Box-type substation spraying quality detection method based on image processing

Abstract

The present application relates to the field of image processing, more particularly, the present application relates to a box-type substation spraying quality detection method based on image processing, the method comprises: first, obtaining the image to be detected, calculating the pixel point structure tensor component after pretreatment; then based on the component, extracting the gradient energy field and the gradient main direction angle field, and then calculating the gradient radial consistency representing the geometric characteristics of the defect; combining the image to be detected with the gradient energy field, calculating the saturation and flatness index representing the optical highlight characteristics, and the intensity-gradient cross-correlation representing the spatial form of the defect; fusing the three features to generate the final saliency map; finally, threshold segmentation and post-processing of the saliency map are carried out to obtain the detection result. The present application can effectively distinguish physical spraying defects from optical highlight reflection, and significantly improve the defect detection accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of image processing. More specifically, this invention relates to a method for inspecting the coating quality of prefabricated substations based on image processing. Background Technology

[0002] Prefabricated substations, as crucial units in power systems, operate outdoors for extended periods. The quality of their exterior coating not only affects the equipment's aesthetics but also directly determines its corrosion resistance and service life. A high-quality coating is a critical barrier protecting the substation's exterior from environmental erosion such as rain, salt spray, and ultraviolet radiation. Therefore, rigorous testing of the coating quality before shipment is an indispensable step in ensuring product quality.

[0003] However, several factors contribute to the difficulty of inspection during automated spraying and testing. On the one hand, the spraying process may introduce various physical defects, such as particles, orange peel, fisheyes, or pinholes. These physical defects disrupt the smoothness and continuity of the coating and must be detected. On the other hand, to achieve high corrosion resistance and aesthetics, the coating of prefabricated substations typically has a high gloss. This results in strong specular highlights or optical spots easily appearing on the surface under the lighting conditions of machine vision inspection. These optical reflections also appear as bright or high-contrast spot areas in the image, with morphologies highly similar to actual physical defects.

[0004] Traditional threshold-based detection methods, relying solely on grayscale information, are prone to misclassifying bright speckled reflections as coating defects. Methods based on edge detection or simple morphological analysis also fail to effectively distinguish the edges of reflective spots from the edges of physical defects. This results in existing algorithms facing a common dilemma of high false positive and high false negative rates when inspecting the coating quality of prefabricated substations. Therefore, the industry urgently needs a detection method that can differentiate between physical coating defects and optical reflection interference based on their formation mechanisms to address this technical challenge. Summary of the Invention

[0005] To address the challenge of effectively distinguishing between defects and reflection artifacts and achieving high-precision defect detection, this invention proposes an image processing-based method for inspecting the coating quality of prefabricated substations. This method includes the following steps:

[0006] Acquire the image of the prefabricated substation to be inspected; preprocess the image to be inspected and calculate the structural tensor components of each pixel in the image to be inspected;

[0007] Based on the structural tensor components, the gradient energy field and gradient principal direction angle field of each pixel are extracted;

[0008] Based on the gradient energy field and the gradient principal direction angle field, the gradient radial consistency used to characterize the geometric properties of spraying defects is calculated;

[0009] Based on the image to be detected and the gradient energy field, the saturation and flatness indices used to characterize the optical reflective properties of the sprayed surface are calculated.

[0010] Based on the image to be detected and the gradient energy field, the intensity-gradient cross-correlation used to characterize the spatial morphology of the spraying defects is calculated;

[0011] Based on the gradient radial consistency, the intensity-gradient cross-correlation, and the saturation and flatness indices, a final saliency map is generated by fusion.

[0012] Threshold segmentation and post-processing are performed on the final saliency map to obtain the detection results of the spraying defects.

[0013] This invention constructs a multi-dimensional, complementary defect detection method by parallel computing and ultimately fusing gradient radial consistency, saturation and flatness indices, and intensity-gradient cross-correlation. This enables the detection method to make comprehensive judgments from different perspectives, thereby greatly improving the detection accuracy and robustness of various spraying defects in complex industrial environments, and significantly reducing missed detections and false alarms.

[0014] Preferably, the preprocessing of the image to be detected includes:

[0015] The image to be detected is smoothed by applying a bilateral filter or anisotropic diffusion algorithm to suppress noise while preserving edge structure.

[0016] This invention employs edge-preserving denoising algorithms, such as bilateral filters or anisotropic diffusion algorithms, to effectively filter out image sensor noise while maximally protecting the subtle but crucial edge information of spraying defects from obscuring. This provides a high-quality, high signal-to-noise ratio image foundation for subsequent structural tensor analysis and feature extraction, which is key to ensuring the detection of minute defects.

[0017] Preferably, the extraction of the gradient energy field of each pixel includes:

[0018] The structure tensor is subjected to eigenvalue decomposition, and the largest eigenvalue is obtained as the gradient energy field;

[0019] Alternatively, the second largest eigenvalue can be obtained as the second eigenvalue, and the sum of the largest eigenvalue and the second eigenvalue can be used as the gradient energy field.

[0020] This invention utilizes the eigenvalues ​​of the structure tensor to define gradient energy, which, compared to traditional gradient magnitude calculation, can more stably and accurately reflect the intensity and directionality of structural changes within the pixel neighborhood. This makes the gradient energy field insensitive to noise and can better distinguish between real edge structures and non-directional textures, providing a more reliable and discriminative metric for all subsequent gradient-based feature calculations.

[0021] Preferably, the calculation of the gradient radial consistency used to characterize the geometric properties of spraying defects includes:

[0022] Within a preset neighborhood window, calculate the radial consistency of the gradient for each center pixel:

[0023] ;

[0024] in, For neighboring pixels The value of the gradient energy field at that location, For neighboring pixels The value of the gradient principal direction angle field at that location, To be from the center point to Standard radial angle, For the radial similarity of the center pixel, cos() represents the cosine similarity. This is a preset neighborhood window.

[0025] This invention takes into account the characteristic that defects such as bubbles and pits, which are dot-shaped or cluster-shaped, naturally have a gradient direction that points towards or away from the center, while speckled reflective artifacts do not have this characteristic. Therefore, an index is constructed based on this to effectively distinguish between spraying defects and reflective artifacts.

[0026] Preferably, the saturation and flatness index satisfy one of the following relationships:

[0027] ;

[0028] ;

[0029] in, For normalized grayscale values, For the normalized gradient energy field, and To adjust the coefficient, The original grayscale value. As a hard threshold for saturation, For indicator functions, satisfying Under certain conditions, Select 1 if the value is 1, otherwise select 0. tanh() represents the hyperbolic tangent function, which is the saturation and flatness index.

[0030] This invention takes into account that highly reflective artifacts are specular reflections, and these areas exhibit high brightness but are flat and without internal variations. Therefore, an index is constructed based on this characteristic, which can accurately identify and mark these high-brightness interference areas, providing a key basis for their subsequent removal in the final result.

[0031] Preferably, the calculation of the intensity-gradient cross-correlation used to characterize the spatial morphology of spraying defects includes:

[0032] The peak response based on the image to be detected and the flanking response based on the gradient energy field are obtained respectively.

[0033] The intensity-gradient cross-correlation is obtained by multiplying the peak response and the flank response.

[0034] This invention takes into account that raised coating defects typically exhibit abnormal brightness at the center and drastic changes in their boundaries. Therefore, an index is constructed based on this characteristic, providing a foundation for subsequent elimination of reflection artifacts and accurate defect identification.

[0035] Preferably, a Laplace-Gaussian (LoG) filter is applied to the image to be detected to obtain the peak response of the image to be detected; and a Difference-of-Gaussian (DoG) filter is applied to the gradient energy field to obtain the flank response of the image to be detected.

[0036] Preferably, the fusion to generate the final saliency map includes:

[0037] The radial consistency of the gradient and the cross-correlation of the intensity-gradient are subjected to geometric average or weighted average to obtain the defect confidence map.

[0038] The final saliency map is obtained by multiplying or subtracting the defect confidence map with the reflection suppression mask generated based on the saturation and flatness indices.

[0039] Preferably, the threshold segmentation of the final saliency map includes:

[0040] The final saliency map is binarized using the high percentile thresholding method, Otsu thresholding method, or adaptive thresholding method.

[0041] Preferably, the post-processing includes:

[0042] After thresholding the final saliency map, apply morphological opening and / or morphological closing operations to the resulting binary mask.

[0043] Connectivity analysis is applied to remove connected regions whose area is smaller than a preset area threshold.

[0044] The present invention has the following beneficial effects:

[0045] This invention constructs and integrates three orthogonal feature dimensions: geometry, optics, and spatial morphology, forming a powerful and complementary decision-making system. It can accurately identify the essential characteristics of defects and features specifically designed modules to suppress the strong reflective interference most common in industrial scenarios, thus achieving detection accuracy and anti-interference capabilities far exceeding traditional methods in complex environments.

[0046] Since this method is not designed for a specific defect, but rather starts from the underlying physical and geometric principles, it has a good detection effect on a variety of common spraying defects, such as scratches, bubbles, pits, runs, and uneven coatings, and has strong applicability. Attached Figure Description

[0047] Figure 1 This is a flowchart of the steps of the image processing-based method for inspecting the coating quality of a box-type substation provided in an embodiment of the present invention;

[0048] Figure 2 It is a grayscale image provided in the embodiments of the present invention;

[0049] Figure 3 A defect saliency map provided for an embodiment of the present invention;

[0050] Figure 4 Defect images provided for embodiments of the present invention. Detailed Implementation

[0051] Please see Figure 1 The diagram illustrates the steps of the image processing-based method for inspecting the coating quality of a prefabricated substation provided in Embodiment 1. The method includes the following steps:

[0052] S1: Acquire the image to be detected of the prefabricated substation; preprocess the image to be detected and calculate the structural tensor components of each pixel in the image to be detected.

[0053] Specifically, first, the image of the prefabricated substation to be inspected is acquired and converted into a grayscale image. Figure 2 The image displayed is a grayscale image. Several raised coating defects are present at the corner of the prefabricated substation in this image, and several speckled high-reflectivity artifacts are present in the upper left corner of the image. These speckled high-reflectivity artifacts are very similar to the raised coating defects in color and texture, thus interfering with the detection of coating defects.

[0054] Next, the grayscale image is filtered to obtain a smooth image. For example, the filtering process can use anisotropic diffusion algorithm or bilateral filtering algorithm. This embodiment does not impose any specific limitations.

[0055] Next, the gradient components of the smoothed image in the horizontal and vertical directions are calculated. For example, this can be achieved using... The operator can be used for calculation, or an anisotropic diffusion algorithm can be used for calculation. This embodiment does not impose specific limitations.

[0056] Finally, within a preset tensor calculation window, the product of the gradient components is locally Gaussian weighted averaged to calculate the three components of the structure tensor corresponding to each pixel.

[0057] S2: Based on the structural tensor components, extract the gradient energy field and gradient principal direction angle field of each pixel.

[0058] It should be noted that the structural tensor components cannot adequately describe physical properties. It is necessary to extract key physical properties from the structural tensor, namely the most significant energy and the clearest texture direction of the local structure.

[0059] Preferably, the method for obtaining the gradient energy field and the gradient principal direction angle field of each pixel includes:

[0060] First, based on the structural tensor components, a structural tensor matrix is ​​constructed for each pixel.

[0061] Next, the structural tensor matrix is ​​decomposed into eigenvalues ​​to obtain the largest eigenvalue and the principal eigenvector corresponding to the largest eigenvalue.

[0062] For example, the largest eigenvalue is used as the gradient energy field.

[0063] Optionally, the second largest eigenvalue can be obtained as the second eigenvalue, and the sum of the largest eigenvalue and the second eigenvalue can be used as the gradient energy field.

[0064] The maximum eigenvalue and the second eigenvalue represent the energy situation in the direction where the gradient change is more drastic within the neighborhood of the pixel.

[0065] Next, the angle of the principal feature vector is calculated as the principal gradient direction angle. This principal direction angle represents the direction of the strongest gradient change within the neighborhood of that pixel.

[0066] S3: Based on the gradient energy field and the gradient principal direction angle field, calculate the gradient radial consistency used to characterize the geometric properties of spraying defects.

[0067] It should be noted that spraying defects such as particle protrusions will form a gray-scale hill in the image, and its gradient field should exhibit radial characteristics that radiate from the center outwards. However, reflection artifacts do not exhibit this feature. Therefore, an index for distinguishing reflection artifacts from defects can be constructed based on this physical characteristic.

[0068] As an embodiment of the present invention, gradient radial consistency can be calculated using an energy-weighted method, which satisfies the following relationship:

[0069]

[0070] Among them, the preset consistency calculation window , Center pixel For window Any neighboring pixel within. Neighborhood pixels The gradient energy field at that location, Neighborhood pixels The gradient principal direction angle at that location, From the center point to Standard radial angle, For gradient radial consistency.

[0071] Understandable, It reflects the deviation of the actual gradient direction from the standard radial angle. The larger the value, the less the gradient direction conforms to the radial characteristic of diverging from the center to the surroundings.

[0072] Furthermore, by using gradient energy as weights, the calculation focuses on the high-energy gradient signal while automatically ignoring low-energy background noise, thus solving the technical problem that the true radial signal of tiny spraying defects is diluted by the surrounding low-energy background.

[0073] Optionally, the gradient radial consistency can also be calculated in an unweighted manner, satisfying the following relationship:

[0074]

[0075] in, It is a window The total number of pixels in For gradient radial consistency.

[0076] S4: Based on the image to be detected and the gradient energy field, calculate the saturation and flatness indices used to characterize the optical reflective properties of the sprayed surface.

[0077] It should be noted that, unlike coating defects, the specular reflection produced by the newly coated, high-gloss surface of a prefabricated substation is characterized by extreme brightness and internal flatness. Therefore, these two characteristics can be used to calculate another indicator to distinguish between defects and reflective artifacts.

[0078] As an embodiment of the present invention, the saturation degree and the flatness index satisfy the following relationship:

[0079]

[0080] This involves converting the image to be detected from the RGB color space to the HSV color space. Saturation is used to reflect the brightness of a pixel. The higher the value, the brighter the pixel and the closer it is to saturation. For the normalized gradient energy field, To adjust the coefficient, It is a sigmoid function whose function is to nonlinearly map the gradient energy field to... Interval. The flatness term is used to characterize the flatness at a pixel. The larger the gradient energy field, the more significant the texture, and the smaller the calculated flatness term.

[0081] It's understandable that a high exponent is only obtained when both saturation and flatness are high. This design can effectively locate specular highlights in an image.

[0082] Optionally, the saturation and flatness index can also satisfy the following relationship:

[0083]

[0084] The original grayscale value. It is a hard threshold for saturation. It is an indicator function. Let be the saturation term, where satisfies hour, The value is set to 1 if the saturation level is greater than a preset threshold, and 0 otherwise. This calculation method allows the saturation level to be set to 1 only when the saturation level is greater than a preset threshold, thus enabling the focus to be placed on pixels with high brightness. For the normalized gradient energy field, It is a constant. The flatness term is used to characterize the flatness at a pixel. The larger the gradient energy field, the more significant the texture, and the smaller the calculated flatness term.

[0085] S5: Based on the image to be detected and the gradient energy field, calculate the intensity-gradient cross-correlation used to characterize the spatial morphology of spraying defects.

[0086] It should be noted that spraying defects exhibit the characteristics of grayscale hills, meaning that the edges of physical defects are where the gradient energy is highest, while the energy is lower at the center and far from the defect. Therefore, an intensity-gradient cross-correlation index can be calculated based on this characteristic to distinguish between defects and reflection artifacts.

[0087] Preferred methods for obtaining intensity-gradient cross-correlation include:

[0088] First, the image to be detected is filtered to obtain the peak response of each pixel. For example, the filtering process can be performed using a Laplace-Gaussian (LoG) filter.

[0089] Understandably, the Laplace-Gaussian (LoG) filter detects blobs by calculating the second derivative of a grayscale image. For a grayscale hill formed by a physical defect, its peak has the largest second derivative value.

[0090] Next, a Difference of Gaussian (DoG) filter is applied to the gradient energy field to obtain the flanking response of each pixel.

[0091] Understandably, the edges of the defect are where the gradient energy is highest, while the energy is lower at the center and far from the defect, forming a ring-like structure in the gradient energy field. The DoG filter is a bandpass filter designed to produce a highly specific response to this ring-like structure.

[0092] Next, the intensity-gradient cross-correlation is calculated:

[0093]

[0094] in, For the peak response, For flank response, This is an intensity-gradient cross-correlation.

[0095] It is understandable that the intensity-gradient cross-correlation obtained by multiplying the two will be high only when a pixel is simultaneously located at the intensity peak and the gradient energy around it exhibits a ring structure.

[0096] S6: Based on the gradient radial consistency, the intensity-gradient cross-correlation, and the saturation and flatness indices, the final saliency map is generated by fusion.

[0097] It should be noted that the previous steps extracted indicators to distinguish between defects and reflection artifacts. The next step involves intelligently fusing these three indicators to generate a final saliency map that provides the highest response to real-world spraying defects while exhibiting the strongest suppression of reflection artifacts.

[0098] Preferably, based on the gradient radial consistency, the intensity-gradient cross-correlation, and the saturation and flatness indices, the final saliency map is generated by fusing the data, including:

[0099] First, calculate the defect confidence level.

[0100] The gradient radial consistency, the saturation and flatness indices, and the intensity-gradient cross-correlation are normalized to... scope.

[0101] The defect confidence level satisfies the following relationship:

[0102]

[0103] To ensure radial consistency of the normalized gradient, The normalized intensity-gradient cross-correlation, Indicates the confidence level of the defect.

[0104] Understandable, Geometric features characterizing defects The spatial morphological characteristics of the defect are characterized by using geometric mean to fuse the two features of the defect.

[0105] Next, calculate the reflection suppression mask:

[0106]

[0107] in, These are the normalized saturation and flatness indices. This represents a reflection suppression mask.

[0108] Understandable, Characterizing pixels It is the probability of optical reflection. This represents the probability that it is not an optical reflective material.

[0109] Finally, the significance of the defects is calculated.

[0110] For example, the defect significance can be calculated by multiplying the defect confidence level by the reflection suppression mask, or by weighted summation of the defect confidence level and the reflection suppression mask. Figure 3 This is a defect saliency map. As can be seen from this image, the color is relatively brighter at the location of the spraying defect and relatively darker at other locations. Therefore, the defect saliency map can clearly distinguish the spraying defects.

[0111] S7: Perform threshold segmentation and post-processing on the final saliency map to obtain the detection results of the spraying defects.

[0112] Preferably, the final saliency map is subjected to threshold segmentation and post-processing to obtain the detection result of the spraying defect, including:

[0113] A segmentation threshold is calculated based on the final saliency map. For example, the high percentile thresholding method, the Otsu thresholding method, or the locally adaptive thresholding method can be used to obtain the segmentation threshold.

[0114] The final saliency map is binarized using the segmentation threshold to obtain a binary mask.

[0115] Morphological opening and closing operations are applied to the binary mask. The morphological opening operation removes burrs and small noise points, while the morphological closing operation connects adjacent defective regions that may be broken.

[0116] By applying connected component analysis, the area of ​​each connected region in the binary mask is calculated, and all connected regions with an area smaller than the preset minimum defect area are removed to filter out meaningless minute noise.

[0117] The final defect mask is obtained, which is the detection result of the spraying defect. Figure 4 The image shows a defective area circled in green. The image demonstrates that the present invention can accurately eliminate the interference of speckled reflective artifacts and accurately detect coating defects.

[0118] This concludes the embodiment.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inspecting the coating quality of prefabricated substations based on image processing, characterized in that, include: Acquire the image of the prefabricated substation to be inspected; The image to be detected is preprocessed and the structural tensor components of each pixel in the image to be detected are calculated. Based on the structural tensor components, the gradient energy field and gradient principal direction angle field of each pixel are extracted; Based on the gradient energy field and the gradient principal direction angle field, the gradient radial consistency used to characterize the geometric properties of spraying defects is calculated; Based on the image to be detected and the gradient energy field, the saturation and flatness indices used to characterize the optical reflective properties of the sprayed surface are calculated. Based on the image to be detected and the gradient energy field, the intensity-gradient cross-correlation used to characterize the spatial morphology of the spraying defects is calculated; Based on the gradient radial consistency, the intensity-gradient cross-correlation, and the saturation and flatness indices, a final saliency map is generated by fusion. The final saliency map is subjected to threshold segmentation and post-processing to obtain the detection results of the spraying defects; Saturation and flatness indices satisfy one of the following relationships: ； ； in, For normalized grayscale values, For the normalized gradient energy field, and To adjust the coefficient, The original grayscale value. As a hard threshold for saturation, For indicator functions, satisfying Under certain conditions, Select 1 if the value is 1, otherwise select 0. The saturation and flatness indices are represented by tanh(), which denotes the hyperbolic tangent function. Calculate the intensity-gradient cross-correlation used to characterize the spatial morphology of spray defects, including: The peak response based on the image to be detected and the flanking response based on the gradient energy field are obtained respectively; the peak response and the flanking response are multiplied to obtain the intensity-gradient cross-correlation. A Laplace-Gaussian (LoG) filter is applied to the image to be detected to obtain the peak response of the image; a Difference-of-Gaussian (DoG) filter is applied to the gradient energy field to obtain the flanking response of the image. The process of fusing and generating the final saliency map includes: performing a geometric average or weighted average operation on the gradient radial consistency and the intensity-gradient cross-correlation to obtain the defect confidence level; The final saliency map is obtained by multiplying or subtracting the defect confidence score with the reflection suppression mask generated based on the saturation and flatness indices.

2. The image processing-based method for inspecting the coating quality of a prefabricated substation according to claim 1, characterized in that, The preprocessing of the image to be detected includes: The image to be detected is smoothed by applying a bilateral filter or anisotropic diffusion algorithm to suppress noise while preserving edge structure.

3. The method for inspecting the coating quality of a prefabricated substation based on image processing according to claim 1, characterized in that, The extraction of the gradient energy field of each pixel includes: The structure tensor is subjected to eigenvalue decomposition, and the largest eigenvalue is obtained as the gradient energy field; Alternatively, the second largest eigenvalue can be obtained as the second eigenvalue, and the sum of the largest eigenvalue and the second eigenvalue can be used as the gradient energy field.

4. The method for inspecting the coating quality of a prefabricated substation based on image processing according to claim 1, characterized in that, The calculation used to characterize the gradient radial consistency of the geometric properties of spraying defects includes: Within a preset neighborhood window, calculate the radial consistency of the gradient for each center pixel: ； in, For neighboring pixels The value of the gradient energy field at that location, For neighboring pixels The value of the gradient principal direction angle field at that location, To be from the center point to Standard radial angle, For the radial similarity of the center pixel, cos() represents the cosine similarity. This is a preset neighborhood window.

5. The method for inspecting the coating quality of a box-type substation based on image processing according to claim 1, characterized in that, The threshold segmentation of the final saliency map includes: The final saliency map is binarized using the high percentile thresholding method, Otsu thresholding method, or adaptive thresholding method.

6. The image processing-based method for inspecting the coating quality of a prefabricated substation according to claim 5, characterized in that, The post-processing includes: After thresholding the final saliency map, apply morphological opening and / or morphological closing operations to the resulting binary mask. Connectivity analysis is applied to remove connected regions whose area is smaller than a preset area threshold.

Images