A power tower inclination degree detection method based on a rotating target detection network

Abstract

The application discloses a power tower tilt degree detection method based on a rotating target detection network, and belongs to the technical fields of power line inspection and image detection. The method comprises the following steps: collecting a plurality of power tower images, and making the images into a training set, a verification set and a test set; pre-processing the images in the training set, the verification set and the test set; constructing a rotating target detection model; establishing a loss function of a multi-head branch network, and introducing a regression loss term based on Kalman filtering into the loss function; inputting the pre-processed training set, the verification set and the test set into the rotating target detection model, and training the model; inputting a to-be-detected image after pre-processing into the trained rotating target detection model, performing reasoning, and outputting a tilt degree detection result; the power tower tilt degree and the tower category are double classified, and the power tower target is accurately positioned, and the problem of low calculation precision of the existing rotating intersection ratio is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power line inspection and image detection, and particularly relates to a power tower tilt degree detection method based on a rotating target detection network. BACKGROUND

[0002] As one of important equipment supporting power transmission lines, power towers are often affected by natural disasters such as typhoon and earthquake, resulting in unstable tower base, tower body tilt or even tower body breakage, which affects normal power transmission and supply, and even causes economic loss. However, the complex and harsh natural environment and wide distribution area of power towers increase the difficulty of power transmission line inspection, and the long inspection cycle makes it impossible to effectively maintain abnormal power towers in time. Therefore, in recent years, the power operation and maintenance department has introduced unmanned aerial vehicle (UAV) inspection technology to collect and detect information of power towers by carrying various types of sensing devices on the UAV, but there is a common problem of "paying more attention to data collection and less attention to data analysis".

[0003] With the development of artificial intelligence technology, intelligent analysis and processing of UAV inspection images are gradually replacing manual judgment and becoming an important means to improve the level of power grid operation and maintenance. However, existing researches mainly improve the conventional target detection model based on deep learning by expanding the data set, introducing an attention mechanism into the feature extraction network, and multi-scale and multi-stage feature fusion, but there is a lack of targeted improvement method to effectively improve the model detection accuracy for power towers with characteristics such as multi-angle, multi-scale, multi-pose, and large aspect ratio. SUMMARY

[0004] In view of the deficiencies of the prior art, the purpose of the present application is to provide a power tower tilt degree detection method based on a rotating target detection network.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A power tower tilt degree detection method based on a rotating target detection network, characterized in that it comprises the following steps:

[0007] Collecting a plurality of power tower images and making them into a training set, a verification set and a test set;

[0008] Preprocessing the images in the training set, the verification set and the test set;

[0009] Constructing a rotating target detection model, which comprises a feature extraction network, a multi-scale feature fusion network and a multi-head branch network;

[0010] Establishing a loss function of the multi-head branch network and introducing a regression loss term based on Kalman filtering into the loss function;

[0011] The preprocessed training set, the verification set and the test set are input into the rotating target detection model to train the model;

[0012] After the to-be-detected image is preprocessed, the trained rotating target detection model is input to perform inference, and a tilt degree detection result is output.

[0013] Further, the proportion of the training set, the verification set and the test set is 3:1:1, and the quantity proportion of the normal tower, the slightly abnormal tower and the severely abnormal tower is 21:1:3.

[0014] Further, the image preprocessing step is:

[0015] S21, a sliding window with a size of 600 pixels x 600 pixels and a step of 450 pixels is used to cut the original image, and then the obtained cropped block is resized to 800 pixels x 800 pixels;

[0016] S22, the cropped block is randomly horizontally flipped with a probability of 0.5 for data augmentation, and the RGB three channels are normalized to [-1, 1] respectively, so that the original image satisfies the normal distribution.

[0017] Further, the feature extraction network adopts a 50-layer residual network structure, and a plurality of shortcut connections are introduced in the ordinary convolution to realize the fusion or identity transformation of the features;

[0018] The multi-scale feature fusion network adopts a 5-layer feature pyramid structure, which first maps from bottom to top and retains features of different scales on the basis of the residual network, then transmits deep semantic information layer by layer from top to bottom, and finally combines shallow detail features and deep semantic features through horizontal connection;

[0019] The multi-head branch network includes a classification sub-network, a regression sub-network and a tilt degree sub-network; the classification sub-network and the tilt degree sub-network each consist of 4 Cx3x3 convolution layers and 1 KAx3x3 convolution layer, wherein C represents the number of intermediate layer feature map channels, K represents the number of categories, and A represents the number of anchor boxes; the regression sub-network consists of 4 Cx3x3 convolution layers and 1 5Ax3x3 convolution layer, wherein C represents the number of intermediate layer feature map channels, 5 represents the number of positioning parameters, including the horizontal and vertical coordinates of the center point of the rotating frame, the length and width, and the rotation angle.

[0020] Further, the loss function construction step of the multi-head branch network includes:

[0021] S41, the rotating frame is modeled by Gauss, and the rotation intersection over union between the predicted frame and the real frame is calculated based on the Kalman filtering idea;

[0022] S42, based on the rotation IoU between the predicted box and the real box, a consistency regression loss function based on Kalman filtering is established:

[0023] L kf (Σ1,Σ2)=e 1-KFIoU -1

[0024]

[0025] L reg =L c +L kf

[0026] In the formula, L kf is the IoU loss based on Kalman filtering, L c is the center point distance loss between the predicted box and the real box;

[0027] S43, the tower type loss function and the tilt degree type loss function are established, and finally the loss function of the multi-head branch network is obtained.

[0028] Further, the specific steps of calculating the rotation IoU between the predicted box and the real box are as follows:

[0029] S411, linear transformation is performed on the standard normal two-dimensional Gaussian distribution, so that the transformed two-dimensional Gaussian distribution is inscribed in the rotation box, and the linear transformation formula is as follows:

[0030]

[0031]

[0032] A=RS

[0033] In the formula, R is a rotation matrix used to control the rotation angle of the Gaussian distribution, S is a scale matrix used to control the scaling value of the two-dimensional Gaussian distribution in two dimensions, A is a linear transformation matrix, θ is a rotation angle, s w and s h are and w and h are the length and width of the rotation box, respectively;

[0034] S412, based on the solution of S411, the parameters u and Σ of the two-dimensional Gaussian distribution are calculated:

[0035]

[0036]

[0037] In the formula, Λ is a diagonal matrix containing the eigenvalues of Σ, and (x, y) represents the center point coordinates of the rotation box;

[0038] S413, referring to the Kalman filter correction method, taking the prediction box in the rotating target detection as the estimated value, the real box as the observed value, and the overlapping area as the optimal estimation between the estimated value and the observed value, calculating the Gaussian distribution of the new overlapping area by using the product of the two Gaussian distributions corresponding to the prediction box and the real box

[0039]

[0040]

[0041] wherein μ, μ1, μ2, Σ, Σ1, Σ2 are two parameters of the Gaussian distribution, det(·) is a determinant, and α is a parameter controlling the distance between the two Gaussian distributions.

[0042] S414, when the distance in S1413 is close to 0, then α is approximately equal to a constant, at this time, the area of the intersection rectangular frame of the rotating frame can be calculated according to Σ in

[0043] wherein eig(·) is an eigenvalue of a matrix; and the intersection union ratio between the rotating frames can be finally expressed as:

[0044]

[0045]

[0046] wherein A, A1, and A2 are the areas of the real box, the prediction box, and the circumscribed box of the intersection area, respectively.

[0047] Further, the loss function of the multi-head branch network is:

[0048]

[0049] wherein the first term is a regression loss, N and N pos are the number of all anchor boxes and positive sample anchor boxes, respectively, are the Gaussian distribution functions of the nth prediction box and the nth real box, respectively; the last two terms are a tower category loss and a tilt degree category loss, respectively, t n , t' n are the tower category label and the tilt degree category label of the nth target, respectively, p n , p' n ​​​​​Respectively, the probability distribution of the nth prediction frame after the S-shaped curve function corresponding to the tower category and the inclination degree category, and λ 1, λ 2, λ 3 are hyperparameters, the value range is controlled between 0.01 to 1.

[0050] A power tower inclination degree detection system based on a rotating target detection network, comprising:

[0051] An image acquisition module: acquiring a plurality of power tower images and making a training set, a verification set and a test set;

[0052] An image preprocessing module: preprocessing the images in the training set, the verification set and the test set;

[0053] A model construction module: constructing a rotating target detection model, the model comprising: a feature extraction network, a multi-scale feature fusion network and a multi-head branch network;

[0054] A function construction module: establishing a loss function of the multi-head branch network, and introducing a regression loss term based on Kalman filtering into the loss function;

[0055] A model training module: inputting the preprocessed training set, the verification set and the test set into the rotating target detection model to train the model;

[0056] And an inference detection module: inputting the to-be-detected image after preprocessing into the trained rotating target detection model, performing inference, and outputting the inclination degree detection result.

[0057] A power tower inclination degree detection device based on a rotating target detection network, comprising:

[0058] One or more processors;

[0059] A memory for storing one or more programs;

[0060] When the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0061] A storage medium containing computer executable instructions, when the computer executable instructions are executed by a processor, the processor executes the above method.

[0062] The beneficial effects of the present application are:

[0063] 1、The power tower inclination degree detection method of the present application adds an inclination degree branch on the basis of the original rotating target detection network, realizes the dual classification of the power tower inclination degree and the tower category and the accurate positioning of the power tower target.

[0064] 2. A regression loss term based on Kalman filtering is introduced into the loss function, which solves the problem of low calculation accuracy of the existing rotation intersection ratio by taking advantage of Gaussian modeling. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0066] Figure 1 is a flow chart of the power tower tilt detection method of the present application;

[0067] Figure 2 is a rotating target detection network structure diagram of the present application

[0068] Figure 3 is a PR curve diagram of R 3 det of the present application before and after improvement and the ordinary target detection model (normal);

[0069] Figure 4 is a PR curve diagram of R 3 det of the present application before and after improvement and the ordinary target detection model (mild abnormality);

[0070] Figure 5 is a PR curve diagram of R 3 det of the present application before and after improvement and the ordinary target detection model (severe abnormality). DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort are within the scope of protection of the present application.

[0072] As shown in Figure 1 , a power tower tilt detection method based on a rotating target detection network comprises the following steps:

[0073] S1, collecting power tower images and making them into a training set, a validation set and a test set;

[0074] Collecting images of multiple power towers by a UAV and making the collected images into a training set, a validation set and a test set; wherein the proportion of the training set, the validation set and the test set is 3:1:1, and the quantity proportion of normal towers, mild abnormality towers and severe abnormality towers is 21:1:3.

[0075] S2, pre-processing the images in the training set, the validation set and the test set;

[0076] The steps of image pre-processing are:

[0077] S21, cutting the original image by using a sliding window with a size of 600 pixels x 600 pixels and a step of 450 pixels, and then adjusting the size of the obtained cropped block to 800 pixels x 800 pixels;

[0078] S22, randomly horizontally flipping the cropped block with a probability of 0.5 for data augmentation, and normalizing the three channels of RGB to [-1, 1] respectively, so that the original image satisfies the normal distribution.

[0079] S3, constructing a rotating target detection model;

[0080] As shown in Figure 2 , the rotating target detection model comprises a feature extraction network, a multi-scale feature fusion network and a multi-head branch network.

[0081] The feature extraction network adopts a 50-layer residual network structure, and a plurality of shortcut connections are introduced in the ordinary convolution to realize the fusion or identity transformation of the features. The feature extraction network solves the problem of deep network degradation and improves the feature expression ability of the network.

[0082] The multi-scale feature fusion network adopts a 5-layer feature pyramid structure. On the basis of the residual network, the features of different scales are first mapped from bottom to top and reserved, then the deep semantic information is transmitted from top to bottom layer by layer, and finally the shallow detail features and deep semantic features are organically combined through horizontal connection.

[0083] The multi-head branch network comprises a classification sub-network, a regression sub-network and a tilt degree sub-network, which are respectively used for classification, positioning and tilt degree detection of the power tower;

[0084] The classification sub-network and the tilt degree sub-network each comprise 4 convolution layers of Cx3x3 and 1 convolution layer of KAx3x3, wherein C represents the number of intermediate layer feature map channels, K represents the number of categories, and A represents the number of anchor boxes; the regression sub-network comprises 4 convolution layers of Cx3x3 and 1 convolution layer of 5Ax3x3, wherein C represents the number of intermediate layer feature map channels, 5 represents the number of positioning parameters, including the horizontal and vertical coordinates of the center point of the rotating frame, the length and width, and the rotation angle.

[0085] S4, establishing a loss function of the multi-head branch network, and introducing a regression loss term based on Kalman filtering in the loss function;

[0086] The establishment steps of the loss function are:

[0087] S41, Gaussian modeling is performed on the rotating frame, and a rotation intersection ratio between a predicted frame and a real frame is calculated based on the Gaussian modeling combined with a Kalman filtering idea; specifically,

[0088] S411, a linear transformation is performed on a standard normal two-dimensional Gaussian distribution, so that the transformed two-dimensional Gaussian distribution is inscribed in the rotating frame, and the linear transformation formula is as follows,

[0089]

[0090]

[0091] A = RS

[0092] In the formula, R is a rotation matrix, used to control the rotation angle of the Gaussian distribution, S is a scale matrix, used to control the scaling value of the two-dimensional Gaussian distribution in two dimensions, A is a linear transformation matrix, θ is a rotation angle, s w and s h are and w and h are the length and width of the rotating frame, respectively;

[0093] S412, based on the solution of S411, the parameters u and Σ of the two-dimensional Gaussian distribution are calculated:

[0094]

[0095]

[0096] In the formula, Λ is a diagonal matrix containing the characteristic values of Σ, (x, y) represents the center point coordinates of the rotating frame;

[0097] S413, referring to the Kalman filtering correction method, the predicted frame in the rotating target detection is taken as an estimated value, the real frame is taken as an observed value, and the overlapping area is taken as the optimal estimation between the estimated value and the observed value. The product of the two Gaussian distributions corresponding to the predicted frame and the real frame is used to calculate the Gaussian distribution of the new overlapping area.

[0098]

[0099]

[0100] In the formula, μ, μ1, μ2, Σ, Σ1, Σ2 are two parameters of the Gaussian distribution, det(·) is a determinant, and α is a distance parameter between the two Gaussian distributions.

[0101] ​​S414, when the distance in S1413 is close to 0, then a is approximately equal to a constant, at this time, the area of the intersection region rotating rectangular frame can be calculated according to

[0102]

[0103] where eig(·) is the eigenvalue of the matrix; the intersection-over-union between the rotating frames can finally be expressed as:

[0104]

[0105] are the areas of the real frame, the predicted frame and the intersection region bounding box, respectively.

[0106] S42, a consistency regression loss function based on Kalman filtering is established:

[0107] L kf (Σ1,Σ2)=e 1-KFIoU -1

[0108]

[0109] L reg =L c +L kf

[0110] where L kf is the intersection-over-union loss based on Kalman filtering, and L c is the center point distance loss between the predicted frame and the real frame.

[0111] S43, a tower category loss function and a tilt degree category loss function are established, and finally a loss function of the multi-head branch network is obtained;

[0112] The tower category loss function and the tilt degree category loss function can be expressed as:

[0113] L cls (p n ,t n )=-α n (1-p n ) γ log(p n )

[0114]

[0115] L tilt (p' n ,t' n )=-α' n (1-p' n )​​γ log(p' n )

[0116]

[0117] In the formula, alpha n , alpha' n is a weight parameter for controlling positive and negative samples, p n , p' n is a weight parameter for controlling difficult and easy samples;

[0118] Finally, the regression loss, the tower category loss and the tower inclination degree category loss are added to obtain the loss function of the multi-head branch network:

[0119]

[0120] In the formula, the first term is the regression loss, N and N pos are the number of all anchor frames and positive sample anchor frames respectively, are the Gaussian distribution functions of the nth prediction frame and the nth real frame respectively; the last two terms are the tower category loss and the inclination degree category loss respectively, t n , t' n are the tower category label and the inclination degree category label of the nth target respectively, p n , p' n are the probability distribution of the tower category and the inclination degree category corresponding to the nth prediction frame after the S-shaped curve (Sigmoid) function, and lambda1, lambda2 and lambda3 are hyperparameters, the value range of which is controlled between 0.01 and 1.

[0121] S5, input the preprocessed training set, the verification set and the test set into the rotating target detection model for training.

[0122] S6, after the to-be-detected image is preprocessed, the image is input into the rotating target detection model for inference, and the result is output;

[0123] After the image is input into the rotating target detection model, the multi-scale features of the image are obtained after the feature extraction network and the multi-scale feature fusion network; further, the features are input into the multi-head branch composed of the classification sub-network, the regression sub-network and the inclination degree sub-network, and the power tower category, the coordinates and the inclination degree information are output.

[0124] Comparative experiment:

[0125] The present application takes R 3det (Refined Rotation RetinaNet) is a rotated target detection original model, which improves the model by adding a tilt degree branch and introducing a Kalman filter-based regression loss term in the loss function. In addition, the PR (Precision Recall) curve and mAP (mean Average Precision) quantitative indicators are used to test the performance of different models, as shown in Figure 3 , 4 , 5 and Table 1. Among them, AP n , AP ab1 , AP ab2 are the AP (Average Precision) values of normal, mild abnormality, and severe abnormality, respectively.

[0126] Table 1

[0127]

[0128]

[0129] From the table, it can be seen that the mAP value of the original R 3 det is improved by 9.2%, 6.6%, and 5.2% respectively compared with the conventional target detection in each category, which indicates the superiority of the design of R 3 det network structure and the introduction of rotated target anchor frame. After introducing the multi-head branch, although the AP value of the mild tilt category decreases slightly, its performance improves by 3.2% in the severe tilt category detection. On this basis, the regression loss function is further improved, and Kalman filter is introduced to approximate the estimation of the intersection over union of the rotated frame. The overall precision of the final model can reach 94.5%, among which the normal AP is 90.8%, the mild tilt AP is 98.7%, and the severe tilt AP is 93.7%, which verifies the effectiveness of the Kalman filter loss function.

[0130] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0131] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A method for detecting the degree of tilt of a power tower based on a rotating target detection network, characterized in that, The method comprises the following steps: Collecting a plurality of power tower images and making a training set, a verification set and a test set; Pretreating the images in the training set, the verification set and the test set; Building a rotating target detection model, which comprises a feature extraction network, a multi-scale feature fusion network and a multi-head branch network; Establishing a loss function of the multi-head branch network and introducing a regression loss term based on Kalman filtering into the loss function; Inputting the pretreated training set, the verification set and the test set into the rotating target detection model to train the model; After pretreating a to-be-detected image, inputting the pretreated image into the trained rotating target detection model to perform inference and outputting a tilt degree detection result; The loss function building step of the multi-head branch network comprises: S41, Gaussian modeling is performed on the rotating frame, and the rotating intersection-over-union between the predicted frame and the real frame is calculated based on the Gaussian modeling and the Kalman filtering idea; S42, based on the rotating intersection-over-union between the predicted frame and the real frame, a consistency regression loss function based on Kalman filtering is established: L kf (∑1,∑2) = e 1-KFIoU -1 L reg = L c + L kf In the formula, L kf is the intersection over union loss based on Kalman filtering, L c is the center point distance loss between the predicted frame and the real frame, and KFIoU represents the intersection over union between the rotating frames. S43, a tower category loss function and a tilt degree category loss function are established, and finally the loss function of the multi-head branch network is obtained. 2.The power tower tilt detection method based on the rotating target detection network of claim 1, wherein, The proportion of the training set, the verification set and the test set is 3:1:1, and the quantity proportion of normal towers, lightly abnormal towers and severely abnormal towers is 21:1:

3. 3.The power tower tilt detection method based on the rotating target detection network of claim 1, wherein, The image pretreatment step comprises: S21, a sliding window with a size of 600 pixels*600 pixels and a step of 450 pixels is used to cut the original image, and then the size of the obtained cutting block is adjusted to 800 pixels*800 pixels; S22, the cutting block is randomly horizontally flipped with a probability of 0.5 for data enhancement, and the RGB three channels are respectively normalized to [-1, 1], so that the original image satisfies the normal distribution. 4.The power tower tilt detection method based on the rotating target detection network of claim 1, wherein, The feature extraction network adopts a 50-layer residual network structure, and a plurality of shortcut connections are introduced in the ordinary convolution to realize the fusion or identity transformation of the features; The multi-scale feature fusion network adopts a 5-layer feature pyramid structure, which maps from bottom to top on the basis of the residual network, retains features of different scales, then transmits deep semantic information from top to bottom layer by layer, and finally combines shallow detail features and deep semantic features through horizontal connection; The multi-head branch network comprises a classification subnetwork, a regression subnetwork and a tilt degree subnetwork; the classification subnetwork and the tilt degree subnetwork each comprise four convolution layers of C*3*3 and one convolution layer of KA*3*3, wherein C represents the number of intermediate layer feature map channels, K represents the number of categories, and A represents the number of anchor frames; the regression subnetwork comprises four convolution layers of C*3*3 and one convolution layer of 5A*3*3, wherein C represents the number of intermediate layer feature map channels, and 5 represents the number of positioning parameters, including the horizontal and vertical coordinates of the center point of the rotating frame, the length and width, and the rotation angle. 5.The power tower tilt detection method based on the rotating target detection network of claim 1, wherein, The specific steps of calculating the rotating intersection-over-union between the predicted frame and the real frame are as follows: S411, linear transformation is performed on the standard normal two-dimensional Gaussian distribution, so that the transformed two-dimensional Gaussian distribution is inscribed in the rotating frame, and the linear transformation formula is as follows: A=RS In the formula, R is a rotation matrix for controlling a rotation angle of a Gaussian distribution, S is a scale matrix for controlling a scaling value of a two-dimensional Gaussian distribution in two dimensions, A is a linear transformation matrix, θ is a rotation angle, s w and s h are respectively and w and h are respectively a length and a width of a rotated frame; S412, on the basis of the solving result of S411, calculating the parameters μ and Σ of the two-dimensional Gaussian distribution: In the formula, Λ is a diagonal matrix containing the eigenvalues of Σ, and (x, y) represents the coordinate of the center point of the rotation frame; S413, referring to the Kalman filter correction method, taking the prediction box in the rotating target detection as an estimated value, taking the real box as an observation value, taking the overlapping area as the optimal estimation between the estimated value and the observation value, and calculating the Gaussian distribution of the new overlapping area by using the product of the two Gaussian distributions corresponding to the prediction box and the real box ​ where μ, μ1, μ2, Σ, Σ1, Σ2 are respectively two parameters of the Gaussian distribution, det(·) is the determinant, and α is a control distance parameter between the two Gaussian distributions; S414, when the distance in S1413 is close to 0, then a is approximately equal to a constant, at this time, the area of the intersection region rotating rectangular frame can be calculated according to Σ in S1413 In the formula, eig(·) is the eigenvalue of the matrix; The intersection-over-union between the rotation frames can finally be expressed as: In the formula, are the areas of the true, predicted and intersection bounding boxes, respectively. 6.The power tower tilt detection method based on the rotating target detection network of claim 1, wherein, The loss function of the multi-head branch network is: where the first term is the regression loss, N and N pos are the number of all anchor boxes and positive anchor boxes, respectively, are the Gaussian distribution functions of the nth predicted box and the nth ground truth box, respectively; the last two terms are the tower class loss and the tilt degree class loss, t n ,t' n are the tower class label and the tilt degree class label of the nth target, respectively, p n ,p' n are the probability distributions of the tower class and the tilt degree class of the nth predicted box after the S-shaped curve function, λ1, λ2, λ3 are hyperparameters, and the value range is controlled between 0.01 and 1.

7. A power tower tilt degree detection system based on a rotating target detection network, performing the detection method of any one of claims 1-6, characterized in that, Including: An image acquisition module: acquiring a plurality of power tower images, and making a training set, a verification set and a test set; An image preprocessing module: preprocessing the images in the training set, the verification set and the test set; A model construction module: constructing a rotation target detection model, the model comprising: a feature extraction network, a multi-scale feature fusion network and a multi-head branch network; A function construction module: establishing a loss function of the multi-head branch network, and introducing a regression loss term based on Kalman filtering into the loss function; A model training module: inputting the preprocessed training set, verification set and test set into the rotation target detection model to train the model; And an inference detection module: inputting the to-be-detected image after preprocessing into the trained rotation target detection model, performing inference, and outputting a tilt degree detection result. 8.A power tower tilt degree detection device based on a rotating target detection network, characterized in that, Including: One or more processors; Memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-6.

9. A storage medium containing computer-executable instructions, wherein: When the computer executable instructions are executed by the processor, the processor executes the method of any one of claims 1-6.

Images