A low-quality weld image enhancement method for preserving structure and balancing brightness

Abstract

The application provides a low-quality welding image enhancement method for preserving structure and balancing brightness, and the specific steps are as follows: inputting a low-quality welding original image, forming an initial illumination component image output through a principal component analysis algorithm; processing the initial illumination component image by using a relative total variation model to form an illumination component image output with optimized structure characteristics; dividing the illumination component image with optimized structure characteristics into two paths, one path is directly substituted into a Retinex model together with the original image to obtain a reflection component image, and the other path is processed by a logarithmic transformation enhancement algorithm to obtain an illumination component image with optimized brightness; fusing the reflection component image and the illumination component image with optimized brightness according to the Retinex model to obtain a reconstructed image; and introducing a guide filter to take the initial illumination component image as a guide image to obtain a final enhanced image. The image enhanced by the method has uniform and natural brightness distribution, good structure characteristic information preservation, no artifact or block effect, and clearer image details.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement methods, specifically a method for enhancing low-quality welding images while preserving structure and balancing brightness. Background Technology

[0002] The application of computer vision technology in arc welding process monitoring has become a major trend. This technology is widely used in industrial manufacturing, bringing significant benefits to the field. However, due to the influence of the welding environment, the welding images captured under monitoring are often accompanied by interference from strong arc light. This is because the dynamic range of brightness in ordinary industrial cameras is limited, around 60dB, while the dynamic range of brightness in the actual welding process is much higher than that of industrial cameras. This leads to quality problems such as uneven brightness and blurred structure in the welding images captured by the welding imaging system.

[0003] Currently, there are two main types of methods for improving the image quality of welding under strong arc light interference. One type uses filters to help select the incident wavelength range. Before selecting the filter size, a spectrometer is needed to determine the actual welding spectral range so that the filter size can be selected based on the spectral response. However, the detected spectral responses differ depending on the welding material, meaning that the spectral response must be re-detected and the corresponding filter configured every time the welding material is changed. This method is difficult to guarantee robustness of image quality and is costly. The other type uses image enhancement methods to optimize welding imaging. This method effectively improves image quality without considering material properties. Current welding image enhancement methods mainly include: histogram equalization, dark channel prior theory, and single-scale / multi-scale Retinex methods. In histogram equalization, the algorithm can lead to loss of structural information when processing images with overly concentrated saturation regions. Dark channel prior theory requires high accuracy in transmittance estimation; the accuracy and stability of the obtained transmittance are difficult to guarantee simultaneously, resulting in white edges appearing at structural areas. Methods based on Retinex theory rely heavily on the estimation of illumination components; improper estimation can lead to halos and over-smoothing of structural features. However, these methods neglect the structural characteristics of the welding target. Furthermore, due to the complexity of the welding environment, most of these methods are insufficient for monitoring the welding process.

[0004] The following publicly disclosed existing patent technologies were found through a search:

[0005] A computer vision-based welding image enhancement method [Application No.: CN202111198499.8] includes: acquiring a weld connected region image corresponding to an X-ray weld image to be processed; acquiring a grayscale inverse image corresponding to the weld connected region image; acquiring a suppression value image corresponding to the weld connected region image using the pixel value of each pixel in the grayscale inverse image; obtaining a final undercut probability map of the weld connected region using the pixel value of each pixel in the weld connected region image and its distance to the center line of the weld connected region; acquiring a masking region corresponding to the final undercut probability map of the weld connected region; calculating and swapping the pixel values ​​of each column of pixels in the masking region using a calculation template to obtain a distance undercut image corresponding to the weld connected region, thereby obtaining an undercut enhancement image corresponding to the weld connected region. This method can obtain a visually clear undercut image, facilitating undercut detection by humans or computers.

[0006] Upon comparison, the methods used in this patent application and the prior art documents are quite different, and the technical problems they aim to solve are also different. The technical differences between the prior art documents and this patent application are significant. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a low-quality welding image enhancement method for preserving structure and balancing brightness, which solves the problems of uneven brightness and loss of structural feature information in welding images caused by arc light.

[0008] The technical problem solved by this invention is achieved through the following technical solution:

[0009] A method for enhancing low-quality weld images while preserving structure and balancing brightness, characterized by the following specific steps:

[0010] Step 1: The input low-quality original welding image is used to extract the first principal component in the welding image as the initial illumination component through the principal component analysis algorithm, and the initial illumination component image is output.

[0011] Step 2: Process the initial illumination component image output from Step 1 using a relative total variation model to generate an illumination component image output with optimized structural features;

[0012] Step 3: The structure feature-optimized illumination component image output from Step 2 is further processed in two paths. One path is directly substituted into the Retinex model along with the original image to obtain the reflection component image, and the other path is processed through the logarithmic transformation enhancement algorithm to obtain the brightness-optimized illumination component image.

[0013] Step 4: Based on the Retinex model, fuse the reflection component map obtained in Step 3 with the brightness-optimized illumination component map to obtain the reconstructed image;

[0014] Step 5: Introduce a guide filter. Use the initial illumination component image output from Step 1 as the guide image to filter the reconstructed image obtained in Step 4, thereby obtaining the final enhanced image.

[0015] Furthermore, the specific steps of Step 1 are as follows:

[0016] Principal component analysis (PCA) is used to process the input low-quality welding raw image. The extracted first principal component is used as the initial illumination component map in the Retinex model, and the input image is used as the raw image. I Calculate the mean of matrix data μ The matrix is ​​first zero-mean normalized; then the matrix is ​​transposed and multiplied by itself to construct the covariance matrix; finally, the eigenvalues ​​and corresponding eigenvectors of the covariance matrix are obtained, and the largest eigenvalue is selected by sorting them. λ max and its corresponding eigenvectors EV max ; to feature vector EV max Compared with the original image I Dot product yields the first principal component of the input image, which serves as the initial illumination component map. L Output.

[0017] Furthermore, the specific steps of Step 2 are as follows:

[0018] The initial illumination component map is analyzed using a relative total variation model. L The process involves calculating the total change and intrinsic change of the pixel window to enhance the contrast between texture and structure. The relative total variation model combines the total change and intrinsic change of the pixel window to form an optimization objective function, as follows: In the formula, ||*|| F And ||*||1 represent the Firobeau norm and the 1-norm, respectively; L This is the initial illumination component diagram; It is an illumination component image with optimized structural features obtained relative to the total variation model; p Represents each pixel in the image traversal; and These represent the horizontal and vertical first-order derivative operators, respectively. and They represent the solutions L The first derivatives in both directions; and Represents the weight matrix; γ These are the coefficients of the balancing equations, where the weight matrix is ​​calculated by the following formula: In the formula, The center is located at p The area of ​​the point; i Index each pixel within this region; The standard deviation is expressed as σ Gaussian kernel; |*| represents the absolute value symbol; ε Given a very small positive constant, the illumination component image with optimized structural features is formed by solving the optimization objective function of formula (1). Output.

[0019] Furthermore, the specific steps of Step 3 are as follows:

[0020] The illumination component image optimized from the structural features output in the previous step. Then process it in two paths, one of which is compared with the original image. I We can directly substitute the data into the Retinex model to obtain the reflection component map. R The formula is as follows: The symbol 'e' represents a function;

[0021] Another approach uses a logarithmic transformation enhancement algorithm to obtain a brightness-optimized illumination component map. The formula is as follows: .

[0022] Furthermore, the specific steps of Step 4 are as follows:

[0023] Based on the Retinex model, the illumination component map output from the previous step is fused. and reflection component diagram R To obtain the reconstructed image I f The formula is as follows: .

[0024] Furthermore, the specific steps of Step 5 are as follows: using a guided filter to further optimize the reconstructed image. I f To improve the quality of the image and obtain the final enhanced image, the initial illumination component image is compared with the original illumination component image. L As the guide image required for the guided filter, the window radius of the guided filter is set to 3 and the regularization parameter is 0.001. The image after the guided filter is applied not only retains the original structural features but also has a uniform brightness distribution. At the same time, it can remove the noise, halo and block effect brought to the reconstructed image by the previous steps.

[0025] The advantages and positive effects of this invention are:

[0026] Compared with existing image enhancement techniques, the method of this invention is more suitable for enhancing low-quality welding images acquired under strong arc welding light interference. The enhanced image has a uniform and natural brightness distribution, better preservation of structural feature information, no artifacts or block effects, and clearer image details. Specifically, it has the following advantages:

[0027] (1) The method of the present invention can still effectively estimate the illumination component in the Retinex model under the interference of strong arc welding light, and retain as much edge, texture and other information in the original image as possible through structural feature optimization;

[0028] (2) The method of the present invention further optimizes the brightness distribution of the illumination component and reconstructs it by fusing it with the extracted reflection component. The generated image can restore and display the surrounding scene that was flooded by the strong arc welding light.

[0029] (3) The method of the present invention uses the initial illumination component as a guide map to perform guided filtering on the reconstructed image, which effectively removes noise, halo and block effect caused by the previous processing to the reconstructed image, and effectively improves the image quality. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0031] Figure 2 This is a comparison diagram of the image processing results when the arc welding light is weak in this embodiment of the invention and the processing results of other existing enhancement methods;

[0032] Figure 3 This is a comparison diagram of the image processing results when the arc welding light is strong in this embodiment of the invention and the processing results of other existing enhancement methods;

[0033] Figure 4 This is a bar chart showing the information entropy of the original image and the results of all enhancement methods in the embodiments of the present invention;

[0034] Figure 5 This is a histogram showing the structural similarity of the results of all enhancement methods in the embodiments of the present invention;

[0035] Figure 6 This is a histogram of the average brightness of the original image and the results of all enhancement methods in the embodiments of the present invention;

[0036] Figure 7 This is an average gradient histogram of the original image and the results of all enhancement methods in the embodiments of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] Figure 1 The diagram illustrates a flowchart of a low-quality welding image enhancement method for preserving structure and balancing brightness according to an embodiment of the present invention. First, the input low-quality welding image is processed using Principal Component Analysis (PCA) to extract the first principal component of the input image as the initial illumination component, forming an initial illumination component image output. Then, the initial illumination component image is processed using a Relative Total Variation (RTV) model to enhance the structural features of the image, forming a structure-optimized illumination component image. Next, the structure-optimized illumination component image is processed in two paths: one path directly substitutes it into the Retinex model along with the original image to obtain the reflection component map, and the other path uses a logarithmic transformation enhancement algorithm to obtain a brightness-optimized illumination component map. Then, the brightness-optimized illumination component map and the reflection component map are fused into the Retinex model to obtain the reconstructed image. Finally, a guided filter is introduced, using the initial illumination component as a guide image, to remove noise, halos, and block effects introduced by the previous steps into the reconstructed image, resulting in the final enhanced image.

[0039] Combination Figure 1 As shown, the present invention provides a method for enhancing low-quality weld images by preserving structure and balancing brightness. The specific implementation steps are as follows:

[0040] (1) The PCA algorithm is used to process the input low-quality welding original image, and the extracted first principal component is used as the initial illumination component map in the Retinex model. The input image is used as the original image. I Calculate the mean of matrix data μ The matrix is ​​first zero-mean normalized; then the matrix is ​​transposed and multiplied by itself to construct the covariance matrix; finally, the eigenvalues ​​and corresponding eigenvectors of the covariance matrix are obtained, and the largest eigenvalue is selected by sorting them. λ max and its corresponding eigenvectors EV max ; to feature vector EV max With the original image I The dot product, which is the first principal component of the input image, serves as the initial illumination component map. L Output.

[0041] (2) Initial illumination component map obtained by RTV model L The process involves processing the data. Contrast between texture and structure is enhanced by traversing and calculating the windowed total variations (WTV) and windowed inherent variations (WIV) of the pixel window. The RTV model combines WTV and WIV to form an optimization objective function, as follows:

[0042] In the formula, ||*|| F And ||*||1 represent the Firobeau norm and the 1-norm, respectively; L This is the initial illumination component diagram; It is an illumination component image with optimized structural features obtained relative to the total variation model; p Represents each pixel in the image traversal; and These represent the horizontal and vertical first-order derivative operators, respectively. and They represent the solutions L The first derivatives in both directions; and Represents the weight matrix; γ The coefficients of the balancing equation (in this embodiment) γ (Take a value of 0.25), where the weight matrix is ​​calculated by the following formula: In the formula, The center is located at p The area of ​​the point; i Index each pixel within this region; The standard deviation is expressed as σ Gaussian kernel; |*| represents the absolute value symbol; ε Given a very small positive constant (0.001 in this embodiment), the illumination component image with optimized structural features is formed by solving the optimization objective function of formula (1). Output.

[0043] (3) Optimize the illumination component image of the structural features output in the previous step. Then process it in two separate steps.

[0044] Along the way with the original image I We can directly substitute the data into the Retinex model to obtain the reflection component map. R The formula is as follows:

[0045] The symbol 'e' represents a function;

[0046] Another approach uses a logarithmic transformation enhancement algorithm to obtain a brightness-optimized illumination component map. The formula is as follows: .

[0047] (4) Based on the Retinex model, fuse the illumination component map output in the previous step. and reflection component diagram R To obtain the reconstructed image I f The formula is as follows:

[0048]

[0049] (5) Guided filtering is used to further optimize the reconstructed image. I f The quality is improved to obtain the final enhanced image. This patent uses the initial illumination component image... L As the guiding image required for the guided filter, the window radius of the guided filter is set to 3, and the regularization parameter is 0.001. The image after the guided filter process not only retains the original structural features and the brightness distribution remains uniform, but also removes noise, halo, and blockiness introduced by the previous steps into the reconstructed image.

[0050] The specific experimental comparison data are as follows:

[0051] This invention employs the method of this patent to enhance actual images acquired during the arc welding process, and compares them experimentally with six commonly used image enhancement algorithms: Single-scale Retinex (SSR), Multi-scale Retinex (MSR), Multi-scale Retinex with Color Restoration (MSRCR), Histogram Transformation (HT), Log Transformation (LT), and Dark Channel Prior (DCP).

[0052] Because the intensity of the arc welding light varies with the welding current during welding, the range of the arc welding light covering the video monitoring scene also varies. Using the method of this patented invention and the commonly used methods described above, three welding images each under two states—weak and strong arc welding light—were processed, and the processing effects were compared from a subjective observation perspective. Figure 2The image shows the processing effect of three welding images (Image1, Image2, Image3) when the signal is weak. Figure 3 The image shows the processing effect of three welding images (Image4, Image5, Image6) under relatively strong conditions. Figure 2 and 3 The specific illustrations for (a) to (h) are as follows: (a) Original figure; (b) SSR; (c) MSR; (d) MSRCR; (e) HT; (f) LT; (g) DCP; (h) This method. From the experimental results figures, it can be intuitively seen that... Figure 2 In (b) and (c), the overall contrast of the image is enhanced, and some information in the dark areas of the image is also restored. However, the actual lighting is uneven, resulting in an unnatural restored image, where the brightness distribution is still concentrated on the arc light and some halo artifacts appear at the edges. Figure 2 (d) Although a color restoration factor was introduced to adjust the proportions of the three channel components, the final image showed good color restoration, but the overall brightness was very low. Figure 2 The brightness enhancement in (e) is too drastic, exposing the entire dark area and resulting in the loss of many structural details. The recovered image appears unnatural, and pixel blocks caused by noise hinder the observability of the welded image. Logarithmic transformation balances the brightness distribution between dark and bright areas, improving the result. Figure 2 (f) This is more in line with human color vision characteristics; however, contrast and sharpness are reduced. After dark channel prior processing, the brightness of the recovered image becomes darker, such as... Figure 2 As shown in (g). Figure 2 As shown in (h), the results obtained by this patented method have high contrast, abundant structural information, uniform brightness distribution in layers, no block effect or halo artifacts, and the enhanced image is clearer than the image obtained by other methods. While restoring the image features within the position before and after the welding torch, it also makes the weld bead and the surrounding scene clearer. Figure 3 Methods (b) and (c) are inaccurate in their illumination estimation, resulting in halo artifacts appearing in dark areas near the edges of strong arc lights; in addition, the improvement in contrast comes at the cost of suppressing the brightness of the light areas, resulting in a loss of detail. Figure 3 (d) The brightness is too low to meet the requirements for observation. Figure 3 (e) and (f) contain rich brightness information; however, the introduction of a large amount of noise degrades image quality and causes blockiness. Figure 3 The brightness optimization in (g) is small, and the restored image shows some white edges. Figure 3 (h) This patented method compensates for halo artifacts caused by insufficient lightmap estimation, which not only increases contrast but also preserves more useful information, enhances the image to have sufficient structural contours and appropriate brightness, and effectively avoids the blocking effect.

[0053] Furthermore, the results obtained by the method of this invention and the commonly used methods described above are compared using objective evaluation metrics. The enhancement effects of each method are compared using four objective image quality evaluation metrics: entropy, structural similarity (SSIM), mean brightness (MB), and mean gradient (MG). A higher entropy indicates that the enhanced image retains more details. A SSIM closer to 1 indicates better preservation of structural features after enhancement. MB reflects the visual effect of image brightness; a higher value is better. MG reflects image sharpness and is the mean of the grayscale change rate; a higher value is better.

[0054] Using the four image quality evaluation metrics mentioned above, Figure 2 and Figure 3 The experimental results from the six original images were evaluated and calculated, and the results were plotted as a visual bar chart, such as... Figures 4 to 7 As shown, each bar chart group corresponds to six data points from Image1 to Image6 from left to right. Figure 4 The entropy values ​​of six original images and images enhanced by SSR, MSR, MSRCR, HT, LT, DCP, and the method of this patent (Our) are shown. Figure 5 The SSIM values ​​of images enhanced by SSR, MSR, MSRCR, HT, LT, DCP, and Our methods are shown. Figure 6 The MB values ​​of Original and SSR, MSR, MSRCR, HT, LT, DCP and Our enhanced images are displayed. Figure 7 The MG values ​​of Original and SSR, MSR, MSRCR, HT, LT, DCP and Our enhanced images are displayed.

[0055] The entropy and SSIM performance of this patented method are superior, demonstrating that the processed image contains more useful information. The MB and MG performance of this patented method are second only to the HT and LT methods. It should be clarified that the enhanced brightness should meet a certain hierarchical distribution, rather than only considering the maximum brightness level. Based on the actual enhanced image effects compared from the above subjective observation angles, the HT method suffers from loss of structural details and unnatural restored images, while the LT method suffers from reduced contrast and sharpness. This patented method, while ensuring brightness meets the field of view, achieves a more reasonable brightness distribution and improves the clarity of structural information. Therefore, combining subjective and objective evaluations, this patent is more effective in enhancing low-quality welding images under different intensities.

[0056] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A low-quality weld image enhancement method for preserving structure and balancing brightness, characterized in that: The specific steps are: Step 1: input a low-quality welding original image, and extract the first principal component in the welding image as an initial illumination component through a principal component analysis algorithm to form an initial illumination component image output; Step 2: processing the initial illumination component image output in Step 1 by using a relative total variation model to form an illumination component image output with optimized structural features; Step 3: processing the illumination component image output with optimized structural features in Step 2 in two ways, one of which is directly substituted into a Retinex model together with the original image to obtain a reflection component map, and the other of which is processed by a logarithmic transformation enhancement algorithm to obtain a brightness-optimized illumination component map; Step 4: according to the Retinex model, fusing the reflection component map and the brightness-optimized illumination component map obtained in Step 3 to obtain a reconstructed image; Step 5: introducing a guided filter, taking the initial illumination component image output in Step 1 as a guide image, and filtering the reconstructed image obtained in Step 4 to obtain a final enhanced image.

2. The low-quality weld image enhancement method for preserving structure and balancing brightness of claim 1, wherein: The specific steps of Step 1 are: The principal component analysis algorithm is used to process the input low-quality welding original image, and the first principal component extracted is used as an initial illumination component graph in the Retinex model, and the input image is used as an original image I The mean value of the matrix data is calculated μ The matrix is zero-mean processed; Then, the matrix is transposed and multiplied with the matrix itself to construct a covariance matrix; Then the eigenvalues and the corresponding eigenvectors of the covariance matrix are obtained λ max and the corresponding eigenvectors EV max ; the eigenvectors EV max are multiplied with the original image I , and the first principal component of the input image is obtained as the initial illumination component map L output.

3. The low-quality weld image enhancement method for preserving structure and balancing brightness of claim 1, wherein: The specific steps of Step 2 are: The initial illumination component map is analyzed using a relative total variation model. L The process involves calculating the total change and intrinsic change of the pixel window to enhance the contrast between texture and structure. The relative total variation model combines the total change and intrinsic change of the pixel window to form an optimization objective function, as follows: (1) where ||*|| and ||*||1 represent the Frobenius norm and 1-norm, respectively; F where ||*|| and ||*||1 represent the Frobenius norm and 1-norm, respectively; L is the initial illumination component map; is the structure feature optimized illumination component map obtained by the relative total variation model; p represents each pixel point of the image traversal; and respectively represent the horizontal and vertical first derivative operators, and respectively represent the first derivative of two directions of solving L ; represents the weight matrix; γ is the coefficient of balancing the two terms, wherein the weight matrix is calculated by the following formula: (2) wherein, is a region centered at p ; i is an index of each pixel point in the region; is a Gaussian kernel with a standard deviation of σ ; |*| represents the absolute value symbol; ε is a very small positive number, and by solving the optimization objective function of formula (1), a structure feature optimized illumination component image is output.

4. The low-quality weld image enhancement method for preserving structure and balancing brightness of claim 1, wherein: The specific steps of Step 3 are: the structure feature optimized illumination component image output in the previous step is divided into two paths again, The original image is directly substituted into the Retinex model to obtain the reflection component image I together R The formula is as follows: (3) The symbol e represents a function; Another path is processed by a logarithmic transformation enhancement algorithm to obtain a luminance-optimized illumination component map The formula is as follows: (4)。 5. The low-quality weld image enhancement method for preserving structure and balancing brightness of claim 1, wherein: The specific steps of Step 4 are: According to the Retinex model, the light component map output in the previous step is fused and the reflection component map R to obtain a reconstructed image I f The formula is as follows: (5)。 6. The low-quality weld image enhancement method for preserving structure and balancing brightness of claim 1, wherein: The specific steps of Step 5 are: Guided filter is used to further optimize the reconstructed image I f The quality of the final enhanced image is obtained by using the initial light component image L As the guide image required by the guided filter, the window radius of the guided filter is set to 3 and the regularization parameter is 0.

001. The image after the guided filter processing not only retains the original structural features and uniform brightness distribution, but also removes the noise, halo and blocking effect brought by the previous step.

Images