pixel value, μ is the local mean, σ is adjusted by the signal-to-noise ratio feature fn, and || || 2 represents the Euclidean norm; perform weighted smoothing operation Is(x, y) =ΣW(i, j)·I(i, j) / ΣW(i, j), where (i, j) are the local window coordinates; apply local contrast correction Ic(x, y) = Is(x, y)·(μg / μl)η, where μg is the global mean, μl is the local mean, and η is the adaptive adjustment factor, and obtain balanced images IB_A and IB_B.

[0049] Get the balanced images IB_A and IB_B, calculate the multi-directional gradient response Gθ(x, y) = cos(θ)·ΨI / Ψx + sin(θ)·ΨI / Ψy, where θ is the gradient direction angle and Ψ is the partial derivative; construct the directional enhancement kernel Kθ = exp(-d 2 / 2ρ 2 )·cos(θ), where d is the spatial distance, ρ is the spatial scale parameter, and exp is the exponential function; perform direction-selective convolution operation Ie(x, y) = max_θ{I(x, y) * Kθ}, where * represents the convolution operation, and obtain edge-enhanced images IEE_A and IEE_B.

[0050] Obtain edge enhanced images IEE_A and IEE_B, calculate local statistical features μl(x, y) and σl(x, y); construct adaptive compression coefficient α(x, y) = α0·(1 + k·σl(x, y) / σg), where α0 is the basic compression coefficient, k is the adjustment coefficient, and σg is the global standard deviation; calculate dynamic adjustment factor β(x, y) = β0·exp(-|μl(x, y) - μg| / σg), where β0 is the basic adjustment factor; perform pixel-level adaptive compression operation y(x, y) = α(x, y)·log(1 + β(x, y)·x(x, y)) / log(1 + β(x, y)) to obtain preprocessed images IA' and IB'.

[0051] This embodiment achieves high-quality optimization of aerial images in complex scenes through multi-level adaptive image enhancement and preprocessing strategies. A complete image quality assessment system is constructed by calculating image clarity features, illumination uniformity features, and signal-to-noise ratio features; based on these features, an adaptive enhancement function is designed to achieve targeted image enhancement; local weighted smoothing and contrast correction are performed through an adaptive weight matrix to effectively suppress noise while maintaining edge details; finally, the optimal adjustment of the dynamic range is achieved through pixel-level adaptive compression operations. This embodiment improves the quality of aerial images, especially when dealing with complex scenes such as uneven illumination, low contrast, and noise interference, and shows excellent adaptability and stability.

[0052] According to one aspect of the present application, step S13 is further:

[0053] S131, based on the multi-scale feature pyramid, construct a structure tensor at each scale layer, calculate the eigenvalues ​​and eigenvectors; based on the eigenvalues ​​and eigenvectors, construct a local structure descriptor; based on the local structure descriptor, perform feature aggregation on each detection point to obtain a structure feature matrix;

[0054] S132. Based on the structural feature matrix, calculate the direction vector field and construct a direction consistency measure; based on the direction consistency measure, perform non-maximum suppression to obtain the local main direction; based on the local main direction, use a tensor voting mechanism to enhance the direction coherence and obtain a direction feature set;

[0055] S133, based on the directional feature set, construct an adaptive sampling window to extract local context features; based on the local context features, apply non-local mean filtering to obtain context enhancement features;

[0056] S134, constructing a multidimensional feature descriptor based on the context enhancement feature, and calculating the feature importance weight; performing weighted feature fusion based on the feature importance weight to obtain a multidimensional feature tensor;

[0057] S135. Based on the multidimensional feature tensor, a predetermined number of key points are extracted for each wire, and a feature representation of each key point is calculated; based on the feature representation of each key point, feature normalization and dimensionality reduction are performed to obtain an enhanced wire feature set.

[0058] In one embodiment of the present application, multi-scale feature pyramids PA and PB are obtained, and a structure tensor T(x, y) = [Ix 2 , IxIy; IxIy, Iy 2 ], where Ix and Iy are image gradients; calculate eigenvalues ​​λ1, λ2 and eigenvectors v1, v2; construct a local structure descriptor s(x, y) = [λ1, λ2, v1·v2, trace(T)], where trace is the trace of the matrix; perform feature aggregation on each detection point to obtain the structural feature matrices SMA and SMB.

[0059] Get the structural feature matrices SMA and SMB, calculate the directional vector field ν(x, y) = λ1v1, where λ1 is the main eigenvalue and v1 is the main eigenvector; construct the directional consistency measure c(x, y) = Σw(i, j)·|ν(x, y)·ν(i, j)| / ||ν(x, y)||·||ν(i, j)||, where w(i, j) is the spatial weight; perform non-maximum suppression to obtain the local main direction; use the tensor voting mechanism to enhance the directional coherence and obtain the directional feature sets DA and DB.

[0060] Get the directional feature sets DA and DB, and construct an adaptive sampling window W(x, y) = exp(-d 2 / 2σ 2 )·(1 + κ·c(x, y)), where d is the spatial distance, σ is the scale parameter, and κ is the direction consistency adjustment factor; extract the local context feature h(x, y) = Σ(W(i, j)·f(i, j)), where f(i, j) is the original feature; apply non-local mean filtering to enhance feature consistency and obtain context enhanced features CA and CB.

[0061] Obtain context-enhanced features CA and CB, construct a multidimensional feature descriptor m(x, y) = [s(x, y), d(x, y), h(x, y)], where s is the structural feature, d is the directional feature, and h is the context feature; calculate the feature importance weight w = softmax(φ(m)), where φ is the feature mapping function; perform weighted feature fusion to obtain multidimensional feature tensors MA and MB.

[0062] Obtain multidimensional feature tensors MA and MB, extract n key points {pi1, pi2, ..., pin} for each wire Li; calculate the feature representation fij = [xij, yij, sij, oij, mij] of each point, where xij, yij are spatial coordinates, sij is the scale response, oij is the directional feature, and mij is the multidimensional feature; perform feature normalization and dimensionality reduction to obtain enhanced wire feature sets FA and FB.

[0063] This embodiment realizes the all-round characterization and enhancement of wire features based on multi-dimensional feature extraction and fusion. Local structural features are extracted through structural tensor analysis, eigenvalues ​​and eigenvectors are calculated, and a complete structural descriptor is constructed; directional coherence is enhanced through directional vector field and tensor voting mechanism, and the continuity characteristics of wires are effectively captured; context features are extracted using adaptive sampling windows, and feature consistency is enhanced through non-local mean filtering; finally, a comprehensive feature representation containing spatial position, scale response, directional features and structural information is generated through multi-dimensional feature fusion and normalization processing. This embodiment not only improves the discriminability and robustness of features, but also shows strong adaptability when dealing with abnormal situations such as wire breakage and deformation.

[0064] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0065] S21, obtaining a wire feature set of the same image in the fusion feature matrix, and calculating a position code; based on the position code, performing a feature transformation operation to obtain a position-aware feature matrix;

[0066] S22. Based on the position-aware feature matrix, a spatial constraint matrix is ​​constructed, the attention score is calculated, and a fused feature vector is obtained.

[0067] In one embodiment of the present application, the wire feature set Li of the same image in the fusion feature matrix Z is obtained, and the position encoding PE(Li) = [sin(ωpdi), cos(ωpdi)] is calculated, where di is the spatial distribution feature of the wire, and ωp is the position encoding frequency parameter; feature transformation operations K = WKf(Li) + PE(Li), Q = WQf(Li) + PE(Li) and V = WVf(Li) are performed, where WK, WQ, WV are learnable transformation matrices, f(Li) is the feature representation of the wire Li, and the position-aware feature matrix F is obtained.

[0068] Get the position-aware feature matrix F and construct the spatial constraint matrix Mij = exp(-||pi - pj||2 / σ 2), where pi and pj are the coordinates of the wire endpoints and σ is the spatial scale parameter; calculate the attention score A(Q, K, V) = softmax(QKT / sqrt(d) + M) V, where d is the feature dimension and M is the spatial constraint matrix, and get the fused feature vector f'.

[0069] This embodiment realizes the position perception and contextual relationship modeling of the wire features through the spatial perception self-attention mechanism. First, by calculating the position code, the spatial position information of the wire is effectively encoded into the feature representation; then the features are mapped through a learnable transformation matrix, and the position-aware features are generated in combination with the position code; on this basis, a spatial constraint matrix is ​​introduced to modulate the attention calculation, so that the attention mechanism can better capture the spatial relationship between the wires. This embodiment enables the system to accurately grasp the spatial distribution and relative position relationship of the wires, effectively solving the difficulties of traditional methods in dealing with large-scale deformations and changes in perspective. At the same time, through the adaptive feature fusion mechanism, the discriminability of feature expression is improved, so that the system can better handle the matching problem of wires between different images.

[0070] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0071] S31. Based on the fused feature vectors of different images, a multi-scale feature representation is constructed to obtain a scale feature set; based on the scale feature set, the spatial dependency is calculated to obtain a spatial correlation matrix; based on the spatial correlation matrix, an adaptive feature conversion is performed to obtain a transformation feature matrix; based on the transformation feature matrix, cross-scale attention is calculated to obtain an inter-layer attention tensor; based on the inter-layer attention tensor, a multi-layer attention weight set is generated;

[0072] S32, based on the multi-layer attention weight set, perform a dynamic weight fusion operation to obtain a fusion weight vector; based on the fusion weight vector, calculate a weighted sum to obtain a cross-image fusion feature;

[0073] S33. Based on the cross-image fusion features, the forward matching probability and the reverse matching probability are calculated; based on the forward matching probability and the reverse matching probability, a bidirectional consistency check is performed to obtain a consistency enhancement feature.

[0074] In one embodiment of the present application, the fused feature vectors f' of different images are obtained, denoted as fA and fB respectively, and the mutual attention matrix Cl = tanh((WlfA)T(WlfB)) is calculated in each layer l, where Wl is the learnable weight matrix of the lth layer; after adding the relative position constraint Rl, normalization Al = softmax(Cl + Rl) is performed, where Rl is the relative position constraint matrix, and a multi-layer attention weight set A is obtained.

[0075] Get the multi-layer attention weight set A, perform the dynamic weight fusion operation w = sigmoid(FC([A1; ...; AL])), where FC is the fully connected layer and L is the number of attention layers; calculate the weighted sum fAB = Σ(wi·Ai), where wi is the fusion weight of the i-th layer, and get the cross-image fusion feature fAB.

[0076] Get the cross-image fusion feature fAB, calculate the forward matching probability CA→B = softmax(fAB) and the reverse matching probability CB→A = softmax(fBAT); perform a bidirectional consistency check C = min(CA→B, CB→AT) to obtain the consistency enhancement feature C.

[0077] This embodiment achieves high-precision matching of wire features across images by constructing a multi-layer mutual attention network and a bidirectional consistency check mechanism. First, the mutual attention matrix is ​​calculated at each layer, and relative position constraints are introduced for modulation. The feature correlations of different scales and abstract levels are captured through a multi-layer structure; then, the adaptive integration of attention at different layers is achieved through a dynamic weight fusion operation; finally, the forward matching probability and the reverse matching probability are calculated, and a bidirectional consistency check is performed, which effectively improves the reliability of the matching. This embodiment improves the accuracy of wire registration, especially when dealing with challenging scenes such as complex backgrounds and partial occlusions, showing strong robustness, while reducing the probability of mismatching.

[0078] According to one aspect of the present application, step S31 is further:

[0079] S311, constructing a multi-scale decomposition function based on the fused feature vector; performing feature decomposition based on the multi-scale decomposition function to obtain feature representation at each scale level; normalizing the feature representation at each scale level to obtain a scale feature set;

[0080] S312, based on the scale feature set, calculating the spatial displacement vector between the feature point pairs, constructing a spatial correlation function; based on the spatial correlation function, constructing a spatial constraint matrix, and generating a spatial correlation matrix;

[0081] S313. Based on the scale feature set and the spatial correlation matrix, an adaptive conversion function is constructed to calculate a feature response matrix; based on the feature response matrix, a feature enhancement operation is performed to obtain a transformed feature matrix.

[0082] S314. Based on the transformation feature matrix, a cross-scale attention calculation function is constructed to calculate the inter-layer attention score; based on the inter-layer attention score and the transformation feature matrix, a multi-layer attention response is constructed to obtain an inter-layer attention tensor;

[0083] S315. Based on the inter-layer attention tensor, construct an attention aggregation function and calculate the final attention score; based on the final attention score, perform attention normalization on each layer to obtain a multi-layer attention weight set.

[0084] In one embodiment of the present application, the fused feature vectors fA and fB are obtained, and a multi-scale decomposition function ψl(f) = f * Gl + ηl·▽ is constructed. 2 f, where Gl is the Gaussian kernel, ▽ 2 is the Laplace operator, ηl is the scale coefficient; feature decomposition is performed at L scale levels: sl = ψl(f) + ρl·s(l-1), where ρl is the inter-layer coupling factor, sl and s(l-1) are feature representations at different scale levels; normalization is performed on each layer of features to obtain scale feature sets SA and SB.

[0085] Get the scale feature sets SA and SB, calculate the spatial displacement vector d(i, j) = pi - pj between the feature point pairs, where pi and pj are the spatial coordinates of the feature points; construct the spatial correlation function φ(d) = exp(-||d||2 / 2σ 2 )·(1+κ·cos(θij)), where σ is the spatial scale parameter, κ is the directional modulation coefficient, and θij is the directional angle; generate the spatial constraint matrix R(i, j) = φ(d(i, j)) and obtain the spatial correlation matrix R.

[0086] Obtain the scale feature sets SA and SB and the spatial correlation matrix R, construct the adaptive conversion function τ(f, R) = f +λ·tanh(μ·(R * f)), where λ is the conversion intensity coefficient and μ is the nonlinear modulation factor; calculate the feature response matrix E(i, j) =τ(fi, Rij)·τ(fj, Rji)T; perform the feature enhancement operation T = norm(E +β·E 2 ), where β is the second-order enhancement coefficient, and the transformed feature matrices TA and TB are obtained.

[0087] Obtain the transformation feature matrices TA and TB, construct the cross-scale attention calculation function ω(Tl, Tk) = softmax(Tl·WlkT·Tk), where Wlk is the inter-layer mapping matrix; calculate the inter-layer attention score αlk = ω(Tl, Tk)·(1 +γ·|lk|)-1, where γ is the inter-layer attenuation factor; integrate the multi-layer attention responses L(l, k) = αlk·(Tl + Tk) / 2 to obtain the inter-layer attention tensor L.

[0088] Get the inter-layer attention tensor L and construct the attention aggregation function a(L) = Σ(πl·Ll), where πl is the layer weight coefficient; calculate the final attention score A = sigmoid(ξ·a(L)), where ξ is the scaling factor; perform attention normalization on each layer to obtain a multi-layer attention weight set A.

[0089] This embodiment achieves accurate matching of wire features across images through multi-scale feature representation and inter-layer attention mechanism. First, a hierarchical feature representation is constructed through a multi-scale decomposition function, and an inter-layer coupling factor is introduced to achieve effective feature transfer; a spatial constraint relationship between feature points is established through a spatial correlation function to ensure the spatial consistency of the matching; an adaptive conversion function is used to enhance the feature response, and the discriminability of the feature is improved through second-order enhancement; finally, through cross-scale attention calculation and inter-layer attention fusion, effective integration of multi-layer features is achieved. This embodiment improves the accuracy and robustness of wire registration, especially when dealing with large-scale changes and complex deformations, showing excellent performance.

[0090] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0091] S41. Based on the consistency enhancement feature, a multi-dimensional similarity feature is constructed to obtain a similarity feature tensor; based on the similarity feature tensor, local structural constraints are extracted to obtain a structural constraint matrix; based on the structural constraint matrix, topological relationship features are calculated to obtain a topological feature set; based on the similarity feature tensor, the structural constraint matrix and the topological feature set, a multi-constraint fusion function is constructed to obtain a comprehensive constraint tensor; based on the comprehensive constraint tensor, an initial matching matrix is ​​generated;

[0092] S42. Construct topological consistency constraints, direction consistency constraints, and distance consistency constraints, optimize the initial matching matrix, and obtain an optimized matching matrix.

[0093] In one embodiment of the present application, a consistency enhancement feature C is obtained, and the feature similarity sim(fi, fj) and the geometric similarity sim(gi, gj) are calculated; a weighted fusion operation Sij = βf·sim(fi, fj) + βg·sim(gi, gj) is performed, where βf and βg are adaptive weight coefficients, and an initial matching matrix S is obtained.

[0094] The initial matching matrix S is obtained and optimized based on the topological consistency constraint T(Xi)≈T(Xj), the directional consistency constraint |θi -θj| <ε, and the distance consistency constraint D(pi, pj) <τ, where T(·) is the topological feature extraction function, θi and θj are the directional angles, D(·) is the distance metric function, ε and τ are the threshold parameters, and the optimized matching matrix X is obtained.

[0095] This embodiment adopts a multi-constraint matching matrix generation and optimization strategy to achieve high-precision wire matching result output. The initial matching matrix is ​​constructed by fusing feature similarity and geometric similarity, and an adaptive weight coefficient is introduced for dynamic balance; in the optimization stage, multi-dimensional constraints and optimization of the matching results are achieved by applying topological consistency constraints, direction consistency constraints, and distance consistency constraints. This embodiment not only ensures the accuracy of local matching, but also maintains the consistency of the global geometric structure, effectively solves complex problems such as deformation and scale changes in wire registration, and improves the reliability and accuracy of the registration results.

[0096] According to one aspect of the present application, step S41 is further:

[0097] S411. Obtain consistency enhanced feature C, construct feature distance measurement function df(fi,fj) = ||fi - fj||2·(1 - cos<fi,fj>), where cos<·, ·> is cosine similarity; calculate local shape descriptor ds(pi,pj) = Σwk·||▽kpi - ▽kpj||2, where ▽k is k-order differential operator and wk is weight coefficient; fuse multi-dimensional similarity D(i,j) = [df(fi,fj), ds(pi,pj), C(i,j)] to obtain similarity feature tensor D.

[0098] S412, obtain the similarity feature tensor D, and construct the local structure tensor S(p) =Σwij·(pi - p)(pj -p) T , where wij is the spatial weight; calculate the structural similarity measure gs(Si, Sj) = ||Si - Sj|| F / max(||Si|| F ,||Sj|| F ), where ||·|| F is the Frobenius norm; performs non-local structure contrast g(i, j) = exp(-gs(Si, Sj) 2 / 2σ 2 ), and obtain the structural constraint matrix G.

[0099] S413. Obtain the structural constraint matrix G, construct the local neighborhood graph Ne(i) = {j | ||pi - pj|| <ε}; calculate the topological eigenvector t(i) = [deg(i), ecc(i), cen(i)], where deg is degree centrality, ecc is eccentricity, and cen is betweenness centrality; perform the topological similarity measure τ(i, j) = exp(-||t(i) - t(j)||2 / 2ν 2 ), and obtain the topological feature set T.

[0100] S414. Obtain the similarity feature tensor D, the structural constraint matrix G and the topological feature set T, construct a multiple constraint fusion function h(i, j) = [D(i, j), G(i, j), T(i, j)]; calculate the constraint weight w = softmax(φ(h)), where φ is the feature mapping function; perform weighted fusion H(i, j) = Σwk·hk(i, j) to obtain the comprehensive constraint tensor H.

[0101] S415. Obtain the comprehensive constraint tensor H and construct a matching score function s(i, j) = ω·H(i, j) + (1-ω)·∏ k∈N(i) max(H(k, l)), where ω is the global-local balance factor, N(i) is the neighborhood set of i, and is the ∏ product operation; perform matching score normalization: S(i, j) = s(i, j) / max(s(i, :), s(:, j)), and obtain the initial matching matrix S.

[0102] This embodiment achieves high-reliability wire matching through multi-dimensional similarity features and constraint fusion. By constructing feature distance measurement functions and local shape descriptors, multi-dimensional similarity measurement is achieved; through local structure tensor analysis and non-local structure comparison, a complete structural constraint system is established; topological feature vectors are introduced to capture global structural information, and the unified integration of constraints is achieved through multiple constraint fusion functions; finally, the matching score function is used to balance global and local information to generate high-quality initial matching results. This embodiment effectively improves the accuracy and reliability of matching, especially when dealing with complex scenes and abnormal situations, and shows strong robustness.

[0103] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0104] S51. Based on the optimized matching matrix, construct multi-dimensional evaluation features to obtain an evaluation feature set; based on the evaluation feature set, calculate the local consistency measure to obtain a consistency feature matrix; based on the consistency feature matrix, extract the global stability feature to obtain a stability feature tensor; fuse the evaluation feature set, the consistency feature matrix and the stability feature tensor to generate a comprehensive evaluation tensor; based on the comprehensive evaluation tensor, construct a reliability scoring function to generate a reliability scoring matrix;

[0105] S52: Based on the reliability scoring matrix and the preset adaptive threshold, conditional judgment is performed on each position to obtain a final matching result.

[0106] In one embodiment of the present application, the optimized matching matrix X is obtained, and the comprehensive reliability score R(i, j) = α1·Sij +α2·Cij +α3·Gij is calculated, where Sij is the feature similarity score, Cij is the consistency score, Gij is the geometric consistency score, α1, α2, and α3 are adaptive weight coefficients, and the reliability scoring matrix R is obtained.

[0107] Get the reliability score matrix R, set the adaptive threshold η, and perform conditional judgment on each position (i, j): when R(i, j) >η and R(i, j) is the maximum value of both its row and column, set the corresponding position as a matching point; otherwise, set it as a non-matching point to obtain the final matching result M.

[0108] By constructing multi-dimensional evaluation features and calculating local consistency metrics, this embodiment can carefully evaluate each element of the matching matrix, capture more detailed information, and thus improve the accuracy of matching; the local consistency metric ensures the consistency of the matching results within the local range and reduces the possibility of mismatching. Based on the consistency feature matrix, the global stability feature is extracted, which can effectively evaluate the stability of the matching results within the global range; by comprehensively considering local and global features, it is ensured that the matching results are not only consistent within the local range, but also stable within the global range, thereby improving the reliability of the matching results. The evaluation feature set, the consistency feature matrix, and the stability feature tensor are fused to generate a comprehensive evaluation tensor, which can comprehensively evaluate all aspects of the matching results, generate a reliability scoring matrix for each matching result, and provide an important reference for subsequent decision-making. Based on the reliability scoring matrix and the preset adaptive threshold, conditional judgment is performed on each position, and the screening criteria of the matching results can be dynamically adjusted. This adaptive threshold judgment mechanism can be flexibly adjusted according to the actual situation, thereby improving the accuracy and adaptability of the matching results. This embodiment can generate a final matching result with high precision and high reliability, and can provide strong guarantees in terms of both local consistency and global stability to ensure the accuracy and reliability of the matching results.

[0109] According to one aspect of the present application, step S51 is further:

[0110] S511, obtain the optimized matching matrix X, and calculate the feature similarity evaluation score sf(i, j) = (1 + ||fi- fj||2 / δf) -1 , where δf is the feature distance threshold; construct the spatial transformation consistency evaluation sc(i, j) = exp(-||Ti(pi) - pj||2 / 2σs 2), where Ti is the local affine transformation and σs is the spatial scale parameter; calculate the directional continuity evaluation sd(i, j) = cos(θi - θj), where θi and θj are local direction angles; combine the multi-dimensional evaluation features E(i, j) = [sf(i, j), sc(i, j), sd(i, j)] to obtain the evaluation feature set E.

[0111] S512, obtain the evaluation feature set E, and construct the local neighborhood consistency function ψ(i, j) = Σ k∈N(i) Σ l∈N(j) X(k,l)·exp(-||dik - djl||2 / 2σl 2 ), where dik and djl are relative displacement vectors, and σl is the local scale parameter; calculate the structure preservation metric μ(i, j) = |log(ri / rj)|, where ri and rj are local radii; fuse the local evaluation indicators L(i, j) = [ψ(i, j), μ(i, j), E(i, j)] to obtain the consistency feature matrix L.

[0112] S513, obtain the consistency feature matrix L, construct the global shape descriptor h(P) = [κ(P), τ(P), ρ(P)], where κ is the curvature feature, τ is the torsion feature, and ρ is the density feature; calculate the shape similarity measure ds(hi,hj) = ||hi - hj||2·(1 - cos<hi,hj>); perform stability evaluation q(i,j) = exp(-ds(hi,hj) 2 / 2σq 2 )·(1 +η·|Ne(i) - Ne(j)|)-1, where Ne is the number of effective matches and σq is the stability scale parameter, and the stability characteristic tensor Q is obtained.

[0113] S514. Obtain the evaluation feature set E, the consistency feature matrix L and the stability feature tensor Q, and construct a feature fusion network v(i, j) = [E(i, j), L(i, j), Q(i, j)]; calculate the adaptive weight coefficient w = softmax(φ(v)), where φ is a nonlinear mapping function; perform weighted fusion V(i, j) = Σwk·vk(i, j) to obtain a comprehensive evaluation tensor V.

[0114] S515. Obtain the comprehensive evaluation tensor V and construct a reliability scoring function r(i, j)=λg·V(i, j)+λl·Σ k∈N(i)max(V(k,l)), where λg,λl are global-local balance factors; apply adaptive threshold adjustment r'(i,j) = sigmoid(ζ·(r(i,j) – r*)), where r* is the average reliability score and ζ is the scaling factor; perform regional consistency correction R(i,j) = r'(i,j)·∏ k∈Ω(i，j) (1 + ε·r'(k, l)), where Ω is the local area and ε is the association strength parameter, and the reliability score matrix R is obtained.

[0115] This embodiment achieves high-precision screening of matching results through multi-dimensional evaluation features and comprehensive reliability metrics. First, a complete evaluation system is constructed through feature similarity evaluation, spatial transformation consistency evaluation, and directional continuity evaluation; local consistency of matching is evaluated through local neighborhood consistency function and structure preservation metric; global stability is evaluated using global shape descriptors and shape similarity metrics; and finally, the final matching screening is achieved through reliability scoring function and regional consistency correction. This embodiment effectively improves the reliability of the registration results, especially when dealing with noise interference and abnormal matching, showing excellent discrimination ability.

[0116] In one embodiment of the present application, a method for wire registration between drone aerial images based on deep learning includes the following steps:

[0117] S1. Construct mathematical model;

[0118] Get two UAV aerial images I_A and I_B with overlapping areas. M wires L_A have been detected in image I_A:={L_A 1 , L_A 2 , ..., L_A M}, whose index is A:={1, 2, ..., M}. N wires L_B:={L_B 1 , L_B 2 , ..., L_B N}, whose index is B:={1, 2, ..., N}. Each wire detected in the two images is given as a set of points. After matching, the matching matrix P∈[0, 1] of the wires in the two images is finally output. M×N .

[0119] S2, extracting features of different wires in the same image based on the self-attention mechanism;

[0120] The feature extraction part of the ResNet18 model is used to extract features from the image, and then the points on the wire are sampled on the feature map according to their positions in the image. Each point on the wire can obtain 512-dimensional features. Assuming that there are n points on each wire, each wire corresponds to a feature f_i∈R n×512In order to integrate the features of all points on each wire for subsequent feature matching, a self-attention module is used to integrate the global information of the wire.

[0121] The self-attention module associates different positions of a single sequence, allowing inputs to interact with each other, and the dimensions of the output and input remain the same. In the self-attention module, first a linear transformation is performed on the input f_i to generate three matrices K, Q, and V: K=W K f_i;Q=W Q f_i;V=W V f_i; where W K , W Q , W V represents the linear transformation layer, and the final output is: f_o=Attention(Q, K, V)=Softmax((QK T ) / sqrt(d_k))V; where d_k is the number of columns of matrix K. Since the self-attention mechanism can utilize global information, the first dimension of f_o is used as the extracted wire feature, that is, f_o∈R 1×512 .

[0122] Through the wire feature extraction module, each wire L_i∈{L_A, L_B} in the image I∈{I_A, I_B} can obtain a 512-dimensional feature f_o I , and use it as the initial state s_i of each wire I ←f_o I These wire features are then fed into a feature matching module, which consists of L (L=7) identical layers that jointly process two feature sets. Each layer consists of a self-attention module and a cross-attention module to update the representation of each wire. In each module, a multi-layer perceptron (MLP) is used to update the state s_i I to update.

[0123] For the self-attention module, the state s_i I Update method is as follows: s_i I ←s_i I +MLP([s_i I | y_i I←I ]), where [∙|∙] represents the superposition of two vectors in dimension; y_i I←I =∑ J∈{I_A，I_B} Softmax(α_ik IJ ) j Ws_j {I_A，I_B} ; where Softmax(x_i )=e x_i / (∑ je x_j ), W is the mapping matrix, k∈{I_A, I_B}. α_ij IJ is the attention score between lines Li and Lj. In the self-attention module, each wire pays attention to all wires in the same image. For each line Li∈{LA, LB}, the current state xi I Transformed into key vector k_i and query vector q_i through linear transformation, the attention score α_ij between lines L_i and L_j is defined as: α_ij=q_i T k_j.

[0124] S3, extract the features of the same wire in different images based on the mutual attention mechanism;

[0125] For the criss-cross attention module, each wire in an image attends to all wires in another image, and the state x_i I Update method is as follows: s_i I ←s_i I +MLP([s_i I | y_i I←{I_A，I_B}\I ]); y_i I←{I_A，I_B}\I =∑ J∈{I_A，I_B} Softmax(α_ik IJ ) j Ws_j {I_A，I_B} ; α_ij = k_i T After the self-attention module and the cross-attention module form a multi-layer network, a lightweight output head is used to predict the feature matching situation based on the updated state.

[0126] S4, obtaining the corresponding relationship between the wires through feature matching;

[0127] Calculate the similarity matrix S∈R between the wires in the two images M×N :S_ij=Linear(s_i A ) T Linear(s_i B ) ; all (i, j)∈A×B; where the similarity can be considered as the possibility that this pair of wires are projected from the same 3D wire in space, where Linear(∙) represents the linear transformation layer, the data dimension remains unchanged, s_i A and s_i B Indicates the result after the initial state of the wires in images I_A and I_B is updated by the multi-layer network. In addition to calculating the similarity matrix, the matchability score σ_i∈[0,1] of each wire is also considered: σ_i=Sigmoid(Linear(s_i )), where Sigmoid(x)=1 / (1+e -x), is a common activation function that maps variables to between 0 and 1. The matchability score indicates the probability that the current wire has a corresponding wire. The closer the value is to 1, the higher the probability. Combining the similarity matrix and the matchability score, we get the wire matching matrix P: P_ij=σ_i A σ_j B Softmax(S_(k_1 j) ) i Softmax(S_(ik_2 )) j , k_1∈A, k_2∈B.

[0128] The loss function of the model training process consists of three parts, corresponding to the correctly matched and unmatched wires. The first part of the loss function is to calculate the log-likelihood loss of the predicted correspondence P_ij for the correctly matched wires, where P_ij is the element in the predicted assignment matrix; the second part is to calculate the predicted unmatchability score 1-σ_i for the unmatched wires in image A. A The log-likelihood loss of A is the matchability score of wire i in image A; the third part is to calculate the predicted unmatchability score 1-σ_j for the unmatched wires in image B B The log-likelihood loss of B is the matchability score of wire j in image B. It is shown as follows:

[0129] loss=-1 / L∑ l (1 / |M|∑ (i，j)∈M logP_ij+1 / 2|A*|∑ i∈A* log(1-σ_i A )+1 / 2|B*|∑ j∈B* log(1-σ_j B )) ;

[0130] Where M represents the correctly matched wires, A* represents the unmatched wires in image A, and B* represents the unmatched wires in image B.

[0131] S5. Give matching results based on the transmission line samples, effectively realize the correspondence and information fusion between the wires;

[0132] The matching matrix P takes into account the similarity and matching scores of the corresponding wires. In the prediction stage, for a pair of wires, when both wires are matchable and the similarity between them is higher than other wires, then the pair of wires is related. When the value of P_ij is greater than the threshold (set to 0.1) and greater than the other values ​​of its column and row, then the pair of wires is confirmed to be matched.

[0133] This embodiment uses the self-attention mechanism to help the model understand and focus on the internal features of the wires, while the mutual attention mechanism can focus on the features of the same wire in different images, thereby improving the accuracy of matching. By accurately matching wires, data from different images can be integrated to achieve comprehensive fusion and in-depth analysis of information, providing data support for safety assessment and maintenance strategies of power lines. Based on accurate wire matching results, power companies can make more scientific and reasonable maintenance plans and resource allocation decisions, which helps to build a more intelligent, efficient and reliable power system.

[0134] According to one aspect of the present application, a wire registration system between drone aerial images based on deep learning includes:

[0135] at least one processor; and,

[0136] a memory communicatively connected to at least one of the processors; wherein,

[0137] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the deep learning-based wire registration method between drone aerial images described in any of the above embodiments.

[0138] The present invention realizes high-precision automatic registration of wires between drone aerial images through a multi-stage adaptive processing strategy. Through the organic combination of adaptive image preprocessing mechanism, nonlinear multi-scale feature extraction, position-aware self-attention feature fusion, multi-layer mutual attention matching network and multi-constraint optimization strategy, the system can effectively handle various complex situations in aerial scenes, such as illumination changes, perspective differences, local deformation, etc. In practical applications, the present invention improves the accuracy of wire detection and registration, reduces computational complexity, and improves processing efficiency. In particular, it shows excellent robustness and scalability when processing large-scale aerial data. The end-to-end design of the system not only simplifies the actual deployment process, but also provides good real-time performance, providing reliable technical support for intelligent inspection of power systems.

[0139] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for wire registration between UAV aerial images based on deep learning, characterized in that: The steps include: S1. Obtain at least two original drone aerial images containing overlapping areas, and perform adaptive dynamic range compression processing on them to obtain preprocessed images; based on the preprocessed images, construct a nonlinear multi-scale feature pyramid, extract the wire features of each scale layer, and obtain an enhanced wire feature set; Based on the enhanced wire feature set, a fusion feature matrix is ​​constructed through dynamic feature fusion; S2, based on the fusion feature matrix, obtain the position-aware feature matrix through spatial-aware self-attention processing; Based on the position-aware feature matrix, the enhanced self-attention score is calculated to obtain the fused feature vector between different wires; S3. Based on the fused feature vector, a multi-layer mutual attention network is constructed to obtain a multi-layer attention weight set; based on the multi-layer attention weight set, feature aggregation is performed to obtain cross-image fusion features; Based on cross-image fusion features, consistency enhancement features are obtained through bidirectional consistency checks; S4. Based on the consistency enhancement feature, a multi-constraint matching matrix is ​​constructed to obtain an initial matching matrix; Optimize the initial matching matrix to obtain an optimized matching matrix; S5. Based on the optimized matching matrix, the matching reliability score is calculated to obtain a reliability scoring matrix; based on the reliability scoring matrix and a preset threshold, the final matching result is screened.

2. The method for wire registration between drone aerial images based on deep learning according to claim 1, characterized in that: Step S1 is further as follows: S11, obtaining at least two original drone aerial images containing overlapping areas; and evaluating and obtaining an image quality feature vector based on the image quality of the original drone aerial images; Based on the image quality feature vector, adaptive enhancement is performed to obtain an enhanced image; Based on the enhanced image, illumination equalization and noise suppression are performed to obtain a balanced image; Based on the balanced image, edge feature enhancement is performed to obtain an edge enhanced image; Based on the edge enhanced image, adaptive dynamic range compression is performed to obtain a preprocessed image; S12. Based on the preprocessed image, construct an improved Laplacian pyramid; based on the Laplacian pyramid, perform compression operation on each layer to obtain a compressed feature map; Based on the compressed feature map, pyramid decomposition operation is performed to obtain the Laplacian feature map; based on the Laplacian feature map, a multi-scale feature pyramid is generated; S13. Based on the multi-scale feature pyramid, extract local structural features to obtain a structural feature matrix; based on the structural feature matrix, calculate the directional consistency features to obtain a directional feature set; based on the directional feature set, extract and fuse context information to obtain context enhancement features; Based on the context-enhanced features, a multi-dimensional feature tensor is integrated to finally generate an enhanced wire feature set; S14, calculating the dynamic fusion weight based on the enhanced wire feature set; Based on the dynamic fusion weights, a weighted sum operation is performed to obtain a fusion feature matrix.

3. The method for wire registration between drone aerial images based on deep learning according to claim 2 is characterized in that: Step S2 is further as follows: S21, obtaining a wire feature set of the same image in the fusion feature matrix, and calculating a position code; based on the position code, performing a feature transformation operation to obtain a position-aware feature matrix; S22. Based on the position-aware feature matrix, a spatial constraint matrix is ​​constructed, the attention score is calculated, and a fused feature vector is obtained.

4. The method for wire registration between drone aerial images based on deep learning according to claim 3 is characterized in that: Step S3 is further as follows: S31, constructing a multi-scale feature representation based on the fused feature vectors of different images to obtain a scale feature set; Based on the scale feature set, the spatial dependency is calculated to obtain the spatial correlation matrix; Based on the spatial correlation matrix, adaptive feature conversion is performed to obtain a transformed feature matrix; Based on the transformed feature matrix, calculate the cross-scale attention and obtain the inter-layer attention tensor; Generate a multi-layer attention weight set based on the inter-layer attention tensor; S32, based on the multi-layer attention weight set, perform a dynamic weight fusion operation to obtain a fusion weight vector; based on the fusion weight vector, calculate a weighted sum to obtain a cross-image fusion feature; S33, calculating the forward matching probability and the reverse matching probability based on the cross-image fusion features; Based on the forward matching probability and the reverse matching probability, a bidirectional consistency check is performed to obtain a consistency enhancement feature.

5. The method for wire registration between drone aerial images based on deep learning according to claim 4, characterized in that: Step S4 is further as follows: S41. Based on the consistency enhancement feature, a multi-dimensional similarity feature is constructed to obtain a similarity feature tensor; based on the similarity feature tensor, local structural constraints are extracted to obtain a structural constraint matrix; Based on the structural constraint matrix, the topological relationship features are calculated to obtain the topological feature set; based on the similarity feature tensor, the structural constraint matrix and the topological feature set, a multi-constraint fusion function is constructed to obtain the comprehensive constraint tensor; Based on the comprehensive constraint tensor, an initial matching matrix is ​​generated; S42. Construct topological consistency constraints, direction consistency constraints, and distance consistency constraints, optimize the initial matching matrix, and obtain an optimized matching matrix.

6. The method for wire registration between drone aerial images based on deep learning according to claim 5, characterized in that: Step S5 is further as follows: S51. Based on the optimized matching matrix, construct a multi-dimensional evaluation feature to obtain an evaluation feature set; based on the evaluation feature set, calculate a local consistency measure to obtain a consistency feature matrix; Based on the consistency feature matrix, the global stability features are extracted to obtain the stability feature tensor; The evaluation feature set, consistency feature matrix and stability feature tensor are integrated to generate a comprehensive evaluation tensor; based on the comprehensive evaluation tensor, a reliability scoring function is constructed to generate a reliability scoring matrix; S52: Based on the reliability scoring matrix and the preset adaptive threshold, conditional judgment is performed on each position to obtain a final matching result.

7. The method for wire registration between drone aerial images based on deep learning according to claim 6, characterized in that: Step S11 is further as follows: S111, obtaining at least two original drone aerial images containing overlapping areas, calculating image clarity features, illumination uniformity features, and signal-to-noise ratio features; combining the image clarity features, illumination uniformity features, and signal-to-noise ratio features to obtain an image quality feature vector; S112, constructing an adaptive enhancement function based on the original UAV aerial image and its corresponding image quality feature vector; Based on the adaptive enhancement function, an enhancement operation is performed on each pixel to obtain an enhanced image; S113, constructing an adaptive weight matrix based on the enhanced image, performing weighted smoothing operation to obtain a smoothed image; performing local contrast correction based on the smoothed image to obtain a balanced image; S114, based on the balanced image, calculating multi-directional gradient responses and constructing a directional enhancement kernel; Based on the direction enhancement kernel, a direction selective convolution operation is performed to obtain an edge enhanced image; S115, calculating local statistical features based on the edge enhanced image and constructing an adaptive compression coefficient; Based on the adaptive compression coefficient, a dynamic adjustment factor is calculated; Based on the edge enhanced image and the dynamic adjustment factor, a pixel-level adaptive compression operation is performed to obtain a preprocessed image.

8. The method for wire registration between drone aerial images based on deep learning according to claim 6, characterized in that: Step S13 is further as follows: S131. Based on the multi-scale feature pyramid, a structure tensor is constructed at each scale layer, and eigenvalues ​​and eigenvectors are calculated; based on the eigenvalues ​​and eigenvectors, a local structure descriptor is constructed; Based on the local structure descriptor, feature aggregation is performed on each detection point to obtain a structure feature matrix; S132. Based on the structural feature matrix, calculate the direction vector field and construct a direction consistency measure; based on the direction consistency measure, perform non-maximum suppression to obtain the local main direction; based on the local main direction, use a tensor voting mechanism to enhance the direction coherence and obtain a direction feature set; S133, constructing an adaptive sampling window based on the directional feature set to extract local context features; Based on the local context features, non-local mean filtering is applied to obtain context enhanced features; S134, constructing a multi-dimensional feature descriptor based on the context-enhanced feature and calculating the feature importance weight; Based on the feature importance weights, weighted feature fusion is performed to obtain a multi-dimensional feature tensor; S135, extracting a predetermined number of key points for each wire based on the multidimensional feature tensor, and calculating a feature representation of each key point; Based on the feature representation of each key point, feature normalization and dimension reduction are performed to obtain the enhanced wire feature set.

9. The method for wire registration between drone aerial images based on deep learning according to claim 6, characterized in that: Step S31 is further as follows: S311, constructing a multi-scale decomposition function based on the fused feature vector; performing feature decomposition based on the multi-scale decomposition function to obtain feature representation at each scale level; normalizing the feature representation at each scale level to obtain a scale feature set; S312, based on the scale feature set, calculating the spatial displacement vector between the feature point pairs, and constructing a spatial correlation function; Based on the spatial correlation function, a spatial constraint matrix is ​​constructed to generate a spatial correlation matrix; S313, based on the scale feature set and the spatial correlation matrix, construct an adaptive conversion function and calculate the feature response matrix; Based on the feature response matrix, a feature enhancement operation is performed to obtain a transformed feature matrix; S314. Based on the transformation feature matrix, a cross-scale attention calculation function is constructed to calculate the inter-layer attention score; Based on the inter-layer attention scores and the transformed feature matrix, a multi-layer attention response is constructed to obtain the inter-layer attention tensor. S315. Based on the inter-layer attention tensor, construct an attention aggregation function and calculate the final attention score; Based on the final attention score, attention normalization is performed on each layer to obtain a set of multi-layer attention weights.

10. A deep learning-based wire registration system for UAV aerial images, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the deep learning-based wire registration method between drone aerial images as described in any one of claims 1 to 9.

Images