An insulator ultraviolet image reconstruction method based on physical-data double driving

Abstract

This invention discloses a method for reconstructing ultraviolet images of insulators based on a physics-data dual-drive approach. The method includes: acquiring an initial ultraviolet image of the insulator and preprocessing it to obtain a preprocessed ultraviolet image; calculating the temporal pose of the camera based on the preprocessed ultraviolet image, then calculating the pixel shift at each exposure time to generate a motion blur kernel processed by sub-pixel weights and kernel regularization; establishing a physics-driven loss function based on the motion blur kernel; designing a multi-scale regularized U-shaped deep network reconstruction model based on the physics-data dual-drive approach based on the physics-data dual-drive approach; and outputting the reconstructed ultraviolet image of the insulator through the multi-scale regularized U-shaped deep network reconstruction model. This invention enables simultaneous optimization of motion blur and physical constraints, thereby obtaining a high-resolution and interpretable reconstructed image.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing technology, and specifically to a method for reconstructing ultraviolet images of insulators based on a physical-data dual-drive approach. Background Technology

[0002] The overhead contact system, a crucial component of electrified railways, provides energy to trains and ensures their normal operation. Insulators play a vital role in the power sector, and their proper functioning is essential for railway safety. During power line inspections, the use of drones or inspection vehicles equipped with lightweight ultraviolet imaging equipment to acquire ultraviolet image data of insulator operating conditions has become an important method for digital operation and maintenance. However, during image data acquisition, image blurring caused by high-speed relative motion between the camera and the target results in severe motion blur in the captured images, leading to loss of image details, blurred edges, and a decreased signal-to-noise ratio. This directly affects subsequent analysis of insulator corona discharge.

[0003] Currently, methods for handling image degradation caused by motion blur include traditional physics-driven methods and deep learning data-driven methods. However, these existing methods cannot achieve simultaneous optimization of motion blur and physical constraints, affecting the resolution and interpretability of the reconstructed image. Summary of the Invention

[0004] In view of this, the present invention provides a method for reconstructing ultraviolet images of insulators based on physical-data dual-drive, so as to achieve synchronous optimization of motion blur and physical constraints, thereby obtaining a high-resolution and interpretable reconstructed image.

[0005] A method for reconstructing ultraviolet images of insulators based on a physical-data dual-drive approach includes:

[0006] Step S1: Obtain an initial ultraviolet image of the insulator with motion blur, which is collected by the high-speed inspection vehicle or generated during aerial photography of high-voltage transmission lines. Then, preprocess the initial ultraviolet image of the insulator to obtain a preprocessed ultraviolet image of the insulator.

[0007] Step S2: Based on the preprocessed ultraviolet image of the insulator, calculate the temporal pose of the camera, then calculate the pixel shift at each exposure time, and then generate a motion blur kernel processed by sub-pixel weights and kernel regularization.

[0008] Step S3: Based on the motion blur kernel, the physical mechanism of image imaging is transformed into mathematical constraints that can be embedded in the deep learning framework. At the same time, the energy conservation and spectral characteristics of ultraviolet imaging are combined to establish a physical driving loss function.

[0009] Step S4: Based on the physical-driven loss function established in step S3, a multi-scale regularized U-shaped deep network is introduced to design a reconstruction model based on physical-data dual-driven multi-scale regularized U-shaped deep network. The reconstructed ultraviolet image of the insulator is output through the multi-scale regularized U-shaped deep network reconstruction model.

[0010] The insulator ultraviolet image reconstruction method based on a physics-data dual-driven approach provided by this invention calculates the temporal pose of the camera based on the preprocessed insulator ultraviolet image, then calculates the pixel shift at each exposure moment, and generates a motion blur kernel processed by sub-pixel weights and kernel regularization. Based on the motion blur kernel, the physical mechanism of image imaging is transformed into mathematical constraints that can be embedded in a deep learning framework. Simultaneously, combining the energy conservation and spectral characteristics of ultraviolet imaging, a physics-driven loss function is established. This ensures that the reconstructed image, while matching the results, also follows the physical energy conservation and imaging geometry laws, effectively avoiding the physical distortion that may occur with purely data-driven methods. It achieves simultaneous optimization of motion blur and physical constraints. This physics-driven loss function is embedded in a multi-scale regularized U-shaped deep network reconstruction model, enabling the generation of high-resolution and interpretable reconstructed images. Furthermore, this invention has higher processing efficiency and can meet the real-time detection requirements of high-speed inspection and UAV inspection. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the ultraviolet image reconstruction method for insulators based on physical-data dual-drive provided in an embodiment of the present invention. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0013] Please see Figure 1 The embodiments of the present invention provide a method for reconstructing ultraviolet images of insulators based on physical-data dual-drive, including steps S1 to S4:

[0014] Step S1: Acquire an initial ultraviolet image of the insulator with motion blur, which is collected by the high-speed inspection vehicle or generated during aerial photography of high-voltage transmission lines. Then, preprocess the initial ultraviolet image of the insulator to obtain a preprocessed ultraviolet image of the insulator.

[0015] The initial ultraviolet (UV) images of insulators are acquired using high-speed inspection vehicles equipped with UV cameras or during aerial photography of high-voltage transmission lines. These initial UV images exhibit motion blur. Preprocessing of the initial UV images is necessary, employing the PTPv2 protocol to synchronize the camera and IMU clocks, ensuring a timestamp deviation of <1ms. The input initial UV image of the insulator is in 8 / 12-bit RAW format. Exposure and shutter information are synchronously recorded from camera metadata and synchronized with IMU data. During exposure, cubic spline interpolation is performed on the IMU data to compensate for the asynchronous opening and closing of the shutter.

[0016] The preprocessing of the initial ultraviolet image of the insulator specifically includes:

[0017] Step S11: Acquire dark field images at different temperatures, construct a temperature-dependent dark current lookup table, and perform dark current correction on the initial insulator UV image based on the current temperature. The expression is:

[0018]

[0019] in, This is the initial ultraviolet image of the insulator. This represents the image after dark current correction. For pixel coordinates, For temperature Dark-field images captured at the time, For temperature coefficient, The current temperature. This is a reference temperature.

[0020] Dark current correction can effectively eliminate false brightness signals in low grayscale areas.

[0021] Step S12: After dark current correction, the ultraviolet image of the insulator still suffers from differences in the spatial distribution of image brightness due to issues such as vignetting of the ultraviolet camera lens, non-uniformity of sensor pixel response, and uneven illumination of the optical system. To eliminate these differences, a flat-field image acquired by a uniform light source using a calibration plate or integrating sphere is typically used. To perform flat field correction, the expression is:

[0022]

[0023] in, The image after flat-field correction. It is a flat-field image. This is a dark field image. This represents the average pixel value of the flat field image.

[0024] Step S13: Based on the Symlets wavelet basis, perform multi-scale decomposition and soft thresholding on the flat-field corrected image to obtain the processed image.

[0025] Even after planar correction, the image still retains some minor noise. A three-level decomposition using the Symlets wavelet basis is employed. Wavelet transform is then used to decompose the insulator's ultraviolet image into different scales and frequencies, effectively separating noise (typically distributed in the high-frequency subband) from the useful signal. Soft thresholding is applied to the high-frequency coefficients to achieve initial noise reduction. Further local smoothing is then applied to the wavelet-denoised image to eliminate any remaining block artifacts or minor noise. Simultaneously, the image's own structural information is used as a guide to protect and enhance edges.

[0026] S14. The ROI is extracted from the processed image to reduce background interference, enhance the network’s sensitivity to the features of defects, textures and light spots, improve the training stability of the network to be reconstructed in the future, detect the insulator region in real time, and dynamically adjust the ROI position through Kalman filtering. Motion blur modeling is performed only on the insulator region to make the IMU trajectory correspond to the main body of the image. Mirror reflection is used to fill the edge region, and finally the preprocessed ultraviolet image of the insulator is obtained.

[0027] Step S2: Based on the preprocessed ultraviolet image of the insulator, calculate the temporal pose of the camera, then calculate the pixel shift at each exposure time, and finally generate a motion blur kernel processed by subpixel weights and kernel regularization.

[0028] Specifically, step S2 includes:

[0029] Step S21: Based on the motion parameters of the preprocessed ultraviolet image of the insulator, obtain the synchronized angular velocity and acceleration output by the IMU. Combined with the initial velocity and initial attitude of the camera, correct the IMU drift and zero bias error, and calculate the temporal pose of the camera through inertial integration. The expression is:

[0030]

[0031] in, It is a fleeting moment. yes Timing pose of the camera It is a rotation matrix describing the camera's pose. This represents the translation vector of the camera's displacement during the exposure.

[0032] In this process, based on the angular velocity and acceleration output from the IMU synchronized in step S1, combined with the camera's initial velocity and attitude, a two-step correction is required to reduce errors and correct IMU drift and bias errors. The first step employs a short-time extended Kalman filter, which utilizes motion continuity constraints to estimate and compensate for IMU measurement noise in real time. The second step involves tightly coupled vision-inertial fusion, optimizing the raw visual and inertial data at the underlying level. Specifically, when the ultraviolet camera possesses visible features, optical flow or visual odometry is used to assist in correcting gyroscope drift. Finally, the camera's temporal pose is obtained through inertial integration, with a sampling frequency typically between 200 and 1000 Hz to ensure the temporal continuity of the pose data.

[0033] Step S22: Based on the temporal pose of the camera, the three-dimensional pose change is mapped onto the imaging plane to obtain the pixel shift at each exposure moment.

[0034] Specifically, based on the temporal pose of the camera obtained in step S21, the three-dimensional pose change is mapped onto the imaging plane to obtain the pixel shift at each exposure moment. , , , These represent the instantaneous displacement of pixels in the insulator image in the horizontal and vertical directions, respectively. If camera distortion exists, the pixel coordinates are first mapped to the normalized plane using intrinsic parameters and a distortion correction model.

[0035] Step S23: Based on the pixel shift at each exposure moment obtained in step S2.2, generate a motion blur kernel processed by sub-pixel weights and kernel regularization, with the following expression:

[0036]

[0037] in, For motion blur kernel, The number of discrete sampling moments for the exposure time. For Dirac functions, For the first The horizontal displacement at each sampling time. For the first The vertical displacement at each sampling time.

[0038] First, discrete pixel shift points are smoothed in a sub-pixel distribution form to reduce noise and sampling errors, improve the continuity of the blur kernel, and ensure the normality of the kernel function. Second, in order to prevent artifact stretching, abnormal jumps along the main motion direction can be smoothed.

[0039] The generated motion blur kernel is used to define the blur intensity based on the cumulative motion amplitude of the pixels during the exposure period. The expression is:

[0040]

[0041] in, Indicates the fuzzy intensity. The total exposure time of the camera, Indicates the exposure time At that time, the instantaneous displacement of the pixel in the horizontal direction; Indicates the exposure time At that time, the instantaneous displacement of the pixel in the vertical direction; Indicates time The differential.

[0042] Among them, the temporal pose of the camera and the motion blur kernel obtained in step S21 and blur intensity Together, these elements constitute the physical degradation model of the insulator's ultraviolet image. To improve the accuracy and computational efficiency of the physical degradation model, and considering that the effective signal in the ultraviolet image is only in the insulator region, an adaptive ROI module needs to be constructed. ROI extraction reduces computational burden and background interference. The workflow of the adaptive ROI module is as follows:

[0043] First, using the UV response diagram The grayscale distribution is analyzed, and thresholding or adaptive Otsu segmentation is used to filter out highlight areas; secondly, the blur intensity is combined with... The process involves filtering out high-ambiguity, low-signal regions and retaining target regions with high signal-to-noise ratios. Finally, the positional changes of the ROI are predicted based on the IMU pose, achieving stable tracking across frames.

[0044] The subsequent step S3 is executed only within the adaptive ROI module, which improves computational efficiency and convergence speed. After optimization by the adaptive ROI module, it can achieve millisecond-level real-time execution on embedded GPUs such as Jetson.

[0045] Step S3: Based on the motion blur kernel, the physical mechanism of image imaging is transformed into mathematical constraints that can be embedded in the deep learning framework. At the same time, the energy conservation and spectral characteristics of ultraviolet imaging are combined to establish a physical driving loss function.

[0046] Specifically, the physical mechanism of image imaging is transformed into mathematical constraints that can be embedded in a deep learning framework based on a motion blur kernel. Combined with the energy conservation and spectral characteristics of ultraviolet imaging, this ensures that the reconstructed image, while matching the results, also adheres to the physical energy conservation and imaging geometric laws. Finally, multiple losses are used to jointly optimize both the physical consistency and visual quality of the image. Specifically, the expression for the physics-driven loss function is:

[0047]

[0048] in, The physical-driven loss function, Let the reprojection consistency loss function be... The ultraviolet energy conservation loss function is... Let the edge consistency loss function be... For the total variational regularization term , , For weight hyperparameters.

[0049] With motion fuzzy kernel UV images of the original insulator Using IMU data as the core input, a reprojection consistency loss is designed. In implementation, a Fast Fourier Transform (FFT) is used to perform the convolution operation, and the convolution theorem is used for calculation in the frequency domain. This process transforms the IMU data into the kernel matrix, which is the motion blur kernel. To ensure differentiability, the generation process needs to be implemented as a differentiable operation chain. Alternatively, when using a fixed core, the generation process can be... Treat it as a constant tensor. Specifically, the reprojection consistency loss function. The expression is:

[0050]

[0051] in, This is the original ultraviolet image of the insulator. This is the reconstructed ultraviolet image of the insulator. This represents the square of the L2 norm.

[0052] Considering the imaging characteristics of insulators in specific ultraviolet bands, the radiant energy of the scene should be conserved. To ensure the overall energy consistency of the reconstructed image, the ultraviolet energy conservation term can be differentiable. Specifically, the ultraviolet energy conservation loss function... The expression is:

[0053]

[0054] in, for The corresponding ultraviolet energy operator; The image is a real-world ultraviolet image of an insulator without blur or noise. When no real value is available, edge estimation obtained by fusing multiple frames or simple deconvolution can be used as a substitute. for The corresponding ultraviolet energy operator, This represents the L1 norm.

[0055] With the reprojection consistency loss function and the ultraviolet energy conservation loss function already established, to avoid non-physical distortions or displacements of the insulator contours and discharge spot boundaries in the reconstructed ultraviolet image during image degradation, and to ensure that the reconstructed image maintains consistency with the real scene in terms of edge structure, a total variational regularization term penalizes unnecessary detail oscillations and noise in the image, promoting a smoother and more natural reconstructed ultraviolet image of the insulator. To maintain structural clarity and suppress noise artifacts, an edge consistency loss function is also introduced.

[0056] Edge consistency loss function The expression is:

[0057]

[0058] in, The gradient operator is implemented using the Sobel operator through convolution;

[0059] Total variational regularization term The expression is:

[0060]

[0061] in, for The partial derivative, express Gradient along the horizontal direction, for The partial derivative, express Gradient along the vertical direction.

[0062] In practice, step S3 can be implemented using a dedicated GPU acceleration architecture, supporting batch training and automatic differentiation. All loss terms are normalized and gradient pruning is used to ensure numerical stability, and they are linked with the ADMM optimizer through pluggable interfaces (hooks) to achieve joint optimization driven by both physics and data during network training.

[0063] Step S4: Based on the physical-driven loss function established in step S3, a multi-scale regularized U-shaped deep network is introduced to design a reconstruction model based on physical-data dual-driven multi-scale regularized U-shaped deep network. The reconstructed ultraviolet image of the insulator is output through the multi-scale regularized U-shaped deep network reconstruction model.

[0064] The network input of the reconstruction model simultaneously receives the original ultraviolet image of the insulator. Motion blur kernel With fuzziness intensity The three inputs are concatenated along the channel dimension to form a four-channel input tensor, which is then fused and feature extracted by a 3×3 convolution layer at the front end of the network.

[0065] In this embodiment, the multi-scale regularized U-shaped deep network reconstruction model includes an encoder and a decoder. The encoder introduces a multi-scale feature extraction module to effectively compensate for the lack of feature details in the original convolutional layer during feature extraction. The decoder introduces an attention module to solve the problem of inconsistent blur levels at different locations in high-speed motion blurred images.

[0066] Specifically, the encoding end consists of four levels of convolutional downsampling layers, each embedding a multi-scale feature extraction unit (MFE). The first branch is a 1×1 convolution, while the second, third, and fourth branches are all 3×3 convolutions. The dilation rates of the second, third, and fourth branches are 1, 2, and 3, respectively, which can extract features from different receptive fields. The outputs of each branch are concatenated and fused with the 1×1 convolution to finally form a scale-adaptive feature map H. The encoding end of this invention can solve the problem that fine feature information is easily lost in traditional convolutional parts and effectively capture detailed features across multiple scales.

[0067] On the decoding side, to address the challenges of convolutional operations when processing high-speed motion-blurred images, such as inconsistencies in global and local restoration caused by illumination variations, blurring, and differences in target size, a parallel attention mechanism is introduced. This mechanism combines positional attention and channel attention in parallel to simultaneously focus on the position and channels of deep features in the image and apply differentiated weights. By combining the outputs of the two attention mechanisms, effective feature extraction is enhanced, with particular attention paid to blurred regions in high-speed motion images, while irrelevant background information is reduced.

[0068] In the decoding stage, feature maps are generated using a positional attention mechanism. Feature maps are generated through channel attention mechanism. The expression for the parallel attention mechanism is:

[0069]

[0070] in, This represents the decoded features after fusion. and This is the self-learning weight matrix. This represents element-wise multiplication;

[0071] Fusion decoding features After ReLU activation and BatchNorm normalization, the reconstructed ultraviolet image of the insulator is finally generated. .

[0072] After incorporating a parallel attention mechanism at the decoding end, a 1×1 convolutional layer is configured at the end of the network, primarily to reduce the number of channels in the output feature map. This network adopts the Sigmoid activation function, which transforms the network's output values ​​to the [0, 1] interval, forming a probability tensor. This probability tensor is used to calculate the network's loss value during adversarial training. Optimization employs the Alternating Directional Multiplier (ADMM) method, alternately updating network parameters and physical constraint weights in each iteration to achieve synergistic optimization of physical consistency and data-driven feature learning. To ensure the physical consistency of the reconstructed image, the network training phase jointly minimizes the total loss function. , where the total loss function The expression is:

[0073] ;

[0074] in, This is a perceptual loss function used to maintain deep semantic consistency; To counteract the loss function, the PatchGAN discriminator is used to improve realism; This is a smoothing term used to remove artifacts; , , This is the weight value.

[0075] As a specific example, after outputting the reconstructed ultraviolet image of the insulator in step S4, the ultraviolet image reconstruction result of the insulator is evaluated, and post-processing, physical consistency verification, and image reconstruction quality evaluation are performed on the image respectively.

[0076] First, the reconstructed ultraviolet image of the insulator undergoes post-processing. First, gamma correction is used to adjust the global contrast of the output image. Then, histogram equalization is applied to normalize the brightness distribution of each channel, enhancing local details. Next, the ultraviolet brightness information is mapped to a color heatmap (e.g., blue-red gradient) to make the discharge spot distribution more prominent. Finally, using the ROI extraction process from step S14, invalid background interference is removed, and only the main body of the insulator and the discharge area are displayed and stored, thus generating a reconstructed image that is easy for both manual and algorithmic detection.

[0077] After post-processing, to verify the consistency between the reconstructed insulator image and the degraded model at the physical level, two physical evaluation metrics, reprojection error and energy error, were used. The reprojection error metric measures the difference between the reconstructed image after physical degradation convolution and the original blurred image; a smaller reprojection error indicates that the reconstructed model better conforms to physical imaging laws. The energy error metric assesses whether the reconstruction process maintained the conservation of ultraviolet radiation energy. Furthermore, by visualizing the correspondence between the blur intensity mapping and the gradient map of the reconstruction result, the model's adaptive recovery capability in highly blurred regions can be verified.

[0078] Finally, while ensuring that the reprojection error and energy error meet the physical consistency requirements, the similarity between the reconstructed image and the original sharp image is quantified by calculating metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). PSNR reflects the signal-to-noise ratio between the reconstructed image and the original sharp image; a higher PSNR value indicates less distortion. Structural similarity measures the consistency between structure and texture.

[0079] If the reconstructed insulator UV image meets the evaluation criteria, the final result is output; if it does not meet the evaluation criteria, the original insulator UV image is re-inputted until the criteria are met. The method of this invention employs a modular parallel design in its architecture, efficiently achieving GPU-accelerated inference through CUDA parallel computing and TensorRT quantization deployment. This method enables the system to achieve a single-frame latency of less than 60ms on an embedded platform, supporting high-precision insulator UV image reconstruction and defect detection in high-speed inspection vehicle or UAV inspection scenarios, and enabling real-time processing.

[0080] In summary, the insulator ultraviolet image reconstruction method based on a physics-data dual-driven approach described in the above embodiments calculates the temporal pose of the camera based on the preprocessed insulator ultraviolet image, then calculates the pixel shift at each exposure moment, and generates a motion blur kernel processed by sub-pixel weights and kernel regularization. Based on the motion blur kernel, the physical mechanism of image imaging is transformed into mathematical constraints that can be embedded in a deep learning framework. Simultaneously, combining the energy conservation and spectral characteristics of ultraviolet imaging, a physics-driven loss function is established. This ensures that the reconstructed image, while matching the results, also follows the physical energy conservation and imaging geometry laws, effectively avoiding the physical distortion that may occur with purely data-driven methods. It achieves simultaneous optimization of motion blur and physical constraints. This physics-driven loss function is embedded in a multi-scale regularized U-shaped deep network reconstruction model, enabling the generation of high-resolution and interpretable reconstructed images. Furthermore, this invention offers higher processing efficiency and can meet the real-time detection requirements of high-speed inspection and UAV inspection.

[0081] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for reconstructing ultraviolet images of insulators based on a physical-data dual-drive approach, characterized in that, include: Step S1: Obtain an initial ultraviolet image of the insulator with motion blur, which is collected by the high-speed inspection vehicle or generated during aerial photography of high-voltage transmission lines. Then, preprocess the initial ultraviolet image of the insulator to obtain a preprocessed ultraviolet image of the insulator. Step S2: Based on the preprocessed ultraviolet image of the insulator, calculate the temporal pose of the camera, then calculate the pixel shift at each exposure time, and then generate a motion blur kernel processed by sub-pixel weights and kernel regularization. Step S3: Based on the motion blur kernel, the physical mechanism of image imaging is transformed into mathematical constraints that can be embedded in the deep learning framework. At the same time, the energy conservation and spectral characteristics of ultraviolet imaging are combined to establish a physical driving loss function. Step S4: Based on the physical-driven loss function established in step S3, a multi-scale regularized U-shaped deep network is introduced to design a reconstruction model based on physical-data dual-driven multi-scale regularized U-shaped deep network. The reconstructed ultraviolet image of the insulator is output through the multi-scale regularized U-shaped deep network reconstruction model. Specifically, step S2 includes: Step S21: Based on the motion parameters of the preprocessed ultraviolet image of the insulator, obtain the synchronized angular velocity and acceleration output by the IMU. Combined with the initial velocity and initial attitude of the camera, correct the IMU drift and zero bias error, and calculate the temporal pose of the camera through inertial integration. The expression is: in, It is a fleeting moment. yes Timing pose of the camera It is a rotation matrix describing the camera's pose. The translation vector represents the displacement of the camera during the exposure period; Step S22: Based on the temporal pose of the camera, the three-dimensional pose change is mapped onto the imaging plane to obtain the pixel shift at each exposure moment; Step S23: Based on the pixel shift at each exposure moment obtained in step S22, generate a motion blur kernel processed by sub-pixel weighting and kernel regularization, with the following expression: in, For motion blur kernel, The number of discrete sampling moments for the exposure time. For Dirac functions, For pixel coordinates, For the first The horizontal displacement at each sampling time. For the first Displacement in the vertical direction at each sampling time; In step S3, the expression for the physics-driven loss function is: in, The physical-driven loss function, Let the reprojection consistency loss function be... The ultraviolet energy conservation loss function is... Let the edge consistency loss function be... For the total variational regularization term , , These are weight hyperparameters; Reprojection consistency loss function The expression is: in, This is the original ultraviolet image of the insulator. This is the reconstructed ultraviolet image of the insulator. Represents the square of the L2 norm; Ultraviolet energy conservation loss function The expression is: in, for The corresponding ultraviolet energy operator, These are blur-free and noise-free ultraviolet images of insulators in real-world scenarios. for The corresponding ultraviolet energy operator, Represents the L1 norm; Edge consistency loss function The expression is: in, For gradient operators; Total variational regularization term The expression is: in, for The partial derivative, for The partial derivative; In step S4, the total loss function of the multi-scale regularized U-shaped deep network reconstruction model is... The expression is: ; in, For the perceptual loss function, To counteract the loss function, For smoothing terms, , , These are weight values; The multi-scale regularized U-shaped deep network reconstruction model simultaneously receives the original ultraviolet images of the insulators at its network input. Motion blur kernel With fuzziness intensity The three elements are then concatenated along the channel dimension, where the blur intensity... The expression is: in, Indicates the fuzzy intensity. The total exposure time of the camera, Indicates the exposure time At that time, the instantaneous displacement of the pixel in the horizontal direction; Indicates the exposure time At that time, the instantaneous displacement of the pixel in the vertical direction; Indicates time The differential; The multi-scale regularized U-shaped deep network reconstruction model includes an encoder and a decoder. The encoding end consists of four levels of convolutional downsampling layers, each of which embeds a multi-scale feature extraction unit. The first branch is a 1×1 convolution, while the second, third, and fourth branches are all 3×3 convolutions. The dilation rates of the second, third, and fourth branches are 1, 2, and 3, respectively. The outputs of each branch are concatenated and then fused with the 1×1 convolution to finally form a scale-adaptive feature map H. A parallel attention mechanism is introduced in the decoding end. This mechanism combines the position attention mechanism and the channel attention mechanism in parallel to simultaneously pay attention to the position and channel of deep features in the image and perform differentiated weight allocation.

2. The insulator ultraviolet image reconstruction method based on physical-data dual-drive according to claim 1, characterized in that, In step S1, the initial ultraviolet image of the insulator is preprocessed, specifically including: Step S11: Acquire dark field images at different temperatures, construct a temperature-dependent dark current lookup table, and perform dark current correction on the initial insulator UV image based on the current temperature. The expression is: in, This is the initial ultraviolet image of the insulator. This represents the image after dark current correction. For temperature Dark-field images captured at the time, For temperature coefficient, The current temperature. For reference temperature; Step S12: Perform flat-field correction on the image after dark current correction using the flat-field plot. The expression is: in, The image after flat-field correction. It is a flat-field image. This is a dark field image. The average pixel value of the flat field image; Step S13: Perform multi-scale decomposition and soft thresholding on the flat-field corrected image based on the Symlets wavelet basis to obtain the processed image; S14. The ROI is extracted from the processed image, and the ROI position is dynamically adjusted by Kalman filtering. Motion blur modeling is performed only on the insulator region to make the IMU trajectory correspond to the main body of the image. Mirror reflection is used to fill the edge region, and finally the preprocessed ultraviolet image of the insulator is obtained.

3. The insulator ultraviolet image reconstruction method based on physical-data dual-drive according to claim 1, characterized in that, In the decoding stage, feature maps are generated using a positional attention mechanism. Feature maps are generated through channel attention mechanism. The expression for the parallel attention mechanism is: in, This represents the decoded features after fusion. and This is the self-learning weight matrix. This represents element-wise multiplication; Fusion decoding features After ReLU activation and BatchNorm normalization, the reconstructed ultraviolet image of the insulator is finally generated. .

Images