A method and system for image quality enhancement

Abstract

The application discloses an image quality enhancement method and system, and relates to the technical field of intelligent detection of power systems, and comprises the following steps: inputting a pretreated high-voltage wire accident scene picture into a multi-scale convolutional neural network to obtain a feature map; inputting the feature map into a luminance branch, a contrast-texture branch and an electromagnetic-noise removal branch respectively to obtain feature maps corresponding to the branches; inputting the feature maps of all the branches into a trained gating network to obtain fusion weights of each branch; and performing weighted summation on the feature maps of each branch according to the fusion weights of each branch to obtain an enhanced image. The application can enhance the quality of high-voltage wire accident scene images in a targeted manner.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for power systems, and in particular to an image quality enhancement method and system. Background Technology

[0002] With the rapid development of the power industry, high-voltage power line accidents occur frequently, making rapid and accurate analysis of accident sites crucial. Drones and unmanned robots are widely used for inspection and photography of high-voltage power line accidents due to their flexibility and efficiency. However, the accident site environment is complex, and images captured by drones and unmanned robots are often affected by factors such as insufficient light, smoke, and strong electromagnetic interference, resulting in low image quality.

[0003] Two existing technologies are currently available. Technology one, patent CN112241765A, uses multi-scale convolution and attention mechanisms to quickly and efficiently obtain classification results while achieving high-precision verification results. Technology two, patent CN114863099A, uses multi-branch asymmetric convolution for cloud image segmentation, achieving good segmentation results even when daytime and nighttime cloud image data are imbalanced. However, these existing technologies do not perform adaptive image analysis for high-voltage scenarios, resulting in poor image processing quality. Therefore, how to specifically enhance the quality of high-voltage power line accident scene images is an important problem that urgently needs to be solved. Summary of the Invention

[0004] The present invention provides an image quality enhancement method and system, but it cannot specifically enhance the quality of images at high-voltage power line accident sites.

[0005] This invention provides an image quality enhancement method, comprising the following steps:

[0006] Obtain feature maps from images of high-voltage power line accident scenes;

[0007] The first feature map is obtained by separating the electric arc light from the background in the feature map using an asymmetric mapping function; the second feature map is obtained by enhancing the image contrast of the feature map using the CLAHE algorithm and gradient guidance; the third feature map is obtained by learning the noise distribution in the feature map using residual learning and using frequency domain masking for denoising and suppressing smoke artifacts.

[0008] Weights are assigned to the first, second, and third feature maps and then fused to obtain a quality-enhanced image.

[0009] Furthermore, before acquiring the feature map of the high-voltage power line accident scene image, the method further includes: preprocessing the captured high-voltage power line accident scene image;

[0010] The preprocessing steps include: converting the high-voltage power line accident scene images from the time domain to the frequency domain using Fourier transform, and filtering the frequency domain to remove noise;

[0011] The filtering result is converted from the frequency domain to the time domain using the inverse Fourier transform to obtain the filtered time-domain image.

[0012] The filtered temporal input is adaptively convolved, and the convolution kernel parameters are dynamically adjusted according to the noise intensity to suppress residual noise.

[0013] Furthermore, the quality-enhanced image is obtained using an image enhancement network, which includes:

[0014] The dynamic multi-task encoder is used to input pre-processed high-voltage power line accident scene images into the dynamic multi-task encoder to obtain feature maps of the high-voltage power line accident scene images.

[0015] The brightness branch, contrast-texture branch, and electromagnetic-denoising branch are used to extract features from the feature maps of high-voltage power line accident scene images to obtain the first, second, and third feature maps, respectively.

[0016] A gating network is used to fuse the first, second, and third feature maps into the trained gating network to obtain a quality-enhanced image.

[0017] Furthermore, the step of inputting the preprocessed high-voltage power line accident scene image into a dynamic multi-task encoder to obtain a feature map of the high-voltage power line accident scene image includes the following specific steps:

[0018] Based on preprocessed high-voltage power line accident scene images, illumination and noise distribution features are extracted using a lightweight convolutional neural network;

[0019] Dynamically adjust the weights of channel attention and spatial attention to enhance fault regions; adjust the feature intensity of each channel through multiplication operations based on channel attention; highlight fault regions through element-wise multiplication or weighted summation based on spatial attention.

[0020] Obtain feature maps of high-voltage power line accident scene images.

[0021] Furthermore, the specific training steps for the trained gating network include:

[0022] Obtain images of high-voltage power line accident scenes under different accident types;

[0023] By training a gating network with high-voltage power line accident scene images under different accident types, the weights of the brightness branch, contrast-texture branch, and electromagnetic-denoising branch in the fusion process are automatically adjusted. Based on the fusion weight of each branch, the feature maps of each branch are weighted and summed to obtain the optimal fusion strategy under different accident types.

[0024] This invention provides an image quality enhancement system, comprising:

[0025] The image acquisition module is used to acquire feature maps of high-voltage power line accident scene images;

[0026] The branch extraction module is used to separate the electric arc intensity light from the background in the feature map through an asymmetric mapping function to obtain the first feature map; the image contrast of the feature map is enhanced by the CLAHE algorithm and gradient guidance to obtain the second feature map; the noise distribution in the feature map is learned by residual learning and denoised and smoke artifacts are suppressed by frequency domain masking to obtain the third feature map.

[0027] The image enhancement module is used to assign weights to the first, second, and third feature maps and fuse them to obtain a quality-enhanced image.

[0028] This invention provides an image quality enhancement method and system, which, compared with the prior art, have the following advantages:

[0029] The first feature map is obtained by separating the electric arc light from the background in the feature map using an asymmetric mapping function; the second feature map is obtained by enhancing the image contrast of the feature map using the CLAHE algorithm and gradient-guided convolution; the third feature map is obtained by learning the noise distribution in the feature map using residual learning and performing noise reduction and smoke artifact suppression using frequency domain masking; then, the first, second, and third feature maps are weighted and fused to obtain a quality-enhanced image; this quality-enhanced image combines the three prominent feature maps and selectively acquires the most important accident scene images during the weight allocation and fusion process. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the system architecture provided in an embodiment of the present invention;

[0031] Figure 2 A flowchart of electromagnetic noise modeling and suppression provided for embodiments of the present invention;

[0032] Figure 3 A flowchart of a dynamic multi-task encoder provided in an embodiment of the present invention;

[0033] Figure 4 The flowchart of the multi-branch collaborative encoder provided in the embodiments of the present invention. Detailed Implementation

[0034] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0035] See Figure 1 This invention provides an image quality enhancement method, comprising the following steps:

[0036] Step 1: Obtain feature maps of high-voltage power line accident scene images.

[0037] Step 2: Separate the electric arc light from the background in the feature map using an asymmetric mapping function to obtain the first feature map; enhance the image contrast of the feature map using the CLAHE algorithm and gradient guidance to obtain the second feature map; learn the noise distribution in the feature map using residual learning and perform noise reduction and smoke artifact suppression using a frequency domain mask to obtain the third feature map.

[0038] Step 3: Assign weights to the first, second, and third feature maps and fuse them to obtain a quality-enhanced image.

[0039] This invention presents a low-quality image enhancement method for high-voltage power line accident sites in complex environments (such as strong electromagnetic interference, dynamic lighting, and smoke / flame noise). It is applicable to real-time processing of power accident images acquired by drones, inspection robots, and other equipment, providing high-precision visual support for arc detection, conductor break location, and emergency repairs. This invention proposes a system that integrates electromagnetic noise suppression, dynamic attention allocation, and multi-task collaborative enhancement, addressing key challenges in high-voltage accident image enhancement.

[0040] The technical solution of this invention is explained in detail below:

[0041] 1. Electromagnetic noise suppression front end:

[0042] like Figure 2 As shown, it includes two parts: a frequency domain filtering module and an adaptive convolution kernel. The frequency domain filtering module detects periodic stripe noise through Fourier transform and generates a frequency domain mask for filtering. The adaptive convolution kernel dynamically adjusts the convolution kernel parameters according to the noise intensity to suppress residual noise.

[0043] 2. Dynamic multi-task encoder:

[0044] like Figure 3As shown, through this dynamic attention adjustment process, the system can adaptively optimize image enhancement effects based on image content and scene conditions, resulting in better presentation of key areas and details. This mainly includes the following four modules:

[0045] (1) Scene-aware attention module:

[0046] This module dynamically adjusts the channel and spatial attention weights based on the illumination / noise distribution (such as arc intensity and smoke concentration) of the input image.

[0047] Prioritize enhancing faulty areas (such as breakpoints) for use by subsequent decoders.

[0048] (2) Extraction of illumination / noise distribution features:

[0049] A lightweight convolutional neural network is used to perform preliminary feature extraction on the input image. The multi-scale feature extraction layers are parallel 1×1, 3×3, and 5×5 convolutions, focusing on key factors such as light intensity, arc intensity, and smoke concentration.

[0050] These features are compressed into vectors using pooling layers (such as global average pooling) for subsequent attention weight calculation.

[0051] (3) Calculation of dynamic attention weights:

[0052] Channel attention: Attention weights are calculated for each channel (e.g., RGB three channels). A sigmoid activation function is used to perform a non-linear transformation on the channel feature vectors to obtain the weight value for each channel. These weight values ​​reflect the importance of different channels in the enhancement process.

[0053] Spatial Attention: In the spatial dimension, a spatial attention map is generated using convolutional operations (such as 3x3 convolution). This map determines which regions require more attention (such as fault regions) by calculating the feature response at each spatial location. Normalization is also performed using the Sigmoid activation function.

[0054] (4) Application of attention:

[0055] The calculated channel attention and spatial attention weights are applied to the feature map respectively. Channel attention adjusts the feature intensity of each channel through multiplication operations, while spatial attention highlights key regions through element-wise multiplication or weighted summation.

[0056] 3. Multi-branch collaborative decoder:

[0057] like Figure 4As shown, through this dynamic allocation process, the gating network can adaptively optimize the fusion strategy of each branch according to the content of the input image and scene conditions, thereby improving the image enhancement effect and robustness. This mainly includes the following parts:

[0058] (1) Gated network receives input:

[0059] The gated network receives output feature maps from the brightness enhancement branch, the contrast-texture branch, and the electromagnetic-denoising branch.

[0060] (2) Learning the optimal fusion strategy:

[0061] Gated networks learn optimal fusion strategies under different accident types (such as arc discharge and wire breakage) through training. This means that it can automatically adjust the importance of each branch in the fusion process based on the specific content of the input image.

[0062] Brightness Enhancement Branch: Introduces an asymmetric mapping function to separate the intense arc light from the background, avoiding overexposure. Contrast-Texture Branch: Integrates the CLAHE algorithm with gradient-guided convolution to enhance the edges of the conductors. Electromagnetic-Denoising Branch: Combines residual learning and frequency domain masking to suppress residual electromagnetic noise and smoke artifacts.

[0063] The specific training and learning process is as follows: end-to-end scene awareness training, the gating network gradually masters the strategy of dynamically adjusting branch weights according to the scene through multi-task supervised learning and curriculum-based training strategies.

[0064] Step 1: Dataset Construction (Scene-Driven Multimodal Labeling).

[0065] 1) Raw data: Collect images of power accident sites, covering: Arc discharge: strong light flares, metal vapor clouds, conductor melting marks; Conductor breakage: flying debris, arc traces, insulator damage; Equipment short circuit: smoke, electromagnetic interference, local overheating marks.

[0066] 2) Annotation System: Professional Enhancement Images: Annotated by professionals to reflect the ideal enhancement effect, including three dimensions of scoring: brightness evenness, texture clarity, and noise level. Scene Tags: Each image is labeled with the accident type (e.g., arc discharge, wire breakage) and severity (mild / moderate / severe).

[0067] Step 2: Loss function design (multi-objective collaborative optimization).

[0068] The training loss of a gating network consists of three parts that collectively guide the learning of weights:

[0069] 1) Content-aware loss (Loss_content): Uses the feature space constraints of a pre-trained VGG network to force the output image to align with expert annotations at the semantic level. .

[0070] Where Φ is the VGG feature extractor, which ensures that the enhanced image is consistent with the annotation in terms of texture, edge and other details.

[0071] 2) Branch-specific loss (L_branch):

[0072] Apply L1 constraints to the output of each branch to prevent skill degradation: .

[0073] For example, in an arc scenario, the output of the luminance branch should be close to the luminance map annotated by experts, while the contrast branch should retain the details of the wire edges.

[0074] 3) Gating regularization term (L_gate):

[0075] Introduce scenario-prior constraint weight distribution; for example, in an arc discharge scenario, the brightness branch weight should be dominant. .

[0076] in For scene-specific weights (such as in the case of an electric arc scene) ), where λ is the regularization intensity coefficient.

[0077] Step 3: Course-based training strategy (progressive ability improvement).

[0078] 1) Phase One (Pre-training):

[0079] With fixed gating network weights, each branch network is optimized individually. Brightness branch: Trains an asymmetric mapping function to separate the arc's intense light from the background. Contrast branch: Adaptively adjusts the CLAHE parameters to enhance the conductor edges. Electromagnetic denoising branch: Learns residual mappings to suppress electromagnetic noise.

[0080] 2) Phase Two (Joint Training):

[0081] Enabling a gated network and using a small learning rate (1e-4) for end-to-end fine-tuning, the total loss is: .

[0082] 3) Phase Three (Combat Training):

[0083] A discriminator network is introduced to distinguish between real and fake images in the enhanced images, thereby improving the naturalness of the output and preventing overfitting.

[0084] (3) Dynamically generate fusion weights:

[0085] Based on the learned fusion strategy, the gating network dynamically generates a fusion weight for each branch. These weights reflect the contribution of each branch to image enhancement.

[0086] The process of dynamically generating fusion weights is as follows:

[0087] 1) Feature extraction: The gated network first receives the image features processed by three branches (brightness, contrast, and electromagnetic denoising). These features contain the different enhancement effects of each branch on the image.

[0088] 2) Understanding the scene: By analyzing these features, the gating network can determine which accident scene the current image belongs to (such as arc discharge or wire breakage).

[0089] 3) Calculate weights: Based on the identified scenario, the gating network assigns an importance score (weight) to each branch. For example, in an arc discharge scenario, the brightness branch, which handles strong light, will receive a higher weight.

[0090] 4) Adjusting weights: To make the weight distribution more reasonable, the network uses a "fine-tuning coefficient" to adjust these scores. In the early stages of training, the coefficient is larger, allowing the network to explore more possibilities; as training progresses, the coefficient gradually decreases, making the weight distribution more stable.

[0091] (4) Weight normalization:

[0092] To ensure the quality of the fused image, the gating network uses the Softmax function to normalize the generated fusion weights. This ensures that the sum of the weights of all branches is 1, guaranteeing the stability of the fusion process.

[0093] (5) Weighted fusion of the outputs of each branch:

[0094] Finally, the gating network performs a weighted sum of the outputs of each branch based on the normalized fusion weights. This yields the final enhanced image, which incorporates the strengths of each branch while avoiding parameter conflicts and redundancy.

[0095] 4. Lightweight real-time processing module:

[0096] During feature extraction, depthwise separable convolution is used to replace standard convolution to meet the real-time inference requirements of drones.

[0097] The key points and protection points of this invention are as follows:

[0098] 1. Electromagnetic noise modeling and suppression:

[0099] For the first time, frequency domain filtering and adaptive convolution are integrated in the preprocessing stage to specifically address stripe noise in high-voltage scenarios.

[0100] 2. Scene-adaptive dynamic attention mechanism:

[0101] The weights of each branch are automatically assigned based on the arc intensity and smoke concentration.

[0102] 3. Multi-task gating fusion:

[0103] Brightness, contrast, and noise reduction branches are dynamically fused through a gating network to avoid parameter conflicts and improve robustness in complex scenarios.

[0104] 4. Power Characteristic Enhancement Module:

[0105] Gradient-guided convolutional layers, specifically designed for conductors and insulators, enhance device edge features.

[0106] The advantages of this invention are:

[0107] 1. Noise suppression: Peak signal-to-noise ratio (PSNR) of stripe noise is improved by ≥8dB (traditional filtering methods only improve it by 3-4dB).

[0108] 2. Fault detection accuracy: Arc detection rate improved to 98.7%, wire breakage location error <0.5 pixels.

[0109] 3. Real-time performance: After lightweight design, the processing latency for 1080P images is <50ms (the latency for the U-Net enhanced model is >120ms).

[0110] 4. Scene adaptability: In the scenario where electric arc and smoke coexist, the Structural Similarity (SSIM) index is improved by 23% compared with the comparative patent model.

[0111] This invention provides an image quality enhancement system, comprising:

[0112] The image acquisition module is used to acquire feature maps of images from high-voltage power line accident scenes.

[0113] The branch extraction module is used to separate the electric arc light and background in the feature map through an asymmetric mapping function to obtain the first feature map; the image contrast of the feature map is enhanced by the CLAHE algorithm and gradient guidance to obtain the second feature map; the noise distribution in the feature map is learned by residual learning and denoised and smoke artifacts are suppressed by frequency domain masking to obtain the third feature map.

[0114] The image enhancement module is used to assign weights to the first, second, and third feature maps and fuse them to obtain a quality-enhanced image.

[0115] A specific example is as follows:

[0116] This embodiment discloses an image quality enhancement method, the specific steps of which are as follows:

[0117] S1, Enhanced Arc Discharge Scene:

[0118] Input: An image of an electric arc taken by a drone (overexposed by strong light + electromagnetic stripe noise).

[0119] Processing flow: electromagnetic front-end noise suppression (frequency domain filtering + adaptive convolution); dynamic attention focusing on discharge points (channel weights increased to 0.9); brightness branch separation of arc light, contrast branch enhancement of conductor texture.

[0120] Results: The arc outline is clear, the details of the background equipment are discernible, and the fault location time is reduced by 60%.

[0121] S2, Enhanced Nighttime Smoke Environment:

[0122] Input: Image of a broken wire obscured by smoke at night.

[0123] Processing flow: Denoising branch suppresses smoke (residual learning + frequency domain masking); Dynamic attention enhances broken edges (spatial weight increased to 0.85); Gated network fuses multi-branch results (brightness weight 0.3, contrast weight 0.6).

[0124] Results: The accuracy of fracture point localization reaches 99%, and the misjudgment rate is less than 1%.

[0125] 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. An image quality enhancement method, characterized in that, Includes the following steps: Obtain feature maps from images of high-voltage power line accident scenes; The first feature map is obtained by separating the electric arc light from the background in the feature map using an asymmetric mapping function; the second feature map is obtained by enhancing the image contrast of the feature map using the CLAHE algorithm and gradient guidance; the third feature map is obtained by learning the noise distribution in the feature map using residual learning and using frequency domain masking for denoising and suppressing smoke artifacts. Weights are assigned to the first, second, and third feature maps and then fused to obtain an enhanced image. The enhanced image is obtained using an image enhancement network, which includes: The dynamic multi-task encoder is used to input pre-processed high-voltage power line accident scene images into the dynamic multi-task encoder to obtain feature maps of the high-voltage power line accident scene images. The brightness branch, contrast-texture branch, and electromagnetic-denoising branch are used to extract features from the feature maps of high-voltage power line accident scene images to obtain the first, second, and third feature maps, respectively. A gating network is used to fuse the first, second, and third feature maps into the trained gating network to obtain a quality-enhanced image. The specific training steps for the trained gating network include: Obtain images of high-voltage power line accident scenes under different accident types; By training a gating network with high-voltage power line accident scene images under different accident types, the weights of the brightness branch, contrast-texture branch, and electromagnetic-denoising branch in the fusion process are automatically adjusted. Based on the fusion weight of each branch, the feature maps of each branch are weighted and summed to obtain the optimal fusion strategy under different accident types.

2. The image quality enhancement method as described in claim 1, characterized in that, Before obtaining the feature map of the high-voltage power line accident scene image, the method further includes: preprocessing the captured high-voltage power line accident scene image; The preprocessing steps include: converting the high-voltage power line accident scene images from the time domain to the frequency domain using Fourier transform, and filtering the frequency domain to remove noise; The filtering result is converted from the frequency domain to the time domain using the inverse Fourier transform to obtain the filtered time-domain image. The filtered temporal input is adaptively convolved, and the convolution kernel parameters are dynamically adjusted according to the noise intensity to suppress residual noise.

3. The image quality enhancement method as described in claim 1, characterized in that, The feature map obtained from the high-voltage power line accident scene image specifically includes: Based on preprocessed high-voltage power line accident scene images, illumination and noise distribution features are extracted using a lightweight convolutional neural network; Dynamically adjust the weights of channel attention and spatial attention to enhance fault regions; adjust the feature intensity of each channel through multiplication operations based on channel attention; highlight fault regions through element-wise multiplication or weighted summation based on spatial attention. Obtain feature maps of high-voltage power line accident scene images.

4. An image quality enhancement system, characterized in that, include: The image acquisition module is used to acquire feature maps of high-voltage power line accident scene images; The branch extraction module is used to separate the electric arc intensity light from the background in the feature map through an asymmetric mapping function to obtain the first feature map; and to enhance the image contrast of the feature map through the CLAHE algorithm and gradient guidance to obtain the second feature map. The noise distribution in the residual feature map is learned and denoised and smoke artifacts are suppressed by frequency domain masking to obtain the third feature map; The image enhancement module assigns weights to the first, second, and third feature maps and fuses them to obtain a quality-enhanced image. The image enhancement module obtains the enhanced image using an image enhancement network, which includes: The dynamic multi-task encoder is used to input pre-processed high-voltage power line accident scene images into the dynamic multi-task encoder to obtain feature maps of the high-voltage power line accident scene images. The brightness branch, contrast-texture branch, and electromagnetic-denoising branch are used to extract features from the feature maps of high-voltage power line accident scene images to obtain the first, second, and third feature maps, respectively. A gating network is used to fuse the first, second, and third feature maps into the trained gating network to obtain a quality-enhanced image. The specific training steps for the trained gating network include: Obtain images of high-voltage power line accident scenes under different accident types; By training a gating network with high-voltage power line accident scene images under different accident types, the weights of the brightness branch, contrast-texture branch, and electromagnetic-denoising branch in the fusion process are automatically adjusted. Based on the fusion weight of each branch, the feature maps of each branch are weighted and summed to obtain the optimal fusion strategy under different accident types.

Images