Blind super-resolution reconstruction method for inspection image of power transmission line

Abstract

The invention belongs to the technical field of power transmission lines, and particularly relates to a blind super-resolution reconstruction method for a power transmission line inspection image, which comprises the following steps: firstly, constructing a power transmission line image analysis unit, and standardizing an original power transmission line inspection image LRin to be subjected to blind super-resolution; using an RMC-SR image processing chip to realize blind super-resolution reconstruction of the inspection image of the power transmission line; on the basis, a closed-loop processing flow of prediction, self-inspection, correction and optimization is formed, so that the chip can dynamically balance local and global information, and high-precision and high-robustness blind super-resolution reconstruction of the inspection image of the power transmission line is realized.

Description

Technical Field

[0001] This invention relates to the field of power transmission line technology, specifically to a blind super-resolution reconstruction method for power transmission line inspection images. Background Technology

[0002] With the continuous expansion of the power grid and the increasing number of transmission lines operating in complex environments such as across regions, over long distances, and at high altitudes, power line inspection methods are gradually shifting from manual inspections to multi-source methods such as drone inspections, intelligent robot inspections, and fixed video surveillance. In transmission line operation and maintenance, high-quality inspection images play a crucial role in tasks such as conductor condition assessment, insulator defect identification, hardware corrosion analysis, foreign object entanglement detection, and line icing monitoring, and are one of the core data sources for ensuring the safe and stable operation of the power grid.

[0003] However, in actual inspection image acquisition, due to factors such as inspection platform vibration, long-distance shooting, wind speed variations, uneven lighting, rain and fog, equipment compression, and imaging resolution limitations, the acquired transmission line inspection images generally suffer from complex and unknown degradation problems, including low resolution, blurriness, noise pollution, significant compression artifacts, and weakened texture. These degradations not only severely affect the discernibility of equipment details but also reduce the accuracy of defect detection algorithms and image recognition models, posing challenges to the safe inspection of power lines.

[0004] Traditional super-resolution reconstruction methods typically rely on fixed interpolation algorithms such as bilinear interpolation and bicubic interpolation, which make it difficult to balance detail restoration and edge sharpness at high magnification. While deep learning-based super-resolution methods perform well under known degradation models, they generally assume that the degradation process is fixed, clear, and modelable. However, the degradation types of transmission line inspection images are complex, diverse, and unpredictable, resulting in insufficient generalization ability of such methods in real-world scenarios. The reconstruction results may exhibit problems such as texture artifacts, over-smoothing, or structural distortion.

[0005] Based on the above, traditional super-resolution reconstruction methods have the following specific problems: (1) Traditional interpolation methods and prior-based super-resolution methods are difficult to effectively recover small key components, such as conductor strands, insulator skirt edges, and hardware bolts, when processing images of power transmission line inspections. Under high magnification, these methods are prone to edge blurring, detail breakage, and obvious artifacts, making it difficult to meet the requirements of power line inspections for high-precision identification of small structures.

[0006] (2) Existing deep learning super-resolution methods usually rely on pre-set degradation models for training. However, actual inspection images are often accompanied by complex and unknown degradation conditions such as motion blur caused by wind vibration, degradation due to rain and fog, compression artifacts of UAVs, and noise superposition. This makes traditional networks less adaptable to real-world scenarios.

[0007] (3) Some existing methods lack adaptive processing mechanisms for local abnormal areas such as reflection, occlusion, and shadow in inspection images, and have not established effective strategies for balancing local details and overall structure. As a result, the super-resolution results are prone to quality degradation in local areas, and it is difficult to balance global contour and detail representation, which cannot meet the requirements of high robustness and high fidelity for power transmission line scenarios. Summary of the Invention

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A blind super-resolution reconstruction method for transmission line inspection images includes the following specific steps: S1, construct a transmission line image parsing unit to analyze the original transmission line inspection image LR to be analyzed using the super-resolution method. in Standardize; S2 uses the RMC-SR image processing chip to achieve blind super-resolution reconstruction of transmission line inspection images.

[0009] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, the specific steps of S1 are as follows: S11, Input transmission line inspection image Extract brightness information and generate a brightness image. Preserves the spatial structure and texture details of the image; S12, for The image edges are filled or cropped to generate a size-normalized image. This ensures that the input dimensions of subsequent processing units are consistent. S13, for Each pixel in Linear normalization is performed to obtain standardized transmission line inspection images. : (1) in and They represent The minimum and maximum pixels are determined to ensure that the normalized pixel values ​​are in the range [0, 1].

[0010] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, the specific steps of S2 are as follows: S21, using a blind super-resolution generation unit to process standardized transmission line inspection images Processing is performed to generate a preliminary super-resolution image. ; S22, through the regional consistency sensing unit... , Local inconsistency region detection is performed to obtain a consistency heatmap for image reconstruction during the inspection. ; S23, using a conditional reflexive correction unit to... and Self-reflexive correction is performed in the mid-deviation area to output a corrected version of the transmission line inspection image. and the confidence level of local correction in the reconstruction of inspection images ; S24, using global equilibrium units to , and Perform overall feature adjustments to generate adjusted transmission line inspection images. ; S25, through the credibility guidance unit... , , and Perform joint analysis and calculate The confidence weight corresponds to each pixel location, thereby generating a pending result image for blind super-resolution reconstruction of power transmission line inspection images. ; S26, in the adaptive optimization unit, and A joint evaluation of consistency and credibility is performed. If the credibility is less than a threshold, inspection image reconstruction feedback is generated. Used to iteratively update the updatable parameters in the above units. , If the confidence level is greater than or equal to the threshold or the maximum number of iterations is reached, then... The output is a blind super-resolution reconstruction result image of a power transmission line inspection. .

[0011] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, in step S21, the specific execution steps of the blind super-resolution generation unit are as follows: S211, Design a Blind-SR Feature Perception Module (BFPM) to achieve hierarchical modeling of input features of transmission line inspection images through a multi-branch parallel structure and cross-scale feature interaction; S212, a blind feature enhancement fusion module (BFEFM) is designed to adaptively adjust the channel weights in IBSRNet according to the degradation degree of the input transmission line inspection image, while enhancing structural information and detail features respectively; S213, a blind super-resolution progressive upsampling method BPUP is designed. By determining the low-resolution neighborhood corresponding to the high-resolution pixel, calculating the neighborhood mean and calculating the weighting coefficient based on distance attenuation, the neighboring pixels are subjected to high-frequency enhancement weighting, thereby achieving detail enhancement and edge fidelity in the blind super-resolution scenario of power transmission line inspection images. S214. Based on the BFPM module, PFEFM module, and BPUP method, a blind super-resolution reconstruction network (IBSRNet) for transmission line inspection images is constructed to receive standardized transmission line inspection images. As input, a preliminary super-resolution image robust to unknown degradation scenarios is generated. .

[0012] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, in step S22, the specific execution steps of the region consistency sensing unit are as follows: S221, for Two BPUP upsampling operations were performed to obtain the upsampled transmission line inspection image. ; S222, for Local error calculation is performed to obtain the local error feature map for the reconstruction of the inspection image. ; S223, for Perform channel averaging and apply Function to generate a consistent heatmap for inspection image reconstruction .

[0013] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, wherein: in step S23, the specific execution steps of the conditional reflexive correction unit are as follows: S231, for and Conditional feature modulation is performed to obtain the correction features for the reconstructed deviation region of the inspection image. ; S232, perform local corrections to obtain the corrected transmission line inspection image. Then, confidence level calculation is performed to obtain the local correction confidence level for the reconstruction of the inspection image. .

[0014] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, in step S24, the specific execution steps of the global balancing unit are as follows: S241, for , and Perform overall feature adjustment to obtain a globally balanced feature map for reconstructing the inspection image. ; S242, Generate the adjusted transmission line inspection image. .

[0015] As a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, in step S25, the specific execution steps of the confidence guidance unit are as follows: S251, for Perform channel stitching to generate inspection image reconstruction confidence weights ; S252, combined with credibility weight and The primary and secondary relationships are used to generate undetermined results for blind super-resolution reconstruction of transmission line inspection images. .

[0016] In a preferred embodiment of the blind super-resolution reconstruction method for transmission line inspection images according to the present invention, wherein: in step S26, the execution steps of the adaptive optimization unit are as follows: S261, according to and Calculate the global consistency residual for image reconstruction during inspection ; S262, perform a joint evaluation of consistency and credibility to obtain credibility. ; S263, Determining Reliability With threshold The size relationship is determined and processed accordingly.

[0017] Compared with existing technologies: Compared to existing technologies, this invention proposes a blind super-resolution reconstruction method for transmission line inspection images, implemented using the RMC-SR chip. This chip integrates a blind super-resolution generation unit, a region consistency sensing unit, a conditional reflexive correction unit, a global balancing unit, a confidence guidance unit, and an adaptive optimization unit, enabling high-quality super-resolution reconstruction of inspection images with unknown degradation. Within the IBSRNet executed by the blind super-resolution generation unit, this invention designs two core modules: BFPM and BFEFM. BFPM combines texture-aware frequency domain convolution (TAConv) and texture-aware dynamic normalization (TADN) to achieve multi-scale structural information extraction and high-frequency detail enhancement, effectively improving the reconstruction of conductors in transmission line inspection images. The chip possesses strong capabilities for restoring the local structure of key components such as insulators and fittings. BFEFM utilizes a blind super-resolution degradation adaptive channel weighting method (BDACW) to dynamically adjust channel responses based on different degradation conditions in transmission line inspection images, such as rain, fog, compression artifacts, and motion blur, suppressing noise and enhancing effective details. Simultaneously, the blind super-resolution progressive upsampling method (BPUP) in IBSRNet, through neighborhood weighting and high-frequency compensation, improves edge clarity and local discernibility while maintaining natural transitions in transmission line inspection images. By forming a closed-loop processing flow of "prediction-self-checking-correction-optimization," the chip can dynamically balance local and global information, achieving high-precision and robust blind super-resolution reconstruction of transmission line inspection images. Attached Figure Description

[0018] Figure 1 This is a flowchart of the technical solution of the present invention; Figure 2 This is a structural diagram of the IBSRNet of the present invention; Figure 3 This is a structural diagram of the BFPM module of the present invention; Figure 4 This is a structural diagram of the BFEFM module of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0020] This invention provides a blind super-resolution reconstruction method for transmission line inspection images, based on the RMC-SR chip. The chip consists of a blind super-resolution generation unit, a region consistency sensing unit, a conditional reflexive correction unit, a global balancing unit, a confidence guidance unit, and an adaptive optimization unit. The method receives and standardizes the original transmission line inspection image (LRin) to be super-reconstructed by the transmission line image parsing unit. First, it generates a preliminary super-resolution image and detects local inconsistencies based on this result. Then, it performs reflexive correction on the deviation regions and adaptively adjusts the overall features to balance local and global information. Finally, it obtains the undetermined result image of the blind super-resolution reconstruction of the transmission line inspection image through confidence guidance. Subsequently, the consistency and reliability of the undetermined blind super-resolution reconstruction result image and the standardized transmission line inspection image are jointly evaluated. When the reliability is lower than the set threshold, an adaptive optimization process is triggered to iteratively update the updatable parameters of the relevant units until the reliability is greater than or equal to the set threshold or the maximum number of iterations is reached. Then, the final blind super-resolution reconstruction result image SRout of the transmission line inspection image is generated, forming a closed loop of "prediction-self-checking-correction-optimization" to achieve high-quality blind super-resolution reconstruction of unknown degraded transmission line inspection images. In this invention, "blind super-resolution" is an abbreviation for "blind super-resolution".

[0021] Please see Figures 1-4 The specific steps are as follows: S1, construct a transmission line image parsing unit to analyze the original transmission line inspection image LR to be analyzed using the super-resolution method. in Standardize it.

[0022] The transmission line image analysis unit is used to process raw, low-resolution transmission line inspection images for blind super-resolution. Standardization and preliminary preprocessing are performed: S11, Input transmission line inspection image Extract brightness information and generate a brightness image. Preserves the spatial structure and texture details of the image; S12, for The image edges are filled or cropped to generate a size-normalized image. This ensures that the input dimensions of subsequent processing units are consistent. S13, for Each pixel in Linear normalization is performed to obtain standardized transmission line inspection images. : (1) in and They represent The minimum and maximum pixels are determined to ensure that the normalized pixel values ​​are in the range [0, 1].

[0023] S2 uses the RMC-SR image processing chip to achieve blind super-resolution reconstruction of transmission line inspection images.

[0024] The RMC-SR image processing chip consists of six functional units. In the first five units, the blind super-resolution generation unit is used for initial super-resolution generation, the regional consistency perception unit is used for local inconsistency region detection, the conditional reflexive correction unit is used for reflexive correction, the global balancing unit is used for overall feature adjustment, and the credibility guidance unit is used to generate the undetermined result map of blind super-resolution reconstruction of transmission line inspection images. The sixth unit is the adaptive optimization unit, which is used to jointly evaluate the consistency and credibility of the undetermined result map of blind super-resolution reconstruction and the standardized transmission line inspection image. When the credibility is detected to be lower than the set threshold, the adaptive optimization process is triggered, and the reconstruction feedback is output to iteratively update the updatable parameters of the aforementioned five units until the credibility is greater than or equal to the set threshold or the maximum number of iterations is reached, and the final blind super-resolution reconstruction result map of the transmission line inspection image is generated.

[0025] The specific steps of S2 are as follows: S21, using a blind super-resolution generation unit to process standardized transmission line inspection images Processing is performed to generate a preliminary super-resolution image. ; In step S21, the specific execution steps of the blind super-resolution generation unit are as follows: In the blind super-resolution generation unit, an Image Blind Super-Resolution Reconstruction Network (IBSRNet) for power line inspection images is executed. The structure of the IBSRNet network is as follows: Figure 2 As shown, the network comprises two core modules: BFPM and BFEFM. BFPM combines texture-aware frequency domain convolution (TAConv) and texture-aware dynamic normalization (TADN) to extract multi-scale structural information and enhance high-frequency details, effectively improving the local structural recovery capability of key components such as conductors, insulators, and hardware in transmission line inspection images. In the BFEFM module, the Blind Super-Digital Degradation Adaptive Channel Weighting (BDACW) method dynamically adjusts the channel response according to different degradation conditions such as rain, fog, compression artifacts, and motion blur in the transmission line inspection images, suppressing noise and enhancing effective details. In the IBSRNet network, the Blind Super-Digital Progressive Upsampling (BPUP) method improves edge sharpness and local recognizability while maintaining the natural transition of transmission line inspection images through neighborhood weighting and high-frequency compensation.

[0026] S211, Design a Blind-SR Feature Perception Module (BFPM) to achieve hierarchical modeling of input features of transmission line inspection images through a multi-branch parallel structure and cross-scale feature interaction; In the BFPM module, texture-aware frequency domain convolution (TAConv) is employed to simultaneously capture multi-scale texture features in transmission line inspection images in both spatial and frequency domains. Addressing the challenges of long-distance shooting, low resolution, and weakened details of key targets such as conductors, insulators, and hardware, texture-aware dynamic normalization (TADN) dynamically adjusts high-frequency weights, significantly improving IBSRNet's ability to perceive low-level details and global texture, as well as its generalization performance. The structure of the BFPM module is as follows: Figure 3 As shown.

[0027] S2111, input the blind super-resolution reconstruction feature map P1 of the transmission line inspection image into the CTM (Conv+TADN+Mish) module for basic feature extraction to obtain the blind super-resolution reconstruction feature map P2 of the transmission line inspection image. S2112, perform TAConv(3x3) operation on P2 to obtain blind super-resolution reconstruction feature map P3 of transmission line inspection image; S2113, perform TAConv(5x5) operation on P2 to obtain blind super-resolution reconstruction feature map P4 of transmission line inspection image; S2114, perform TAConv(7x7) operation on P2 to obtain blind super-resolution reconstruction feature map P5 of transmission line inspection image; S2115, perform TAConv(9x9) operation on P2 to obtain the blind super-resolution reconstruction feature map P6 of the transmission line inspection image.

[0028] S2116, perform SE (Squeeze-and-Excitation) operations on P3, P4, P5, and P6 respectively to obtain P3', P4', P5', and P6'; then perform Add operations on P3', P4', P5', and P6' to obtain the blind super-resolution reconstruction feature map P7 of the transmission line inspection image; S2117, perform DS (downsampling) operation on P7, followed by Conv (3x3) operation to obtain blind super-resolution reconstruction feature map P8 of the transmission line inspection image; S2118, perform US (upsampling) operation on P7, followed by Conv (3x3) operation to obtain blind super-resolution reconstruction feature map P9 of transmission line inspection image; S2119, input P8 and P9 into the CTS (Conv+TADN+SE) module for deep feature extraction to obtain the blind super-resolution reconstruction feature map P10 of the transmission line inspection image; S21190, perform an Add operation on P7 and P10 to obtain the final output blind super-resolution reconstruction feature map P11 of the transmission line inspection image.

[0029] The CTM (Conv+TADN+Mish) module is a feature processing unit consisting of a convolutional layer (Conv), a texture-aware dynamic normalization layer (TADN), and a Mish activation function.

[0030] The CTS (Conv+TADN+SE) module is a feature processing unit consisting of a convolutional layer (Conv), a texture-aware dynamic normalization layer (TADN), and a channel attention mechanism (SE).

[0031] Here, Mish is a commonly used activation function in deep learning, defined as follows:

[0032] Squeeze-and-Excitation (SE) is a commonly used channel attention mechanism in deep learning.

[0033] This invention designs a texture-aware frequency convolution method, TAConv (Texture-Aware Frequency Convolution), which performs multi-scale convolution operations in the spatial and frequency domains on the features of the input image in the module. It enhances the high-frequency texture information of the image by adaptively learning frequency domain filters, and at the same time, combines spatial domain convolution to achieve dynamic fusion of local details and global structure, thereby improving the module's ability to perceive the distribution of small targets and complex structures in power transmission line inspection images.

[0034] Let the convolutional input feature map of the power transmission line inspection image be... The convolutional output feature map of the power transmission line inspection image is ,in This represents a 3D feature map, where the sizes of the three dimensions are respectively... , , , This represents a 3D feature map, where the sizes of the three dimensions are respectively... , , The three dimensions represent the number of channels, height, and width of the feature map, respectively; the execution flow of the TAConv method is as follows: A1. Yes Perform a Fourier transform on each channel to obtain the Fourier domain features. The purpose of Fourier transform is to map spatial domain information to frequency domain representation in order to separate high-frequency textures and low-frequency components in transmission line inspection images. (2) Spatial domain information refers to the direct representation of a transmission line inspection image in pixel coordinate space, that is, the variation of the value of each pixel with its position. High-frequency texture refers to the rapidly changing parts of the transmission line inspection image that contain a lot of details, such as fine textures, edge contours, or metallic reflections; low-frequency components refer to the parts of the transmission line inspection image that change gently and have a clear overall structure, such as the sky background or the ground.

[0035] A2. Yes Perform frequency domain convolution to obtain frequency domain features. Frequency domain convolution can enhance the high-frequency texture information of images, enabling the module to better capture the texture details of small targets and complex structures in power transmission line inspection images. (3) Frequency domain convolution is a convolution operation performed on the frequency plane after the original feature map has undergone Fourier transform. In simpler terms, it weights different frequency components in the frequency domain, enhancing high-frequency texture information and suppressing low-frequency smoothing information.

[0036] A3. Yes Perform spatial domain convolution to obtain spatial domain features. Spatial domain convolution can extract spatial structure information, enabling convolutional features to capture local details and overall structural distribution, providing spatial domain supplementation for high-frequency features enhanced in the frequency domain, achieving effective complementarity between frequency domain and spatial domain features, thereby improving the ability to perceive small targets and complex structural distributions. (4) Spatial domain convolution is a convolution operation performed on the original feature map, which is a weighted sum of pixels or feature values ​​in the local neighborhood. It is mainly used to extract spatial structure information of the image.

[0037] A4. Yes and Nonlinear fusion is performed to obtain the convolutional output feature map of the transmission line inspection image. By analyzing frequency domain features Inverse Fourier transform is performed to recover the spatial representation, and combined with weighted spatial domain features, dynamic fusion of high-frequency enhancement in the frequency domain and spatial local structure is achieved, thereby enabling... It simultaneously preserves high-frequency texture information of small targets and complex structures, as well as global structural information.

[0038] (5)

[0039] In equations (2) to (5), For Fourier transform, For inverse Fourier transform, and These are the convolution kernels for frequency domain convolution and spatial domain convolution, respectively. This represents the convolution operation. This is element-wise multiplication.

[0040] in: Let be the hyperbolic tangent function, and its formula is:

[0041] This invention designs a texture-aware dynamic normalization (TADN) method to perform high-frequency texture enhancement and dynamic normalization on input features in a module. It generates adaptive texture weights by calculating local high-frequency responses and combines batch mean and variance to perform weighted normalization on the features, thereby highlighting the detailed information of key targets such as wires and insulators, while maintaining the stability of the overall structure, and realizing dynamic adjustment and hierarchical modeling of input features.

[0042] Let the normalized input feature map of the transmission line inspection image be... The normalized output feature map of the transmission line inspection image is as follows: The execution flow of the TADN method is as follows: B1. Calculation mean With variance , and Used for measurement The average value and fluctuation of each channel feature are used to ensure that the model maintains consistent feature scale in different inspection scenarios, such as strong light, shadow, and complex backgrounds, which helps to highlight the details of key high-frequency structures such as conductors and insulators. (6) (7) B2. Calculate texture response weights When identifying high-texture areas containing key targets such as conductors and insulators in transmission line inspection images, This is used to adaptively adjust the normalization intensity at each pixel location, thereby enhancing details while maintaining background stability; First, calculate the local high-frequency response. : (8) pass The function will Mapping to the interval (0,1) yields the texture response weights. : (9) in: It is a commonly used nonlinear mapping function, and its formula is:

[0043] B3. Yes Weighted normalization is performed to obtain texture-aware normalized features. ; through texture response weights Adjusting the position of each pixel, normalized features are used to enhance details in high-response areas such as wire edges and insulator outlines, while original features are retained in low-response areas to maintain background stability, thereby achieving the highlighting of details of key targets and the balance of the overall structure. (10) B4. A linear transformation is performed to obtain the normalized output feature map of the transmission line inspection image. Through the analysis of Apply learnable scaling parameters and offset parameters This ensures that the enhanced detailed features can be fully utilized by the module, while maintaining the freedom and stability of feature expression, so as to highlight key targets such as conductors and insulators under different inspection conditions.

[0044] (11)

[0045] In equations (8) to (11), For Gaussian blur operation, It is a safety factor. and These are the learnable scaling parameters and the learnable translation parameters, respectively.

[0046] Among them: Learnable parameters: "Learable" means that the parameters are continuously updated and optimized during the network training process to reach the optimal value. The "learnable" appearing in other places in this invention has the same meaning.

[0047] S212, a Blind Feature Enhancement Fusion Module (BFEFM) is designed to adaptively adjust channel weights in IBSRNet according to the degradation degree of the input transmission line inspection image, while enhancing structural information and detail features respectively. The module automatically strengthens key channels through Blind Super-Resolution Degradation Adaptive Channel Weighting (BDACW) to suppress multiple degradation noises in the transmission line inspection image caused by UAV vibration, wind disturbance, rain and fog and compression.

[0048] The BFEFM module can dynamically adjust its feature response based on different ambiguity and noise conditions, locally enhancing small targets (such as conductors, insulators, and hardware) in sparse effective regions while maintaining the integrity of the global structure. The module structure is as follows: Figure 4 As shown.

[0049] S2121, Perform BDACW operation on the blind super-resolution reconstruction feature map F1 of the transmission line inspection image to obtain the blind super-resolution reconstruction feature map F2 of the transmission line inspection image. S2122, perform TACN (TAConv+Conv+TADN) operation on F2 to obtain the blind super-resolution reconstruction feature map F3 of the transmission line inspection image; S2123, perform TAHN (TAConv+HPConv+TADN) operation on F2 to obtain the blind super-resolution reconstruction feature map F4 of the transmission line inspection image; S2124, Perform BDACW operation on F3 to obtain the blind super-resolution reconstruction feature map F5 of the transmission line inspection image; S2125, perform BDACW operation on F4 to obtain the blind super-resolution reconstruction feature map F6 of the transmission line inspection image; S2126, perform Add operation on F5 and F6, then perform Conv(1x1) operation, and finally output the blind super-resolution reconstruction feature map F7 of the transmission line inspection image.

[0050] Among them, the high-pass convolution HPConv is specifically designed to extract high-frequency information from power line inspection images, including edges, textures, and details, while suppressing low-frequency smooth regions, thereby enhancing the network's ability to perceive image details. It is particularly suitable for degradation adaptation and detail restoration tasks.

[0051] Example: A fixed high-pass convolution kernel is used in the HPConv operation of this invention. :

[0052] Wherein: TACN (TAConv+Conv+TADN) refers to the feature processing operation consisting of texture-aware frequency domain convolution (TAConv), convolution (Conv), and texture-aware dynamic normalization (TADN) in sequence.

[0053] (TAConv+HPConv+TADN) refers to a feature processing operation consisting of texture-aware frequency domain convolution (TAConv), high-pass convolution (HPConv), and texture-aware dynamic normalization (TADN) in sequence.

[0054] This invention designs a blind degenerate adaptive channel weighting method (BDACW) for super-resolution, which dynamically adjusts channel weights based on the degradation level of the input image (such as blurring, noise, compression, etc.) to achieve adaptive enhancement of key features. This method combines the median and median deviation from robust statistics with the spectral energy extracted by multi-band high-pass convolution to comprehensively calculate the degradation sensitivity and structural response of channels, thereby generating degradation-aware weights. BDACW can flexibly adjust channel importance according to different degradation types, effectively suppressing noise and artifacts while enhancing key structures and details, thus improving the degradation adaptation capability of IBSRNet in complex inspection images.

[0055] Let the adaptive input feature map for the degradation of transmission line inspection images be... The adaptive output feature map for power transmission line inspection image degradation is as follows: The execution flow of the BDACW method is as follows: C1. Perform robust channel statistics to obtain the dimensionless instability index. ; It can capture the overall level and local fluctuations of channel features, enhancing key details in transmission line inspection images while suppressing abnormal noise and extreme biases. C11 to For each channel, the median is calculated at the sample and spatial location to obtain the channel median. : (12) This indicates the median calculation operation, which involves sorting a set of data by size and taking the median value as the statistical result. C12 pairs and The absolute value of the difference is taken and the median is calculated to obtain the median absolute deviation. : (13) C13 to After normalization, a dimensionless instability index is obtained. : (14) Among them: robust channel: refers to a channel that is not sensitive to outliers or noise in the feature map.

[0056] C2. Perform learnable multi-band spectrum energy calculation to obtain the total spectrum energy. ; It can reflect the activity level of a channel in different frequency bands, providing a quantitative basis for channel degradation sensitivity, thereby enhancing the feature edges and details of transmission line inspection images while suppressing high-frequency noise; the specific steps for learning multi-band spectral energy calculation are as follows: C21 will Each channel is associated with its corresponding convolution kernel. Perform channel-wise convolution to obtain the convolution response. : (15) In the formula, This indicates a channel-wise convolution operation, i.e., the convolution kernel... copy portion, make Each channel has a corresponding convolutional kernel, and then Each channel is convolved with the corresponding convolution kernel; In equation (15) These represent three convolutional kernels with different frequencies used to process the image. The definition is as follows: : A low-frequency kernel, 3×3 in size, is used to capture smoothing information from the blind super-resolution reconstruction feature map of transmission line inspection images. The intermediate frequency kernel, with a size of 3×3, is used to capture the edge and structural information of the blind super-resolution reconstruction feature map of transmission line inspection images. A high-frequency kernel, 5×5 in size, is used to capture details and noise information in the blind super-resolution reconstruction feature map of transmission line inspection images.

[0057] C22 Convolution Response By taking the absolute value and calculating the global average, the average spectral energy at this scale is obtained. : (16) C23 for all scales By summing the results, we obtain the total energy of the spectrum. : (17) C3. Calculate the full-map spectral baseline and degradation index to obtain the spectral drop index. and image-level degradation intensity ; and It can measure the degradation degree of each channel in a transmission line inspection image at multiple scales, thereby guiding channel weighting to highlight key structures while suppressing artifacts and enhancing IBSRNet's adaptability to complex degraded images. The specific steps for calculating the full-image spectral benchmark and degradation index are as follows: C31 Average spectral energy at each scale The average spectral baseline for the entire image at this scale is obtained by averaging all channels. Then, summate all the scales to obtain the overall baseline. : (18) C32 calculates the spectral drop at each scale. : (19) Spectrum drop refers to, The average spectral energy of a certain channel at a certain scale Relative to the full-map average spectral benchmark at this scale The decrease.

[0058] C33 merges all scales And normalized to obtain the spectrum drop index : (20) C34 according to Calculate image-level degradation intensity : (twenty one) Image-level degradation intensity It is a quantitative indicator used to measure the overall degradation level of a complete power transmission line inspection image. The closer the value is to 1, the more severe the image degradation. In equations (20) to (21), It is a constant, and its value is... , The learnable scaling factor; C4. Calculate the degradation perception channel score and weight to obtain the degradation adaptive output features of the transmission line inspection image. Channel weights can adaptively amplify important feature channels in transmission line inspection images, suppress severely degraded channels and noisy channels, thereby enhancing key details, suppressing artifacts, and improving the model's degradation adaptation capability. The specific execution steps for degradation-aware channel scores and weights are as follows: C41 will and Weighted proportionally, and with Linear combination, with bias added. Obtain the original channel score. : (twenty two) C42 performs degradation gating correction by... By gate coefficient Adjustments are made to obtain the corrected channel score. : (twenty three) in The gating coefficient; C43 Perform temperature scaling, and through The function will Mapping to the interval (0,1) yields the final channel weights. : (twenty four) This is a temperature coefficient used to control the sharpness of the weight distribution; C44 will and Multiplying these together yields the adaptive output feature map of the power transmission line inspection image degradation. : (25) in, These are learnable parameters.

[0059] S213. A blind super-resolution progressive upsampling (BPUP) method is designed. By determining the low-resolution neighborhood corresponding to the high-resolution pixel, calculating the neighborhood mean, and calculating the weighting coefficient based on distance attenuation, the neighboring pixels are enhanced with high frequency weighting. This enables detail enhancement and edge fidelity preservation in the blind super-resolution scenario of power transmission line inspection images, improves the local structure restoration capability and high-frequency detail performance of the generated image, and maintains the smoothness and naturalness of the overall image.

[0060] Let the input image for low-resolution transmission line inspection be... High-resolution transmission line inspection output images are ,in To represent a 3D image, the sizes of the three dimensions involved are as follows: , , To represent a 3D image, the sizes of the three dimensions involved are as follows: , The three dimensions represent the image's height, width, and number of channels, respectively; the upsampling process of the BPUP method is as follows: D1. For high-resolution pixels Calculate the corresponding low-resolution neighborhood coordinate set. ; (26)

[0062] Example: Assuming high-resolution pixels The corresponding set of domain coordinates is:

[0063] D2. Calculate the neighborhood mean : For the set of domain coordinates The average value of all pixels is obtained from this, which can provide a reference benchmark for high-frequency enhancement, enabling the upsampling process to highlight the local details of the transmission line inspection image without destroying the overall smoothness. (27) Represents the set of domain coordinates The total number of pixels.

[0064] D3. For each neighboring pixel Calculate the weighting coefficients This enables local spatial adaptive processing of high-resolution pixels while also taking into account... Enhanced detail and edge fidelity; (28) It is an exponential function; It is a learnable distance attenuation coefficient used to control the intensity of distance attenuation; In express Pixel coordinates within; D4. Combining the neighborhood mean and weighting coefficients, for each high-resolution pixel Weighted calculations are performed and high-frequency enhancement is applied by adding high-frequency enhancement terms. By magnifying local texture details and edge information, the overall smoothness and naturalness of the power transmission line inspection images are maintained, while improving the local structure restoration capability and high-frequency detail performance. (29) These are learnable high-frequency enhancement coefficients used to control the intensity of high-frequency enhancement; D5. Generate complete high-resolution transmission line inspection output images. .

[0066] For all high-resolution pixels Repeat the above steps, assigning the weighted result of each pixel to the corresponding position in sequence, to obtain the complete high-resolution transmission line inspection output image. .

[0067] S214. Based on the BFPM module, PFEFM module, and BPUP method, a blind super-resolution reconstruction network (IBSRNet) for transmission line inspection images is constructed to receive standardized transmission line inspection images. As input, a preliminary super-resolution image robust to unknown degradation scenarios is generated. ; structure as Figure 2 As shown.

[0068] S2141, standardized transmission line inspection images Inputting into the BFPM layer yields the blind super-resolution reconstruction feature map B1 of the first transmission line inspection image; S2142, input B1 into the BFEFM layer to obtain the blind super-resolution reconstruction feature map B2 of the second transmission line inspection image; S2143, input B2 into the BFPM layer, and then perform BPUP upsampling to obtain the blind super-resolution reconstruction feature map B3 of the third transmission line inspection image; S2144, input B3 into the BFEFM layer, and then perform BPUP upsampling to obtain the blind super-resolution reconstruction feature map B4 of the fourth transmission line inspection image; S2145, input B4 into the CONV layer to obtain the final preliminary super-resolution image. .

[0069] S22, through the regional consistency sensing unit... , Local inconsistency region detection is performed to obtain a consistency heatmap for image reconstruction during the inspection. ; In step S22, the specific execution steps of the region consistency awareness unit are as follows: S221, for Two BPUP upsampling operations were performed to obtain the upsampled transmission line inspection image. ; S222, for Local error calculation is performed to obtain the local error feature map for the reconstruction of the inspection image. ; (30) These are the updatable parameters for the regional consistency-aware unit. This represents a four-dimensional tensor, where the sizes of the four dimensions are respectively... , , ; It is a local adaptive gain function. In the referential expression (30) This is used to highlight significant differences in transmission line inspection images, and is defined as follows: (31) in, For local gradient magnitude, , and It is a Sobel convolution kernel with horizontal and vertical gradients; This is the local gain enhancement coefficient, which controls the enhancement intensity; This is a local gain smoothing coefficient used to prevent division by zero and smooth low-gradient regions.

[0070] Among them, the Sobel convolution kernel is a convolution kernel commonly used in deep learning to calculate image gradient information.

[0071] S223, for Perform channel averaging and apply Function to generate a consistent heatmap for inspection image reconstruction ; (31) MeanChannel represents the averaging operation over the channels. express function.

[0072] S23, using a conditional reflexive correction unit to... and Self-reflexive correction is performed in the mid-deviation area to output a corrected version of the transmission line inspection image. and the confidence level of local correction in the reconstruction of inspection images ; In S23, the specific execution steps of the conditional reflexive correction unit are as follows: S231, for and Conditional feature modulation is performed to obtain the correction features for the reconstructed deviation region of the inspection image. ; (33) These are updatable parameters for the conditional reflexive correction unit, used for adaptive adjustment. The deviation area This represents a four-dimensional tensor, where the sizes of the four dimensions are respectively... , , .

[0073] S232, perform local corrections to obtain the corrected transmission line inspection image. Then, confidence level calculation is performed to obtain the local correction confidence level for the reconstruction of the inspection image. ; (34) (35) S24, using global equilibrium units to , and Perform overall feature adjustments to generate adjusted transmission line inspection images. ; In step S24, the specific execution steps of the global balancing unit are as follows: S241, for , and Perform overall feature adjustment to obtain a globally balanced feature map for reconstructing the inspection image. ; (36) This allows for the adjustment of the overall characteristics of equation (36) to focus on Regions with large bias and low confidence; These are updatable parameters for global balancing units, used for feature adjustment.

[0074] S242, Generate the adjusted transmission line inspection image. ; (37) The balance enhancement coefficient is used to control... Enhancement intensity in the low-to-medium confidence region.

[0075] S25, through the credibility guidance unit... , , and Perform joint analysis and calculate The confidence weight corresponds to each pixel location, thereby generating a pending result image for blind super-resolution reconstruction of power transmission line inspection images. ; In step S25, the specific execution steps of the credibility guidance unit are as follows: S251, for Perform channel stitching to generate inspection image reconstruction confidence weights ; (38) in, This indicates that in the channel dimension To splice; These are updatable parameters of the credibility guidance unit, used to generate the credibility weights for each channel. This represents a four-dimensional tensor, where the sizes of the four dimensions are respectively... , , .

[0076] S252, combined with credibility weight and The primary and secondary relationships are used to generate undetermined results for blind super-resolution reconstruction of transmission line inspection images. .

[0077] (39)

[0078] exist During the generation process, with As the primary feature map to preserve the global correction effect, while using and for Supplementary information is provided for areas with lower confidence levels to ensure... Even in low-confidence regions, reliable features can be preserved.

[0079] S26, in the adaptive optimization unit, and A joint evaluation of consistency and credibility is performed. If the credibility is less than a threshold, inspection image reconstruction feedback is generated. Used to iteratively update the updatable parameters in the above units. , If the confidence level is greater than or equal to the threshold or the maximum number of iterations is reached, then... The output is a blind super-resolution reconstruction result image of a power transmission line inspection. ; In step S26, the adaptive optimization unit executes the following steps: S261, according to and Calculate the global consistency residual for image reconstruction during inspection ; (40) in, This refers to Replication in channel dimension Next, to with Alignment for The number of channels.

[0080] S262, perform a joint evaluation of consistency and credibility to obtain credibility. ; (41) S263, Determining Reliability With threshold The size relationship should be determined and processed accordingly; if < Generate inspection image reconstruction feedback Update the updatable parameters of each unit in S22–S25. , Number of iterations Increment by 1, and then continue execution from S22; this process is the adaptive update process. >= or , If the preset maximum number of iterations is used, then... Reset to 1, and at the same time set to... The output is a blind super-resolution reconstruction result image of a power transmission line inspection. , that is to say The processing flow has ended; Inspection image reconstruction feedback The generation process is as follows: (42) in, This represents a four-dimensional tensor, where the sizes of the four dimensions are respectively... , , ; Iterate the convolutional layer to inspect the global features of the image.

[0081] Example: In the solution of the present invention, in The convolution kernel used is ; This represents a four-dimensional tensor, where the sizes of the four dimensions are respectively... , , .

[0082] Parameters in S22–S25 , The update process is as follows: Will Perform sequential splitting along the second dimension, resulting in dimensions that are respectively equal to... , of ,in Consistency perception feedback for reconstructed regions in inspection images; Reflexive correction feedback for image reconstruction conditions during inspection; To reconstruct global balanced feedback for inspection images; Feedback is provided to guide the confidence assessment of the reconstructed inspection images; then, parameter updates are performed. (43) For parameters The updated value, i.e. Reference , ; Reference .

[0083] Example

[0084] The network executed by the internal unit of the RMC-SR image processing chip was trained using a publicly available power line inspection image dataset to determine the learnable parameters and weights of the internal network for practical applications. The dataset was divided into two parts in a 7:3 ratio. The first part was used to train the internal network of the blind super-resolution generation unit. After training, the internal network of the blind super-resolution generation unit remained unchanged. The second part was then used to train the remaining five units. Threshold Set to 0.95, number of iterations Initialize to 1, maximum number of iterations Set it to 300.

[0085] Training begins by initializing the hyperparameters required for training the network, such as the batch size, number of training epochs, and initial learning rate. During training, the Adam algorithm is used to adjust the weights in the network to optimize the loss function, causing it to converge and ultimately bringing the network to equilibrium.

[0086] The RMC-SR image processing chip's six functional units are all implemented by Deep Processing Units (DPUs). The chip as a whole consists of a control core, on-chip storage, and the six DPU functional units. The six DPU units include: a blind super-resolution generation unit, a region consistency perception unit, a conditional reflexive correction unit, a global balancing unit, a confidence-guided unit, and an adaptive optimization unit. After training, the structure definition files of the internal networks of each unit, along with the learned parameters and weights obtained from training, are exported into a deployment format recognizable by the DPU (such as .xmodel or intermediate representation files). Subsequently, the network structure of each unit is fixed into the corresponding DPU logical region through the chip configuration interface, and the corresponding weight parameters are loaded into the on-chip storage of each DPU. After power-on, the chip's control core sequentially schedules the six DPU units to work according to the preset execution logic.

[0087] Assume the input is the original transmission line inspection image to be super-resolution in the blind. The dimensions are 3×256×256 (channels × width × height). The transmission line image analysis unit receives and standardizes the data. This yields a standardized transmission line inspection image with dimensions of 1×256×256. Blind super-resolution generation unit is used to process standardized transmission line inspection images. After processing, a preliminary super-resolution image with dimensions of 3×1024×1024 is obtained. Then, through the regional consistency sensing unit... , By performing local inconsistency region detection, a consistency heatmap of the inspection image with dimensions of 1×1024×1024 is obtained. Then, with the help of a conditional reflexive correction unit, and Self-reflexive correction is performed on the mid-deviation region to obtain a corrected transmission line inspection image with dimensions of 3×1024×1024. and the confidence level of local correction for reconstructing inspection images with dimensions of 1×1024×1024 Next, the global equilibrium unit is used to... , and Overall feature adjustment is performed to generate a transmission line inspection image with dimensions of 3×1024×1024. Then, the credibility guidance unit is used to... , , and Perform joint analysis and calculate The confidence weight corresponding to each pixel position generates a blind super-resolution reconstruction result image of a power line inspection image with a dimension of 3×1024×1024. Finally, in the adaptive optimization unit, and A joint evaluation of consistency and credibility is performed. If the credibility is less than the threshold of 0.95, the output is a 3×(1+3+3+10)×3×3 inspection image reconstruction feedback. Update the updatable parameters of the regional consistency awareness unit, conditional reflexive correction unit, global balancing unit, and credibility guidance unit. , Number of iterations Increment by 1, then continue execution starting from the region consistency-aware unit; if the confidence level is greater than or equal to the threshold of 0.95 or the number of iterations... If it is greater than or equal to 300, then Reset to 1, and set the current value to 1. The output is a blind super-resolution reconstruction result image of a power transmission line inspection. , that is to say The processing flow has ended.

[0088] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for blind super-resolution reconstruction of transmission line inspection images, characterized in that, The specific steps are as follows: S1, construct a transmission line image parsing unit to analyze the original transmission line inspection image LR to be analyzed using the super-resolution method. in Standardize; S2 uses the RMC-SR image processing chip to achieve blind super-resolution reconstruction of transmission line inspection images.

2. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Input transmission line inspection image Extract brightness information and generate a brightness image. Preserves the spatial structure and texture details of the image; S12, for The image edges are filled or cropped to generate a size-normalized image. This ensures that the input dimensions of subsequent processing units are consistent. S13, for Each pixel in Linear normalization is performed to obtain standardized transmission line inspection images. : （1） in and They represent The minimum and maximum pixels are determined to ensure that the normalized pixel values ​​are in the range [0, 1].

3. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 2, characterized in that, The specific steps of S2 are as follows: S21, using a blind super-resolution generation unit to process standardized transmission line inspection images Processing is performed to generate a preliminary super-resolution image. ; S22, through the regional consistency sensing unit... , Local inconsistency region detection is performed to obtain a consistency heatmap for image reconstruction during the inspection. ; S23, using a conditional reflexive correction unit to... and Self-reflexive correction is performed in the mid-deviation area to output a corrected version of the transmission line inspection image. and the confidence level of local correction in the reconstruction of inspection images ; S24, using global equilibrium units to , and Perform overall feature adjustments to generate adjusted transmission line inspection images. ; S25, through the credibility guidance unit... , , and Perform joint analysis and calculate The confidence weight corresponds to each pixel location, thereby generating a pending result image for blind super-resolution reconstruction of power transmission line inspection images. ; S26, in the adaptive optimization unit, and A joint evaluation of consistency and credibility is performed. If the credibility is less than a threshold, inspection image reconstruction feedback is generated. Used to iteratively update the updatable parameters in the above units. , ; If the confidence level is greater than or equal to the threshold or the maximum number of iterations is reached, The output is a blind super-resolution reconstruction result image of a power transmission line inspection. .

4. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 3, characterized in that, In step S21, the specific execution steps of the blind super-resolution generation unit are as follows: S211, Design a blind super-resolution feature perception module (BFPM) to achieve hierarchical modeling of input features of transmission line inspection images through a multi-branch parallel structure and cross-scale feature interaction; S212, a blind super-resolution feature enhancement fusion module BFEFM is designed to adaptively adjust the channel weights in IBSRNet according to the degradation degree of the input transmission line inspection image, while enhancing structural information and detail features respectively. S213, a blind super-resolution progressive upsampling method BPUP is designed. By determining the low-resolution neighborhood corresponding to the high-resolution pixel, calculating the neighborhood mean and calculating the weighting coefficient based on distance attenuation, the neighboring pixels are subjected to high-frequency enhancement weighting, thereby achieving detail enhancement and edge fidelity in the blind super-resolution scenario of power transmission line inspection images. S214, based on the BFPM module, PFEFM module, and BPUP method, constructs an IBSRNet blind super-resolution reconstruction network for transmission line inspection images, receiving standardized transmission line inspection images. As input, a preliminary super-resolution image robust to unknown degradation scenarios is generated. .

5. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 4, characterized in that, In step S22, the specific execution steps of the region consistency awareness unit are as follows: S221, for Two BPUP upsampling operations were performed to obtain the upsampled transmission line inspection image. ; S222, for Local error calculation is performed to obtain the local error feature map for the reconstruction of the inspection image. ; S223, for Perform channel averaging and apply Function to generate a consistent heatmap for inspection image reconstruction .

6. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 5, characterized in that, In S23, the specific execution steps of the conditional reflexive correction unit are as follows: S231, for and Conditional feature modulation is performed to obtain the correction features for the reconstructed deviation region of the inspection image. ; S232, perform local corrections to obtain the corrected transmission line inspection image. Then, confidence level calculation is performed to obtain the local correction confidence level for the reconstruction of the inspection image. .

7. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 6, characterized in that, In step S24, the specific execution steps of the global balancing unit are as follows: S241, for , and Perform overall feature adjustment to obtain a globally balanced feature map for reconstructing the inspection image. ; S242, Generate the adjusted transmission line inspection image. .

8. The method for blind super-resolution reconstruction of transmission line inspection images according to claim 7, characterized in that, In step S25, the specific execution steps of the credibility guidance unit are as follows: S251, for Perform channel stitching to generate inspection image reconstruction confidence weights ; S252, combined with credibility weight and The primary and secondary relationships are used to generate undetermined results for blind super-resolution reconstruction of transmission line inspection images. .

9. A method for blind super-resolution reconstruction of transmission line inspection images according to claim 8, characterized in that, In step S26, the adaptive optimization unit executes the following steps: S261, according to and Calculate the global consistency residual for image reconstruction during inspection ; S262, perform a joint evaluation of consistency and credibility to obtain credibility. ; S263, Determining Reliability With threshold The size relationship is determined and processed accordingly.

Images