Cable lead seal defect image denoising and enhancing method and system, and storage medium

Abstract

The invention relates to the technical field of image processing, in particular to a cable lead seal defect image denoising and enhancing method and system, and a storage medium, and the method comprises the steps: carrying out the multi-stage denoising processing, employing a U-Net architecture to separate the contour of a lead seal main body from background noise, and reserving defect features; two-way reconstruction processing: performing two-way reconstruction on a denoising result and an original noisy image by using a generative adversarial network, and enhancing defect edge continuity and texture authenticity; feature fusion processing, global structure modeling, local detail convolution extraction, multi-source feature fusion and defect contrast optimization. The system comprises a multi-stage denoising module, a double-path reconstruction module and a feature fusion module. A computer program for implementing the method is stored in the storage medium. According to the method, the problems of low defect contrast and fuzzy details under extreme noise in a traditional method are solved, and the precision and reliability of cable lead seal defect detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and system for denoising and enhancing a cable seal defect image, and a storage medium. Background Art

[0002] In the field of intelligent inspection of power equipment, cable seals are key components for ensuring the safe operation of power transmission and transformation systems. Accurately assessing their sealing status is directly related to grid reliability. However, the complex working conditions of industrial sites pose significant challenges to seal defect detection. On the one hand, high-frequency texture noise generated by the metal oxide layer and motion artifacts caused by mechanical vibrations create a complex interference. Existing denoising algorithms can easily misinterpret key features such as microcracks (<100μm) and indentation defects when suppressing noise. On the other hand, the non-uniform light reflection characteristics caused by the multi-material (copper / aluminum) surface of the seals cause a drop in defect contrast by over 60% in visible light imaging. Existing single-modality denoising methods struggle to meet the accuracy requirements for visual defect detection as required by the IEC62895-2017 standard.

[0003] Current mainstream defect detection systems rely heavily on manual visual inspection, which can lead to low efficiency (inspection of a single piece takes ≥ 3 minutes) and high rates of missed detection. While deep learning-based denoising methods (such as DnCNN and FFDNet) have made progress in general scenarios, their adaptability to specific industrial scenarios is significantly insufficient: 1. The noise distribution of the training data does not match that of the actual seal defects, resulting in a metal oxidation artifact removal rate of less than 40%. 2. The network structure lacks prior modeling of the seal's geometric features, resulting in secondary damage such as broken seal indentation strokes and distorted crack paths. More critically, under extreme operating conditions with a signal-to-noise ratio of ≤45dB (such as strong electromagnetic interference), the existing method's microcrack detection rate plummets to 72.3%, failing to meet the zero-tolerance requirement for hidden defects in UHV equipment.

[0004] In view of this, the present invention is proposed. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems existing in the prior art, the present invention provides a cable seal defect image denoising and enhancement method and system, and a storage medium. This method can solve the problem in the prior art of poor ability to identify and separate noise and defective parts in cable seal images.

[0006] To achieve the above object, the technical solution of the present invention is as follows: In a first aspect, a method for denoising and enhancing a cable seal defect image comprises: Multi-level denoising processing to denoise the noisy seal image, separate the main outline of the cable seal from the background noise and retain the defect features; Two-way reconstruction processing, inputting the original noisy image and the denoising result of the multi-level denoising processing into the generative adversarial network for two-way reconstruction, and enhancing the continuity of the defect edge and the authenticity of the surface texture through the adversarial training mechanism; Feature fusion processing, fusing the denoising results of the multi-level denoising processing, the dual-path reconstruction results of the dual-path reconstruction processing and the original noisy image, combining global structure modeling with local detail perception, suppressing residual noise and optimizing the contrast of the defect area, and outputting a denoised and enhanced image of the defect features.

[0007] Furthermore, the multi-stage denoising process includes: Denoising is performed using a deep neural network architecture comprising an encoder and a decoder; batch normalization is used to automatically adjust the input data distribution of each layer of the neural network; Use activation functions to learn activation thresholds for different features; Restore the feature map output by the encoder to the input image size; connect the feature details extracted by the encoder to the corresponding layer of the decoder to avoid image blur; The composite noise is estimated by the noise estimation function and a preliminary denoised image is output.

[0008] Furthermore, the deep neural network architecture is a U-Net architecture, and the encoder includes multiple layers of convolution, each layer using a preset size convolution kernel and step size to scan the image.

[0009] Furthermore, the denoising process is optimized using a loss function, wherein the loss function is a mean square error loss function, and the formula is:

[0010] in, is the noise estimation loss, is the normalization factor, where H and W are the height and width of the image respectively, is the double summation symbol, To estimate the noise at pixel location The pixel value at is the noisy image at pixel position The pixel value at For a flawless seal image at pixel position The pixel value at is the real noise.

[0011] Furthermore, the dual-path reconstruction process includes: Image-to-image conversion block: inputs the multi-level denoising results and the original noisy image, and splices them into a multi-channel feature map; the conversion block includes a generator and a discriminator; The generator uses a deep neural network to repair the image, and the objective function includes pixel-level loss, conditional adversarial loss, and edge-preserving loss; The discriminator compares the texture and edges of the repaired image and the real defect-free image through convolution operations.

[0012] Furthermore, the objective function formula is:

[0013] in, is the weight coefficient of each loss item, Pixel level loss, To fight against loss, Reserve loss for seal edge.

[0014] Furthermore, the feature fusion processing includes: Feature splitting: Split the input into global features and local features through 1×1 convolution; Global feature modeling: using Capture the overall structure of the seal; Local feature extraction: Details are extracted through 3×3 convolution, and feature transfer is optimized after splicing with global features; the loss function includes mean square error loss and defect contrast loss, the latter of which maximizes the difference between defects and background.

[0015] Furthermore, the feature fusion processing further includes: Adaptive weight allocation: The regional complexity is estimated by calculating the variance of the feature map and generating dynamic weight coefficients; the final image is generated through weighted fusion, where the weights of global features and local features are dynamically adjusted by the regional complexity.

[0016] On the other hand, the present invention further provides a cable seal defect image denoising and enhancement system for performing the above-mentioned cable seal defect image denoising and enhancement method, the system comprising: Multi-level denoising module, used to denoise noisy seal images, separate the main contour of the cable seal from the background noise and retain the defect features; A dual-path reconstruction module is used to input the original noisy image and the denoising result of the multi-stage denoising module into the generative adversarial network for dual-path reconstruction, thereby enhancing the continuity of defect edges and the authenticity of surface texture through an adversarial training mechanism; A feature fusion module is provided, which fuses the denoising results of the multi-level denoising module, the dual-path reconstruction results of the dual-path reconstruction module, and the original noisy image, combines global structure modeling with local detail perception, suppresses residual noise, optimizes the contrast of the defect area, and outputs a denoised and enhanced image of the defect features.

[0017] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for denoising and enhancing a cable seal defect image.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method, system, and storage medium for denoising and enhancing cable seal defect images. By preprocessing the image, the main contour of the cable seal and background noise are effectively separated, preliminarily eliminating noise interference in the image. The denoised result and the original noisy image are reconstructed in two ways, and the adversarial training mechanism is used to enhance the continuity of the defect edge and restore the true surface texture, preventing key details from being lost during the denoising process. Finally, the multi-source processing results are integrated, combining global structural modeling with local detail perception to suppress residual noise while enhancing the contrast of the defect area, ensuring that key features such as cracks and scratches are clearly discernible. This achieves a dynamic balance between noise suppression and detail preservation, effectively resolving the problem of defect deletion or blurring in extreme noise environments caused by traditional methods, and providing highly robust image enhancement support for automated quality inspection of cable seals. The method can be widely used in fields requiring high-precision image detection, such as power inspections and industrial equipment maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic flow chart of a cable seal defect image denoising and enhancement method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0021] It should be noted that, unless otherwise specifically stated, the relative arrangements of components and steps, and numerical expressions set forth in these embodiments should not be construed as limiting the scope of the present invention.

[0022] The following description of exemplary embodiments is merely illustrative and is not intended to limit the present invention, its application, or use in any sense. Technologies, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but to the extent applicable, such technologies, methods, and apparatuses should be considered part of this specification.

[0023] See Figure 1 , which is a flow chart of a cable seal defect image denoising and enhancement method proposed by the present invention. The cable seal defect image denoising and enhancement method may specifically include: S1. Multi-level denoising: Perform preliminary denoising on the noisy seal image, separate the main outline of the cable seal from the background noise, and retain the defect features; First input the noisy cable seal image ,in and are the height and width of the image, respectively. 3 represents the red, green, and blue channels of an RGB color image. Noise types include Gaussian noise and metallic reflective artifacts from the seal surface. The U-Net architecture is used, consisting of an encoder and a decoder.

[0024] The encoder contains six convolution layers, each using a 3×3 convolution kernel and a sliding window with a stride of 2 to scan the image. This means the image is slid across the cable seal image in 3×3 regions, with a stride of 2 each time. This allows for rapid coverage of the entire image and inspection for metal reflective noise or cracks in the cable itself. The specific steps include: S11. Batch Normalization (BN): It is used to automatically adjust the input data distribution of each layer of the neural network and maintain stability. For each feature channel in the input data batch, the specific formula is:

[0025] in, For input data, is the mean of the data, is the variance of the data, is a minimum constant, is the standardized eigenvalue;

[0026] in, is the scaling factor, is the translation factor, When processing seal photos under different conditions, first calculate the average reflective intensity of all seal images in the current batch, and remove the overexposed areas (high ) darken, underexposed areas (low ) brighten, by and Preserve the inherent characteristics of metal materials, thereby automatically adjusting the light and dark contrast of the photo, making the brightness of the cracked area consistent with that of normal metal, avoiding local overexposure or excessive darkness.

[0027] S12, PReLU activation: As an improved version of ReLU, the expression of the optimal activation threshold for learning different features is:

[0028] in, For input data, is a learnable negative slope parameter (usually initialized to 0.25), each feature channel has an independent When processing cable seal images, the metal reflective area directly outputs the original value To preserve bright light details, the split region (weak negative response) uses the learned slope Inhibition, that is This adaptive suppression can better preserve the subtle gradient of rust than the standard ReLU (direct return to zero), thereby judging the natural texture and crack defect area of ​​the lead seal in the image.

[0029] S13, transposed convolution: restore the feature map output by the encoder to the input image size, gradually recovering the spatial resolution. Deconvolution is performed on the low-dimensional feature map compressed by the encoder, and the feature map size is gradually expanded to provide a size-matched feature representation for subsequent image reconstruction. S14: Directly connect the metal texture details extracted from the encoder to the corresponding layers of the decoder to avoid blurring the restored image; for example, low-frequency features such as seal outlines and anti-counterfeiting indentations in the original image. Retain the original metal surface texture information captured by the encoder in the early convolutional layers and combine it with the deep semantic features of the decoder to improve the structural integrity of the denoised image; for example, continuous gradients in smooth metal areas and high-frequency details at crack edges.

[0030] The noise estimation function FNE is defined as:

[0031] in, is the U-Net network parameter, is the estimated composite noise, is a noisy seal image; The output of the preliminary denoised image is expressed as:

[0032] in, is the denoised cable seal image, is the input noisy cable seal image, is the estimated composite noise of the output; by subtracting the estimated noise from the noisy image, the preliminary denoising result that retains the outline of the seal body is obtained.

[0033] Mean square error (MSE) loss function, the specific formula is:

[0034] in, is the noise estimation loss, is the normalization factor, where H and W are the height and width of the image respectively, is the double summation symbol, To estimate the noise at pixel location The pixel value at is the noisy image at pixel position The pixel value at For a flawless seal image at pixel position The pixel value at is the real noise; in this step, the output signal is The seal outline is retained but some local reflective artifacts remain.

[0035] S2, two-way reconstruction processing: the original noisy image and the denoising result of step S1 are respectively input into the generative adversarial network for two-way reconstruction, and the continuity of the defect edge and the authenticity of the surface texture are enhanced through the adversarial training mechanism; S21, image to image conversion block, input signal output in step S14 , including the input original noise signal , the original noise image is spliced ​​with the image denoised in step S1, and the image denoised in step S1 is used as the preliminary denoised image; the image-to-image conversion block IT includes: a generator G and a discriminator D.

[0036] Among them, the generator G refers to: the input signal is The 6 channels are stitched together and the stitched image is output. The corresponding formula is:

[0037] in, is the spliced ​​6-channel feature map, is the channel dimension splicing operation, is the initial denoised image, is the original noisy image.

[0038] The tool for image repair is: U-Net structure with the same NE, and the corresponding processing formula is:

[0039] in, To generate a defect-free seal image, is the generator network function; the specific repair process is: first downsample through the encoder, analyze the distribution pattern of the noise area, and then redraw the metal texture based on the defect-free lead seal training data. During the redrawing process, the generator objective function modeling formula is:

[0040] in, is the weight coefficient of each loss item, Pixel level loss, To fight against loss, The pixel level is retained for the seal edge loss. The loss formula is:

[0041] in, is the restored image output by the generator, The generated image is a true, defect-free lead seal image. The pixel-level difference between the generated image and the true, defect-free image is constrained to ensure structural consistency.

[0042] The formula for the adversarial loss is:

[0043] in, is the discriminator network, is the mathematical expectation, The original noisy image is used as a conditional input; it forces the generator to produce textured metal surfaces that are difficult for the discriminator to distinguish.

[0044] The formula for the seal edge retention loss is:

[0045] in, The Sobel operator extracts horizontal and vertical gradients; ensures that the edge details of the generated image are consistent with the real image and prevents edge blurring. The output signal here looks like a flawless lead seal image. .

[0046] S22, the discriminator D uses a 4*4 convolution kernel with a step size of 2 to scan the repaired image, comparing the repaired image with the real defect-free seal image, and focusing on whether the texture of the repaired seal is consistent with the original seal texture, and whether the edge of the seal is blurred or over-smoothed, resulting in distortion. The corresponding formula is:

[0047] After that, the discriminator’s judgment function and the generator G finally output a more realistic and defect-free lead seal image after learning. .

[0048] S3, feature fusion processing: Fusion of the denoising result of step S1, the dual-path reconstruction result of step S2, and the original noisy image, combining global structure modeling with local detail perception, suppressing residual noise and optimizing the contrast of the defect area, and outputting a denoised and enhanced image of the defect features. Specifically, this may include: S31, feature splitting: Split the input into two sets of features through 1×1 convolution: global features P1 and local features P2, which are adapted to feature modeling requirements of different scales.

[0049] S311, global feature modeling: input P1 into the global modeling, P1 Processing, the formula for calculating window attention is:

[0050] in, The query / key / value for P1, is the dimension of the key (K), is the relative position encoding, is, is, To encode relative positions, that is, the relative positions between blocks, the window size is set to 8 × 8. Through the window attention mechanism, global features such as the overall shape of the seal and the reflective patterns of the metal surface are captured, distinguishing the structural differences between defective areas (such as cracks) and the background.

[0051] S312, local feature extraction: The output global features are spliced ​​with the local details extracted by convolution to restore the multi-scale feature interaction; P2 is input into the local modeling, P2 extracts local details through 3×3 convolution, and the two sets of features are spliced. 1×1 convolution and residual connection are used to optimize feature transfer, avoid gradient vanishing, and enhance the expressiveness of the fused features. The final output formula is:

[0052] in, is the feature fusion function, are the learnable parameters of the hybrid architecture.

[0053] The loss function is:

[0054] in, is the defect contrast loss (for seal cracks), the specific formula is:

[0055] in, To generate pixel values ​​of defective areas of the image, is the pixel value of the defective area of ​​the defect-free image, To generate the pixel value of the background area of ​​the image, is the pixel value of the background area of ​​the defect-free image.

[0056] By maximizing the difference between the defect area and the background, the crack visibility is enhanced; and the multi-source features of the multi-level denoising results, the dual-path reconstructed image and the original noisy image are integrated to utilize The global structure of the lead seal is modeled and 3×3 convolution is used to extract local details such as the sealing ring. Combined with the defect contrast loss function, residual noise is further suppressed and the visual difference of defect areas such as cracks is significantly improved. The final output is a denoised and enhanced image with clear defect features suitable for industrial inspection, which solves the problem of low defect contrast and blurred details in extreme noise caused by traditional methods.

[0057] In a second aspect, a cable seal defect image denoising and enhancement system is provided, which is configured to execute the above-mentioned cable seal defect image denoising and enhancement method, comprising: Multi-level denoising module, used to denoise noisy seal images, separate the main contour of the cable seal from the background noise and retain the defect features; A dual-path reconstruction module is used to input the original noisy image and the denoising result of the multi-stage denoising module into the generative adversarial network for dual-path reconstruction, thereby enhancing the continuity of defect edges and the authenticity of surface texture through an adversarial training mechanism; A feature fusion module is provided, which fuses the denoising results of the multi-level denoising module, the dual-path reconstruction results of the dual-path reconstruction module, and the original noisy image, combines global structure modeling with local detail perception, suppresses residual noise, optimizes the contrast of the defect area, and outputs a denoised and enhanced image of the defect features.

[0058] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the above-mentioned method for denoising and enhancing a cable seal defect image is implemented.

[0059] In summary, the present invention has the following advantages: 1. A multi-level denoising architecture is used to separate the seal body from the noise, and cross-layer feature connections are combined to preserve defect details. This addresses the problem of traditional methods that incompletely remove metal oxidation artifacts and damage key features. 2. Generate adversarial networks to reconstruct real textures, combine edge preservation loss to enhance defect edge continuity, and maximize the difference between defects and background through defect contrast loss to improve crack visibility. 3. The collaborative architecture of U-Net, Transformer and GAN is adopted to achieve joint modeling of global structure and local details, improve the detection rate of microcracks under extreme working conditions such as strong electromagnetic interference, and significantly improve the reliability of automated detection.

[0060] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the scope of the technical solutions of the present invention, and all of these should be included in the scope of the claims of the present invention.

Claims

1. A cable seal defect image denoising and enhancement method, characterized in that: include: Multi-level denoising processing to denoise the noisy seal image, separate the main outline of the cable seal from the background noise and retain the defect features; Two-way reconstruction processing, inputting the original noisy image and the denoising result of the multi-level denoising processing into the generative adversarial network for two-way reconstruction, and enhancing the continuity of the defect edge and the authenticity of the surface texture through the adversarial training mechanism; Feature fusion processing, fusing the denoising results of the multi-level denoising processing, the dual-path reconstruction results of the dual-path reconstruction processing and the original noisy image, combining global structure modeling with local detail perception, suppressing residual noise and optimizing the contrast of the defect area, and outputting a denoised and enhanced image of the defect features.

2. The cable seal defect image denoising and enhancement method according to claim 1 is characterized in that: The multi-stage denoising process includes: Denoising is performed using a deep neural network architecture comprising an encoder and a decoder; batch normalization is used to automatically adjust the input data distribution of each layer of the neural network; Use activation functions to learn activation thresholds for different features; Restore the feature map output by the encoder to the input image size; connect the feature details extracted by the encoder to the corresponding layer of the decoder to avoid image blur; The composite noise is estimated by the noise estimation function and a preliminary denoised image is output.

3. The cable seal defect image denoising and enhancement method according to claim 2, characterized in that: The deep neural network architecture is a U-Net architecture, and the encoder includes multiple layers of convolution, each layer using a preset size convolution kernel and step size to scan the image.

4. The cable seal defect image denoising and enhancement method according to claim 2, characterized in that: It also includes optimizing the denoising process using a loss function, which is a mean square error loss function, and the formula is: in, is the noise estimation loss, is the normalization factor, where H and W are the height and width of the image respectively, is the double summation symbol, To estimate the noise at pixel location The pixel value at is the noisy image at pixel position The pixel value at For a flawless seal image at pixel position The pixel value at is the real noise.

5. The cable seal defect image denoising and enhancement method according to claim 1, characterized in that: The dual-path reconstruction process includes: Image-to-image conversion block: inputs the multi-level denoising results and the original noisy image, and splices them into a multi-channel feature map; the conversion block includes a generator and a discriminator; The generator uses a deep neural network to repair the image, and the objective function includes pixel-level loss, conditional adversarial loss, and edge-preserving loss; The discriminator compares the texture and edges of the repaired image and the real defect-free image through convolution operations.

6. The cable seal defect image denoising and enhancement method according to claim 5, characterized in that: The objective function formula is: in, is the weight coefficient of each loss item, Pixel level loss, To fight against loss, Reserve loss for seal edge.

7. The cable seal defect image denoising and enhancement method according to claim 1, characterized in that: The feature fusion processing includes: Feature splitting: Split the input into global features and local features through 1×1 convolution; Global feature modeling: using Capture the overall structure of the seal; Local feature extraction: Details are extracted through 3×3 convolution, and feature transfer is optimized after splicing with global features; the loss function includes mean square error loss and defect contrast loss, the latter of which maximizes the difference between defects and background.

8. The cable seal defect image denoising and enhancement method according to claim 1, characterized in that: The feature fusion processing further includes: Adaptive weight allocation: The regional complexity is estimated by calculating the variance of the feature map and generating dynamic weight coefficients; the final image is generated through weighted fusion, where the weights of global features and local features are dynamically adjusted by the regional complexity.

9. A cable seal defect image denoising and enhancement system, characterized in that: include: Multi-stage denoising processing module: used to perform preliminary denoising on the noisy seal image, separate the main outline of the cable seal from the background noise and retain the defect characteristics; A dual-path reconstruction processing module is used to input the original noisy image and the denoising result of the multi-stage denoising processing module into the generative adversarial network for dual-path reconstruction, thereby enhancing the continuity of defect edges and the authenticity of surface texture through an adversarial training mechanism; Feature fusion processing module: used to fuse the denoising results of the multi-level denoising processing module, the dual-path reconstruction results of the dual-path reconstruction processing module and the original noisy image, combine global structure modeling with local detail perception, suppress residual noise and optimize the contrast of the defect area, and output a denoised and enhanced image of the defect features.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which, when executed by a processor, implements the cable seal defect image denoising and enhancement method according to any one of claims 1 to 8.

Images