An AC surge arrester state detection method based on image semantic segmentation
By detecting the material interface area and suppressing high-gloss pollution, combined with the DeepLabV3+ network, the problem of specular reflection and high-gloss overflow of AC surge arresters in the process of quality inspection on the production line was solved, achieving high-precision pixel-level segmentation and status detection, and meeting the inspection requirements of the production line.
Patent Information
- Application Number
- CN202611299076.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-26
AI Technical Summary
In existing technologies, during the quality inspection of products leaving the factory on an assembly line, the difference in optical properties between the metal and insulating components of AC surge arresters leads to specular reflection and high-gloss overflow, causing semantic segmentation networks to missegment and blur the boundaries at the material interfaces, making it impossible to achieve high-precision pixel-level segmentation and state detection.
By detecting material interface regions and identifying and suppressing specular highlight overflow between metal and insulating components, the system identifies and corrects specular highlight overflow between metal and insulating components. Semantic segmentation is performed using the DeepLabV3+ network, outputting clear pixel-level segmentation results.
It achieves high-precision pixel-level segmentation and status detection of AC surge arresters under strong lighting conditions on the production line, meets the dynamic detection cycle requirements of the production line, avoids loss of texture details, and improves the accuracy and efficiency of detection.
Smart Images

Figure CN122821265B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of AC surge arrester image detection technology, specifically to an AC surge arrester state detection method based on image semantic segmentation. Background Technology
[0002] As a key device for overvoltage protection in power systems, the quality of production and assembly of AC surge arresters and the integrity of their appearance directly affect the safety of power grid operation. Before leaving the factory, the correctness of component assembly, surface defects and dimensional tolerances of surge arresters must be strictly inspected. Traditional manual visual inspection or caliper measurement methods have problems such as low efficiency, strong subjectivity and easy fatigue and missed inspections, which are difficult to meet the requirements of modern intelligent production lines for high throughput, high precision and fully automatic quality inspection. With the development of machine vision and deep learning technologies, surge arrester detection methods based on image recognition have been extensively studied. Among the existing technologies, there are surge arrester recognition methods based on improved YOLOv3-Tiny and knowledge distillation, surge arrester external defect detection methods based on improved YOLOv8, and surge arrester defect detection methods based on UAV inspection. The above methods mainly use target detection networks to identify and locate the entire surge arrester or specific defect areas. In image segmentation, semantic segmentation methods based on DeepLabV3+ have been applied to the automated segmentation of insulator images. Some researchers have cascaded DeepLabV3+ with YOLOv5 to achieve semantic segmentation and defect detection of insulators in aerial images. There are also insulator fault identification methods that combine lightweight MobileNet-SSD with MobileNetV2-DeeplabV3+. In addition, a fast detection method for distribution line insulators and surge arresters based on improved YOLOv4 has also been proposed. The above technologies show that applying deep learning object detection and semantic segmentation technologies to the visual inspection of power equipment has become the mainstream technology direction in this field. However, the aforementioned existing technologies are mainly geared towards the inspection of surge arresters and insulators under drone inspections or outdoor natural light conditions. Their imaging environment differs fundamentally from that of assembly line quality inspection scenarios. In assembly line quality inspection scenarios, standard industrial high-intensity lighting conditions are typically used to ensure imaging consistency, defect visibility, and inspection cycle requirements. However, under these conditions, AC surge arresters are tightly assembled from high-reflectivity metal components (such as polished equalizing rings and flanges) and low-reflectivity insulating components (such as ceramic or silicone rubber skirts). Under standard industrial high-intensity light, the metal surfaces will produce strong specular reflections. The surface of the metal component reflects specular highlights, while the surface of the insulating component does not produce specular reflection due to its material properties. This difference in optical characteristics leads to a specific problem that has not been fully addressed in the existing technology: the specular highlights of the metal surface will physically "overflow" and intrude into the pixel area of the edge of the adjacent insulating component, causing image feature contamination. This highlight overflow effect causes existing semantic segmentation networks to systematically missegment at the material interface between the metal component and the insulating component. Specifically, this manifests as: blurred component outlines, material boundary adhesion, or incorrect segmentation of the metal highlight area as a crack in the insulating component, etc. To address the specific segmentation challenges caused by differences in material optical properties, existing solutions have significant limitations. At the hardware level, global polarizing filter filtering can suppress specular reflection, but it indiscriminately darkens the entire image, resulting in the loss of texture details on the surface of dark insulating parts. Multi-frame high dynamic range synthesis technology can eliminate overexposed areas, but it requires the subject to remain stationary during imaging, which conflicts with the dynamic detection rhythm of the pipeline. At the software level, there are existing methods for removing highlights from metal surfaces, such as threshold segmentation, histogram specification, reflection component separation combined with pixel filling, and deep learning methods based on secondary encoding / decoding structures or attention mechanisms. However, these methods are all general highlight suppression techniques for metal parts, and their processing targets are the highlight areas on the surface of the metal body, rather than the specific physical phenomenon of "metal highlights overflowing into adjacent insulating material pixel areas." In addition, edge connection methods based on geometric priors cannot handle complex umbrella-shaped curve edges. Conventional image inpainting or data enhancement methods can only "blindly" fill or ignore highlight areas, and cannot reliably restore the real material boundaries that are submerged by highlights at the pixel level. Therefore, there is an urgent need for a technical solution that can identify and solve the specific problem of "physical overflow of metallic highlights into the pixel area of insulating material" from the perspective of physical imaging mechanism. Through targeted image preprocessing and the synergistic cooperation of semantic segmentation network, high-precision pixel-level segmentation and status detection of each component of AC surge arrester can be achieved in the process of quality inspection of AC surge arrester in the assembly line. Summary of the Invention
[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an AC surge arrester status detection method based on image semantic segmentation, thereby solving the aforementioned problems in the prior art.
[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an AC surge arrester state detection method based on image semantic segmentation, wherein the AC surge arrester is tightly assembled from metal components and insulating components, and the method includes the following steps: Step S1: Acquire raw images of the AC surge arrester under standard industrial high-intensity lighting conditions on the production line; Step S2: Detect the boundary region between metal and insulating materials in the original image to identify the boundary region between metal and insulating components in the original image; Step S3: Perform specular contamination detection and suppression processing on the material interface area to generate a corrected image; the specular contamination detection and suppression processing is based on the following physical mechanism: under standard industrial strong light illumination, the specular reflection specular light generated on the surface of the metal part physically overflows along the material interface area to the pixel area on the insulating part side, causing the brightness value of the edge pixels on the insulating part side to increase and the saturation value to decrease. Step S4: Input the corrected image into the pre-trained image semantic segmentation network, and output the pixel-level semantic segmentation results of each component of the AC surge arrester; Step S5: Based on the semantic segmentation results, detect the assembly status and surface defects of the AC surge arrester, and output the status detection results.
[0005] Preferably, in step S2, the detection of the material interface area includes: Step S21: Based on the location of the abrupt change in grayscale gradient between the metal component region and the insulating component region in the original image, and combined with the prior material location information of the metal component and the insulating component, the boundary location between the metal component and the insulating component is initially located. Step S22: Centered on the boundary position, extend the boundary line along the normal direction to the metal component side and the insulating component side by 5 to 15 pixels to form a material boundary area.
[0006] Preferably, in step S3, the high light pollution determination and suppression process includes: Step S31: Within the material boundary area, calculate the brightness and saturation values of each pixel. Pixels with a brightness value greater than 220, a grayscale value in the range of 0 to 255, a saturation value less than 30, and an HSV color space saturation in the range of 0 to 100 are identified as pixel contamination points caused by the overflow of specular reflection from the metal mirror. All pixel contamination points constitute a pixel contamination area. Step S32: For each contaminated pixel within the contaminated pixel region, using the contaminated pixel as the center, search for non-contaminated pixels within the material boundary region along a direction parallel to the boundary line with a radius of 3 to 7 pixels. Use the pixel values of the searched non-contaminated pixels to perform weighted interpolation calculations using the following formula. Replace the original pixel value of the contaminated pixel with the interpolation result to complete the pixel value correction.
[0007] in, The pixel value after correction for the pixel contamination point (x, y). Let i be the pixel value of the i-th uncontaminated pixel found in the search. The distance between the uncontaminated pixel and the contaminated pixel is the reciprocal of the spatial distance between them, and n is the total number of uncontaminated pixels found.
[0008] Preferably, in step S4, the image semantic segmentation network is a DeepLabV3+ network; The encoder of the DeepLabV3+ network uses the Xception-65 backbone network to extract image features and captures multi-scale contextual information through the dilated spatial pyramid pooling module, which contains parallel dilated convolutional layers with dilation rates of 6, 12 and 18. The DeepLabV3+ network's decoder sequentially upsamples the low-resolution feature map output by the encoder by 4x, concatenates it with the feature map from the encoder's intermediate layer, and upsamples it by 4x again to restore the resolution of the original image, so as to output pixel-level classification results with clear edges at material boundary regions.
[0009] Preferably, the training steps for the image semantic segmentation network include: Step S41: Obtain multiple sample images containing AC surge arresters. The sample images are labeled with pixel-level semantic tags corresponding to the equalizing ring, flange, porcelain bushing skirt and bolts, and the edge pixels of the material interface area are densely outlined. Step S42: Perform the same material boundary region detection and specular pollution suppression preprocessing as in steps S2 and S3 on the sample images to generate a training sample set; Step S43: Train the image semantic segmentation network using the training sample set. Use a weighted combination of cross-entropy loss function and boundary loss function as the optimization objective. Use the Adam optimizer, set the initial learning rate to 0.001, and the batch size to 8. Determine the network convergence when the loss function value tends to stabilize.
[0010] Preferably, in step S5, detecting the assembly status of the AC surge arrester includes: extracting the outlines of the metal components and the insulation components according to the semantic segmentation results; calculating the minimum Euclidean distance between the outlines of the metal components and the outlines of the insulation components at the preset assembly position as the actual assembly gap; comparing the actual assembly gap with the preset standard assembly gap threshold; and if the actual assembly gap exceeds the allowable deviation range of the standard assembly gap threshold, then the AC surge arrester is determined to be unqualified in assembly.
[0011] Preferably, the metal components include an equalizing ring and a flange, and the insulating components include a porcelain bushing skirt; In step S5, the actual assembly gap includes the first assembly gap between the equalizing ring and the porcelain bushing skirt, and the second assembly gap between the flange and the porcelain bushing skirt. The first assembly gap and the second assembly gap are compared with the corresponding standard assembly gap thresholds. If either assembly gap exceeds the allowable deviation range, the assembly is deemed unqualified.
[0012] Preferably, in step S5, the detection of surface defects in the AC surge arrester includes: Pixel features of the insulating component region are extracted based on the semantic segmentation results. The pixel features are the gray values of each pixel in the insulating component region. Calculate the local mean and local standard deviation of pixel grayscale values within the insulating component area. Mark continuous pixel areas with grayscale values lower than the local mean minus 3 times the local standard deviation as abnormal dark areas. When the continuous area of abnormal dark areas is greater than a preset area threshold, the area is determined to be a crack defect.
[0013] Preferably, in step S1, the standard industrial high-intensity lighting is an LED array light source with an illuminance of 1000 to 2000 lux and a color temperature of 5000 to 6500K; The illuminance range and color temperature range are configured to simultaneously satisfy: generating specular highlights on the surface of the metal component sufficient to trigger the specular contamination determination in step S3, and maintaining the visibility of the umbrella skirt texture details on the surface of the insulating component.
[0014] (III) Beneficial Effects This invention provides a method for detecting the state of AC surge arresters based on image semantic segmentation, which has the following advantages: (1) Existing surge arrester image detection technology is mainly aimed at UAV inspection or outdoor natural lighting scenarios, and does not address the problem of reduced segmentation accuracy caused by differences in material optical properties under strong light illumination on assembly lines; existing general highlight removal methods only suppress the highlights on the surface of the metal body, and do not pay attention to the specific physical phenomenon of "metal highlights overflowing into adjacent insulating material pixel areas". Based on the in-depth analysis of the above physical overflow mechanism, this invention eliminates the highlight pollution in the semantic segmentation input image from the physical imaging level through the synergistic cooperation of "material boundary area detection" and "highlight pollution suppression processing", so that the semantic segmentation network can obtain a high-quality input image after correction, thereby outputting pixel-level segmentation results with clear edges and accurate categories at the material boundary.
[0015] (2) In the prior art, specular suppression and semantic segmentation are usually processed as independent modules, and there is no logical connection between them based on physical mechanism. This invention uses specular pollution suppression based on physical overflow mechanism as a pre-step of semantic segmentation, which upgrades the preprocessing stage from "blind general image enhancement" to "physical cause-based directional correction". The corrected image retains the texture details of the insulating component surface to the maximum extent, while eliminating the pixel pollution of the material boundary area by the metal specular, providing the semantic segmentation network with higher information fidelity input.
[0016] (3) The method of the present invention only requires a single frame image to complete the high light pollution judgment and suppression processing, without the need for multi-frame synthesis, which meets the requirements of the pipeline dynamic rhythm for real-time processing, while avoiding the problem of texture detail loss caused by the global polarizing filter indiscriminately darkening the image, and achieving a balance between metal high light suppression and insulation texture preservation. Attached Figure Description
[0017] Figure 1 A flowchart of the AC surge arrester state detection method based on image semantic segmentation provided by the present invention; Figure 2 A schematic diagram of the module structure of the AC surge arrester state detection method based on image semantic segmentation provided by the present invention; Figure 3 The image provided by this invention shows an AC surge arrester under industrial strong light irradiation. Figure 4 This is a restored image of an AC surge arrester captured under strong industrial light, provided by the present invention. Figure 5 This is a 20×20 pixel dot image showing the high-gloss overflow reflected from the metal mirror surface of an AC surge arrester under industrial strong light irradiation, provided by the present invention. Figure 6 This is a 20×20 pixel image showing the repair of high-gloss overflow from the metal mirror surface of an AC surge arrester under industrial strong light irradiation, provided by the present invention.
[0018] In the image: 1. Uncontaminated pixels in the metal component area; 2. Uncontaminated pixels in the insulating component area; 3. Contaminated pixels at the boundary line before correction; 4. Contaminated pixels at the boundary line after correction. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Please combine Figures 1-6 This invention provides a method for detecting the state of an AC surge arrester based on image semantic segmentation. The AC surge arrester is composed of metal parts and insulating parts tightly assembled. The method includes the following steps: Step S1: Acquire raw images of the AC surge arrester under standard industrial high-intensity lighting conditions on the production line; Step S2: Detect the boundary region between metal and insulating materials in the original image to identify the boundary region between metal and insulating components in the original image; Step S3: Perform specular contamination detection and suppression processing on the material interface area to generate a corrected image; the specular contamination detection and suppression processing is based on the following physical mechanism: under standard industrial strong light illumination, the specular reflection specular light generated on the surface of the metal part physically overflows along the material interface area to the pixel area on the insulating part side, causing the brightness value of the edge pixels on the insulating part side to increase and the saturation value to decrease. Step S4: Input the corrected image into the pre-trained image semantic segmentation network, and output the pixel-level semantic segmentation results of each component of the AC surge arrester; Step S5: Based on the semantic segmentation results, detect the assembly status and surface defects of the AC surge arrester, and output the status detection results.
[0021] In this embodiment: In step S1, the AC surge arrester is tightly assembled from metal components and insulating components. The metal components include an equalizing ring and a flange, and the insulating components include a porcelain bushing skirt. At the quality inspection station on the production line, an industrial camera and a standard industrial high-intensity lighting source are configured. The standard industrial high-intensity lighting is an LED array light source with an illuminance of 1000 to 2000 lux and a color temperature of 5000 to 6500K. This illuminance range and color temperature range are configured to simultaneously meet two conditions: first, to generate specular reflection highlights on the surface of the metal components that are sufficient to trigger subsequent high-light pollution judgment; and second, to maintain the visibility of the details of the skirt texture on the surface of the insulating components. Under the control of the production line cycle, the industrial camera triggers acquisition when the AC surge arrester runs to the shooting station, acquiring the original image of the AC surge arrester. The original image is an RGB color image, and the resolution can be set according to the detection accuracy requirements, such as 1920×1080 pixels or higher. The acquired original image is transmitted to the industrial control computer or server through the image acquisition card for subsequent processing. Before acquiring the original image, the industrial camera is calibrated to correct lens distortion and ensure the accuracy of subsequent dimensional measurements. The camera calibration can use the Zhang Zhengyou calibration method to obtain the camera's intrinsic parameter matrix and distortion coefficients.
[0022] In step S2, sub-step S21: Based on the location of the abrupt change in grayscale gradient between the metal component region and the insulating component region in the original image, and combined with the prior material location information of the metal component and the insulating component, the boundary position between the metal component and the insulating component is initially located. Specifically, the original RGB image is first converted to a grayscale image. The conversion can be done using a weighted average method: Gray = 0.299 × R + 0.587 × G + 0.114 × B. Then, edge detection is performed on the grayscale image. The Canny edge detection operator is preferred. Its low threshold and high threshold can be set to 50 and 150, respectively. The Canny edge detection outputs a binary edge image, where the edge pixels are the locations of abrupt changes in grayscale gradient. Since the components of an AC surge arrester have roughly fixed geometric positional relationships on the production line, a material prior location information database can be pre-established based on the arrester's model and size. This database records the approximate area range of each component (equalizing ring, flange, porcelain bushing skirt, etc.) in the image. The Canny edge detection results are matched with the material prior location information. That is, in the adjacent area of the metal component and the insulating component indicated by the prior location information, the edge pixel points located at the boundary between the two are selected as the initial location boundary. Sub-step S22: Centered on the boundary position, extend 5 to 15 pixels wide along the normal direction of the boundary line towards the metal component side and the insulating component side respectively to form the material boundary area; Specifically, for each boundary pixel located in sub-step S21, the normal direction of the boundary line at that point is calculated. The normal direction can be determined by the direction of the image gradient at that point. The gradient direction is perpendicular to the edge direction, which is the normal direction. The normal direction is extended to the metal part side and the insulating part side along the normal direction to form a rectangular material boundary region. The union of the material boundary regions corresponding to all boundary positions is the material boundary region finally output in step S2. The predetermined pixel width is adaptively determined based on the resolution of the original image and the ratio between the actual size of the surge arrester and the image size. For example, when the image resolution is 1920×1080 and the imaging height of the surge arrester accounts for about 60% of the image height, the predetermined pixel width can be set to 10 pixels.
[0023] In step S3, the high-light pollution detection and suppression processing is based on the following physical mechanism: Under standard industrial strong lighting, specular reflections generated on the surface of metal components (such as equalizing rings and flanges) physically overflow along the material interface area to the pixel area on the insulating component side (such as ceramic sleeve skirt), causing the brightness value of the edge pixels on the insulating component side to increase and the saturation value to decrease. The specific implementation method of high-light pollution detection and suppression processing is as follows: Sub-step S31: Within the material boundary area, calculate the brightness and saturation values of each pixel. Pixels with a brightness value greater than 220 (grayscale value range of 0 to 255) and a saturation value less than 30 (HSV color space saturation range of 0 to 100) are identified as pixel contamination points caused by the overflow of specular reflection from the metal mirror. All pixel contamination points constitute a pixel contamination area. Specifically, the original RGB image is converted to the HSV color space using the following formula:
[0024]
[0025] Where V represents the brightness value, ranging from 0 to 255, and S represents the saturation value, ranging from 0 to 1. Multiplying by 100 gives the value from 0 to 100. Within the material boundary area, the brightness value V and saturation value S of each pixel are judged. If the V of a certain pixel is greater than 220 and the S×100 is less than 30, then the pixel is determined to be a pixel pollution point. All pixels determined to be pixel pollution points constitute a pixel pollution area. It should be noted that the above threshold values of 220 for brightness and 30 for saturation are empirical values for 8-bit RGB images (0 to 255) in HSV space. In practical applications, they can be adjusted adaptively according to specific lighting conditions and camera parameters. For example, the brightness threshold can be set between 210 and 230, and the saturation threshold can be set between 20 and 40. Sub-step S32: For each pixel contamination point in the pixel contamination area, with the pixel contamination point as the center, search for non-contamination pixel points in the neighborhood of the material boundary area along a direction parallel to the boundary line with a radius of 3 to 7 pixels. Use the pixel values of the searched non-contamination pixel points to perform weighted interpolation calculation, and replace the original pixel value of the pixel contamination point with the interpolation result to complete the pixel value correction. Specifically, for each contaminated pixel P(x, y) within the contaminated pixel region, the direction of the boundary line at the location of the point is first determined. The direction of the boundary line can be obtained by rotating the gradient direction of the point by 90 degrees. Then, with P(x, y) as the center, non-contaminated pixel points in the neighborhood are searched along the direction parallel to the boundary line (i.e., the tangent of the boundary line) with a preset radius R (R takes the value of 3 to 7 pixels, preferably 5 pixels). The search is only performed within the material boundary region and does not exceed the boundary of the material boundary region. Suppose there are n uncontaminated pixels found in the search, and their pixel values are respectively The spatial distances between each uncontaminated pixel and the contaminated pixel P(x,y) are respectively The weighted interpolation is then calculated using the following formula:
[0026] in, The pixel value after correction for pixel contamination (x, y), weighted. The search radius R is preferably 5 pixels. This value takes into account both computational efficiency and repair effect. If the radius is too small, it may not be possible to find enough non-contaminated pixels. If the radius is too large, it may introduce pixel information that is too far away, resulting in repair distortion. If the number of non-contaminated pixels n found within the preset radius is less than 3, the search radius is appropriately expanded (e.g., expanded to 10 pixels) and the search is repeated until at least 3 non-contaminated pixels are found. Corrected pixel values Replace the original pixel value of the contaminated pixel point P(x,y) (including replacing the pixel values of the R, G, and B channels) to complete the correction of this pixel point. After completing the above correction for all contaminated pixels in the contaminated pixel area, the corrected image is obtained.
[0027] In step S4, the image semantic segmentation network is a DeepLabV3+ network. The encoder of the DeepLabV3+ network uses the Xception-65 backbone network to extract image features and captures multi-scale contextual information through the dilated spatial pyramid pooling module. The ASPP module contains parallel dilated convolutional layers with dilation rates of 6, 12, and 18, as well as a global average pooling branch. The outputs of each branch are fused through 1×1 convolution. The decoder of the DeepLabV3+ network sequentially upsamples the low-resolution feature map output by the encoder by 4x, concatenates it with the feature map of the encoder's intermediate layer, and then upsamples it again by 4x to restore the resolution of the original image. Specifically, the feature map output by the encoder is first upsampled by 4x through bilinear interpolation, then concatenated with the feature map of the encoder's intermediate layer after 1×1 convolution for dimensionality reduction. The concatenated feature map is then refined by a 3×3 convolutional layer, and finally upsampled by 4x through bilinear interpolation again to restore the same resolution as the original input image. This decoder structure can output pixel-level classification results with clear edges at material boundary regions. The output of the semantic segmentation network is a semantic segmentation map with the same resolution as the original image. Each pixel in the map is assigned a semantic label. The types of semantic labels include: equalizing ring, flange, porcelain umbrella skirt, bolt, and background. The pixel values corresponding to each category can be preset, for example: background=0, equalizing ring=1, flange=2, porcelain umbrella skirt=3, bolt=4. The training process of the image semantic segmentation network is as follows: Sub-step S41: Obtain multiple sample images containing AC surge arresters. The sample images are labeled with pixel-level semantic tags corresponding to the equalizing ring, flange, porcelain bushing skirt and bolt. The annotation accuracy of the semantic tags of each component is pixel-level, and the edge pixels of the material interface area are densely outlined. Specifically, a large number of raw images of AC surge arresters were collected at the assembly line workstation, covering samples of different models, orientations, and lighting conditions. After collection, professional annotators used image annotation tools to perform pixel-level semantic annotation on each sample image. During annotation, the contour edges of the equalizing ring, flange, porcelain bushing skirt, and bolts were accurately depicted. In particular, the edge pixels of the material interface area between the metal parts and the insulating parts were required to be densely annotated, that is, the pixels within 2 to 3 pixels on both sides of the interface line were required to be accurately labeled with their category, so as to ensure that the network can learn the fine segmentation boundary at the material interface. Sub-step S42: Perform the same material boundary region detection and specular pollution suppression preprocessing as in steps S2 and S3 on the sample image to generate a training sample set; That is, for each labeled sample image, step S2 (material boundary region detection) and step S3 (spectral pollution judgment and suppression processing) are executed in sequence to generate a corrected sample image. The corrected sample image and its corresponding pixel-level semantic label map together constitute a training sample. The set of all training samples is the training sample set. Sub-step S43: Train the image semantic segmentation network using the training sample set, using a weighted combination of cross-entropy loss function and boundary loss function as the optimization objective, using the Adam optimizer, setting the initial learning rate to 0.001, setting the batch size to 8, and determining network convergence when the loss function value tends to stabilize. Specifically, the cross-entropy loss function The formula used to measure the difference between the predicted class and the true class for each pixel is as follows:
[0028] in, Represented as the cross-entropy loss function, where C is the total number of classes. This represents the true label (0 or 1) of the pixel belonging to category c. Predict the probability that the pixel belongs to class c for the network; Boundary loss function To enhance the network's segmentation accuracy of material boundary regions, the calculation method is as follows: extract the boundary pixels of the predicted segmentation map and the boundary pixels of the real label map, and calculate the Dice loss between the two. The introduction of the boundary loss function can make the network pay more attention to the edge pixels of the material boundary region. The total loss function is: ,in These are the weighting coefficients for the boundary loss function, ranging from 0.1 to 0.5. During training, the training sample set is divided into training set, validation set and test set in a ratio of 8:1:1. The training set is used to update the network parameters. After each epoch, the validation set is used to evaluate the network performance. When the loss function value of the validation set no longer decreases for 10 consecutive epochs, the network is considered to have converged and training is stopped. Finally, the test set is used to perform the final performance evaluation of the trained network.
[0029] In step S5, the specific implementation methods for assembly status inspection and surface defect inspection are as follows: Assembly status detection: Based on the semantic segmentation results, the outlines of the metal parts and the insulating parts are extracted respectively. The minimum Euclidean distance between the outlines of the metal parts and the outlines of the insulating parts at the preset assembly position is calculated as the actual assembly gap. The actual assembly gap is compared with the preset standard assembly gap threshold. If the actual assembly gap exceeds the allowable deviation range of the standard assembly gap threshold, the AC surge arrester is determined to be unqualified in assembly. Specifically, the contour pixel set of the equalizing ring and the contour pixel set of the porcelain sleeve skirt are extracted from the semantic segmentation results. The Euclidean distance between each pixel on the contour of the equalizing ring and each pixel on the contour of the porcelain sleeve skirt is calculated, and the minimum value is taken as the first assembly gap between the equalizing ring and the porcelain sleeve skirt. Similarly, the minimum Euclidean distance between the flange contour and the porcelain sleeve skirt contour is calculated as the second assembly gap. The first assembly gap and the second assembly gap are compared with the corresponding standard assembly gap thresholds. The standard assembly gap thresholds are preset and stored in the system according to the design drawings and tolerance requirements of the AC surge arrester. When any assembly gap exceeds the allowable deviation range, the AC surge arrester is judged to be unqualified in assembly. Surface defect detection: Pixel features of the insulating component region are extracted based on the semantic segmentation results. The pixel features are the gray values of each pixel in the insulating component region. The local mean and local standard deviation of the pixel gray values in the insulating component region are calculated. Continuous pixel regions with gray values lower than the local mean minus 3 times the local standard deviation are marked as abnormal dark areas. When the continuous area of the abnormal dark area is greater than a preset area threshold, the region is determined to be a crack defect. Specifically, all pixel regions labeled "porcelain umbrella skirt" are extracted from the semantic segmentation results. Within these regions, a sliding window approach is used to calculate the local mean and local standard deviation. The window size can be set to 15×15 pixels. For the center pixel of the window, if its grayscale value is lower than... (in The average grayscale value of the pixels within the window. If the standard deviation of pixel grayscale within the window is 0, then the pixel is marked as a candidate outlier. Connectivity analysis is performed on all candidate outliers, and interconnected candidate outliers are grouped into anomaly regions. The area of each anomaly region (i.e. the number of pixels it contains) is calculated. If the area of an anomaly region is greater than a preset area threshold, then the region is determined to be a crack defect. It should be noted that the preset area threshold can be adjusted according to the image resolution and detection sensitivity requirements. For higher resolution images, the area threshold can be increased accordingly to avoid false alarms caused by noise. The status inspection results include assembly status inspection results (pass, fail and the reason for failure) and surface defect inspection results (whether there are cracks and defects and the location of the defects), which are output to the production management system via a display or transmitted to the production management system via a network.
[0030] It is worth mentioning regarding this embodiment that: Existing surge arrester image detection technologies are primarily geared towards drone inspections or outdoor natural lighting scenarios, failing to address the segmentation accuracy degradation caused by differences in material optical properties under strong illumination in production lines. Current general highlight removal methods only suppress highlights on the metal surface, neglecting the specific physical phenomenon of "metal highlights spilling over into adjacent insulating material pixel areas," such as... Figure 5 As shown, there are contaminated pixels at the boundary line between the uncontaminated pixels in the metal component region and the uncontaminated pixels in the insulating component region before correction. Based on an in-depth analysis of the above-mentioned physical overflow mechanism, this invention eliminates the specular contamination in the semantic segmentation input image from the physical imaging level through the synergistic cooperation of "material boundary region detection" and "spectral contamination suppression processing". Figure 6 As shown, the contaminated pixels at the boundary between the uncontaminated pixels in the metal component region and the uncontaminated pixels in the insulating component region are suppressed and corrected to form corrected contaminated pixels at the boundary, enabling the semantic segmentation network to obtain a corrected high-quality input image, thereby outputting pixel-level segmentation results with clear edges and accurate categories at the material boundary. In existing technologies, specular suppression and semantic segmentation are usually processed as independent modules, lacking a logical connection based on physical mechanisms. This invention uses specular contamination suppression based on physical overflow mechanisms as a pre-processing step for semantic segmentation, upgrading the preprocessing stage from "blind general image enhancement" to "physical cause-based directional correction". The corrected image retains the texture details of the insulating component surface to the maximum extent, while eliminating pixel contamination of material boundary areas by metallic specular highlights, providing the semantic segmentation network with higher information fidelity input. The method of this invention only requires a single frame image to complete the detection and suppression of specular highlight pollution, eliminating the need for multi-frame synthesis. This meets the real-time processing requirements of a pipeline dynamic rhythm, while avoiding the loss of texture details caused by indiscriminate darkening of the image due to global polarizing filter filtering. It achieves a balance between suppressing metallic specular highlights and preserving insulating textures. Figure 3 and Figure 4 As shown, through the identification and suppression of high light pollution, Figure 3 Image correction suppression for loss of mid-texture detail is achieved Figure 4 The image shows a clear surface texture of the surge arrester, which helps to improve the accuracy of surface texture detection of the surge arrester.
[0031] In step S2, the detection of the material boundary area includes: Step S21: Based on the location of the abrupt change in grayscale gradient between the metal component region and the insulating component region in the original image, and combined with the prior material location information of the metal component and the insulating component, the boundary location between the metal component and the insulating component is initially located. Step S22: Centered on the boundary position, extend the boundary line along the normal direction to the metal component side and the insulating component side by 5 to 15 pixels to form a material boundary area.
[0032] In this embodiment, by employing the dual constraints of grayscale gradient abrupt changes and prior material location information, the material boundary between the metal and insulating components can be quickly and accurately located, avoiding the computational overhead and false detection risks associated with blindly searching the entire image. Expanding the boundary line along its normal direction by a predetermined pixel width ensures that the material boundary region completely covers all pixels that might overflow from the metal specular highlights, without omission or over-expansion, providing accurate candidate regions for subsequent specular contamination detection. Compared to existing technologies that perform specular detection across the entire image, this invention reduces the pixel range for specular contamination detection to a small proportion of the total image area—the material boundary region—significantly reducing computational load and improving processing efficiency.
[0033] In step S3, the determination and suppression of high light pollution includes: Step S31: Within the material boundary area, calculate the brightness and saturation values of each pixel. Pixels with a brightness value greater than 220, a grayscale value in the range of 0 to 255, a saturation value less than 30, and an HSV color space saturation in the range of 0 to 100 are identified as pixel contamination points caused by the overflow of specular reflection from the metal mirror. All pixel contamination points constitute a pixel contamination area. Step S32: For each contaminated pixel within the contaminated pixel region, using the contaminated pixel as the center, search for non-contaminated pixels within the material boundary region along a direction parallel to the boundary line with a radius of 3 to 7 pixels. Use the pixel values of the searched non-contaminated pixels to perform weighted interpolation calculations using the following formula. Replace the original pixel value of the contaminated pixel with the interpolation result to complete the pixel value correction.
[0034] in, The pixel value after correction for the pixel contamination point (x, y). Let i be the pixel value of the i-th uncontaminated pixel found in the search. The distance between the uncontaminated pixel and the contaminated pixel is the reciprocal of the spatial distance between them, and n is the total number of uncontaminated pixels found.
[0035] In this embodiment: After the specular reflection from the metal mirror penetrates the edge pixel area of the insulating component, its most significant characteristic is a sharp increase in brightness and a significant decrease in saturation. Although the single threshold determination of brightness can identify overly bright pixels, it is easy to misjudge the natural bright spots on the surface of the insulating component as specular pollution. However, the dual threshold joint determination can accurately distinguish between "high-brightness-low-saturation pixels caused by specular overflow" and "high-brightness-high-saturation or normal texture pixels of the insulating component itself", which significantly reduces the misjudgment rate. Compared with the isotropic neighborhood search interpolation in the prior art, the present invention fully considers the physical characteristics of the specular reflection of the metal mirror overflowing in a strip-like distribution along the material boundary line. By searching for non-contaminated pixels along the direction parallel to the boundary line, it can ensure that the neighboring pixels participating in the interpolation calculation and the contaminated pixels to be repaired have higher consistency in material properties.
[0036] In step S4, the image semantic segmentation network is the DeepLabV3+ network; The encoder of the DeepLabV3+ network uses the Xception-65 backbone network to extract image features and captures multi-scale contextual information through the dilated spatial pyramid pooling module, which contains parallel dilated convolutional layers with dilation rates of 6, 12 and 18. The DeepLabV3+ network's decoder sequentially upsamples the low-resolution feature map output by the encoder by 4x, concatenates it with the feature map from the encoder's intermediate layer, and upsamples it by 4x again to restore the resolution of the original image, so as to output pixel-level classification results with clear edges at material boundary regions.
[0037] In this embodiment, ASPP modules with expansion ratios of 6, 12, and 18 can effectively cover target areas of different scales. The small expansion ratio of 6 captures fine local edge features (such as the precise location of material boundaries), the medium expansion ratio of 12 captures medium-range contextual information (such as the overall shape of a single skirt), and the large expansion ratio of 18 captures global contextual information (such as the structural layout of the entire surge arrester). The fusion of these three features makes the network more robust to the segmentation of various components of the surge arrester, especially suitable for AC surge arrester segmentation tasks composed of tightly assembled components with significant differences in size (small skirt edges and large equalizing rings). The encoder's intermediate layer feature map retains rich spatial location information and edge details. It is then spliced and fused with the upsampled high-level semantic features, so that the segmentation map output by the decoder has clear and sharp boundaries in the material boundary region, providing a precise pixel-level localization basis for subsequent size measurement and defect detection. The weighted interpolation formula uses the reciprocal of the spatial distance as the weight, so that the closer the non-polluted pixel is to the contaminated point of the pixel to be repaired, the greater the contribution. This ensures the spatial continuity and smoothness of the repaired pixel value and avoids abrupt grayscale jumps between the repaired pixel and the surrounding normal pixels.
[0038] The training steps for an image semantic segmentation network include: Step S41: Obtain multiple sample images containing AC surge arresters. The sample images are labeled with pixel-level semantic tags corresponding to the equalizing ring, flange, porcelain bushing skirt and bolts, and the edge pixels of the material interface area are densely outlined. Step S42: Perform the same material boundary region detection and specular pollution suppression preprocessing as in steps S2 and S3 on the sample images to generate a training sample set; Step S43: Train the image semantic segmentation network using the training sample set. Use a weighted combination of cross-entropy loss function and boundary loss function as the optimization objective. Use the Adam optimizer, set the initial learning rate to 0.001, and the batch size to 8. Determine the network convergence when the loss function value tends to stabilize.
[0039] In this embodiment: During the training phase, the same material boundary region detection and specular pollution suppression preprocessing is performed on the sample images as in the detection phase. This ensures the consistency of the distribution of training data and inference data, and avoids the decline in segmentation performance caused by the distribution offset of training and inference data. Since specular pollution suppression processing will change the numerical distribution of some pixels in the material boundary region, if the training data does not undergo the same preprocessing, the network will not be able to learn the pixel distribution features of the corrected image, resulting in a decrease in segmentation accuracy during the inference phase. By employing a weighted combination of cross-entropy loss and boundary loss to optimize the objective, the network significantly enhances its attention to edge segmentation in material boundary regions while maintaining overall pixel classification accuracy. The boundary loss function enables the network to generate clearer classification boundaries on edge pixels at the junctions of metal and insulating materials, effectively solving the problems of blurred segmentation edges and material adhesion caused by high light pollution in existing technologies. Dense edge annotation is applied to edge pixels in material boundary regions, further strengthening the edge supervision signal of the training data, enabling the network to output accurate pixel-level classification results in the most challenging material boundary regions.
[0040] In step S5, the assembly status detection of the AC surge arrester includes: extracting the outlines of the metal parts and the insulation parts according to the semantic segmentation results, calculating the minimum Euclidean distance between the outlines of the metal parts and the outlines of the insulation parts at the preset assembly position as the actual assembly gap, comparing the actual assembly gap with the preset standard assembly gap threshold, and if the actual assembly gap exceeds the allowable deviation range of the standard assembly gap threshold, the AC surge arrester is determined to be unqualified in assembly. Metal components include equalizing rings and flanges, while insulating components include porcelain bushings and skirts. In step S5, the actual assembly gap includes the first assembly gap between the equalizing ring and the porcelain bushing skirt, and the second assembly gap between the flange and the porcelain bushing skirt. The first assembly gap and the second assembly gap are compared with the corresponding standard assembly gap thresholds. If either assembly gap exceeds the allowable deviation range, the assembly is deemed unqualified.
[0041] In this embodiment: the component contour is extracted based on pixel-level semantic segmentation results. Compared with traditional manual caliper measurement or contour extraction based on edge detection, it has higher accuracy and stability. The semantic segmentation network assigns semantic labels to each pixel, and the extraction of component contour is directly based on pixel classification results, avoiding edge breakage or offset problems caused by changes in lighting and noise interference in edge detection. Using the minimum Euclidean distance as the measurement method for assembly clearance can accurately reflect the actual distance between the equalizing ring and the porcelain bushing skirt, and between the flange and the porcelain bushing skirt. The calculation method is clear and reproducible. The judgment logic of comparing the actual assembly clearance with the standard assembly clearance threshold is simple and reliable, and it is suitable for high-throughput automated quality inspection scenarios on production lines. By classifying and inspecting metal components and insulating components separately (first gap and second gap), compared to the existing approach that only performs a rough inspection of the entire surge arrester, the specific location and type of assembly defects can be located more accurately, providing more valuable inspection information for subsequent quality traceability and process improvement.
[0042] Step S5, the inspection of surface defects in the AC surge arrester includes: Pixel features of the insulating component region are extracted based on the semantic segmentation results. The pixel features are the gray values of each pixel in the insulating component region. Calculate the local mean and local standard deviation of pixel grayscale values within the insulating component area. Mark continuous pixel areas with grayscale values lower than the local mean minus 3 times the local standard deviation as abnormal dark areas. When the continuous area of abnormal dark areas is greater than a preset area threshold, the area is determined to be a crack defect.
[0043] In this embodiment, the statistical anomaly detection method using local mean and local standard deviation is adopted. Compared with the defect detection method using a fixed grayscale threshold, it has a stronger scene adaptability. Since the grayscale values of the surge arrester insulation components are different in different areas (such as the top and root of the umbrella skirt), and the insulation material colors of different models of surge arresters may be different, the generalization ability of the fixed grayscale threshold is poor. Using three times the standard deviation as the anomaly detection threshold, combined with the continuous area threshold, further removes isolated noise interference, making the detection results more reliable. This defect detection method is independent of the high light pollution suppression process, and the two together constitute the complete technical chain of this invention from segmentation to detection.
[0044] In step S1, the standard industrial high-intensity lighting is an LED array light source with an illuminance of 1000 to 2000 lux and a color temperature of 5000 to 6500K; The illuminance range and color temperature range are configured to simultaneously satisfy: generating specular highlights on the surface of the metal component sufficient to trigger the specular contamination determination in step S3, and maintaining the visibility of the umbrella skirt texture details on the surface of the insulating component.
[0045] In this embodiment: when the brightness is below 1000 lux, the specular reflection brightness of the metal surface is insufficient to trigger the high-light pollution judgment threshold (brightness value > 220) in step S3, and the "high-light pollution judgment and directional repair" mechanism of the present invention cannot be activated, thus losing the ability to actively control the metal highlights; when the brightness is above 2000 lux, the overall brightness of the image is too high, which may lead to the high-brightness area on the surface of the insulating component being misjudged as high-light pollution, and may also cause the dynamic range of the entire image to be saturated, losing detail information. A color temperature of 5000 to 6500K is the standard daylight color temperature range, which can ensure the authenticity of color reproduction and avoid misjudgment of material color due to color temperature deviation; Limiting the lighting conditions to "simultaneously satisfying two conditions" reflects the non-obvious design choice made by this invention after weighing the conflicting requirements of "triggering specular detection" and "maintaining texture visibility". This limitation further clarifies the application scenario boundary conditions of this invention, making the technical solution of this invention clearly distinguishable from existing surge arrester detection solutions for outdoor natural lighting or low light conditions, and enhancing the basis for demonstrating the inventiveness of this invention.
[0046] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0047] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0048] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for detecting the state of an AC arrester based on image semantic segmentation, characterized in that, An AC surge arrester is assembled from metal components and insulating components in a tightly fitted manner. The method includes the following steps: Step S1: Acquire raw images of the AC surge arrester under standard industrial high-intensity lighting conditions on the production line; Step S2: Detect the boundary region between metal and insulating materials in the original image to identify the boundary region between metal and insulating components in the original image; Step S3: Perform specular contamination detection and suppression processing on the material interface area to generate a corrected image; the specular contamination detection and suppression processing is based on the following physical mechanism: under standard industrial strong light illumination, the specular reflection specular light generated on the surface of the metal part physically overflows along the material interface area to the pixel area on the insulating part side, causing the brightness value of the edge pixels on the insulating part side to increase and the saturation value to decrease. Step S4: Input the corrected image into the pre-trained image semantic segmentation network, and output the pixel-level semantic segmentation results of each component of the AC surge arrester; Step S5: Based on the semantic segmentation results, detect the assembly status and surface defects of the AC surge arrester, and output the status detection results; In step S2, the detection of the material boundary area includes: Step S21: Based on the location of the abrupt change in grayscale gradient between the metal component region and the insulating component region in the original image, and combined with the prior material location information of the metal component and the insulating component, the boundary location between the metal component and the insulating component is initially located. Step S22: Centered on the boundary position, extend the boundary line along the normal direction to the metal component side and the insulating component side by 5 to 15 pixels to form a material boundary area.
2. The method for detecting the status of an AC surge arrester based on image semantic segmentation according to claim 1, characterized in that: In step S3, the determination and suppression of high light pollution includes: Step S31: Within the material boundary area, calculate the brightness and saturation values of each pixel. Pixels with a brightness value greater than 220, a grayscale value in the range of 0 to 255, a saturation value less than 30, and an HSV color space saturation in the range of 0 to 100 are identified as pixel contamination points caused by the overflow of specular reflection from the metal mirror. All pixel contamination points constitute a pixel contamination area. Step S32: For each contaminated pixel within the contaminated pixel region, using the contaminated pixel as the center, search for non-contaminated pixels within the material boundary region along a direction parallel to the boundary line with a radius of 3 to 7 pixels. Use the pixel values of the searched non-contaminated pixels to perform weighted interpolation calculations using the following formula. Replace the original pixel value of the contaminated pixel with the interpolation result to complete the pixel value correction. in, The pixel value after correction for the pixel contamination point (x, y). Let i be the pixel value of the i-th uncontaminated pixel found in the search. The distance between the uncontaminated pixel and the contaminated pixel is the reciprocal of the spatial distance between them, and n is the total number of uncontaminated pixels found.
3. The AC surge arrester state detection method based on image semantic segmentation according to claim 1, characterized in that: In step S4, the image semantic segmentation network is the DeepLabV3+ network; The encoder of the DeepLabV3+ network uses the Xception-65 backbone network to extract image features and captures multi-scale contextual information through the dilated spatial pyramid pooling module, which contains parallel dilated convolutional layers with dilation rates of 6, 12 and 18. The DeepLabV3+ network's decoder sequentially upsamples the low-resolution feature map output by the encoder by 4x, concatenates it with the feature map from the encoder's intermediate layer, and upsamples it by 4x again to restore the resolution of the original image, so as to output pixel-level classification results with clear edges at material boundary regions.
4. The AC surge arrester state detection method based on image semantic segmentation according to claim 1, characterized in that: The training steps for an image semantic segmentation network include: Step S41: Obtain multiple sample images containing AC surge arresters. The sample images are labeled with pixel-level semantic tags corresponding to the equalizing ring, flange, porcelain bushing skirt and bolts, and the edge pixels of the material interface area are densely outlined. Step S42: Perform the same material boundary region detection and specular pollution suppression preprocessing as in steps S2 and S3 on the sample images to generate a training sample set; Step S43: Train the image semantic segmentation network using the training sample set. Use a weighted combination of cross-entropy loss function and boundary loss function as the optimization objective. Use the Adam optimizer, set the initial learning rate to 0.001, and the batch size to 8. Determine the network convergence when the loss function value tends to stabilize.
5. The AC surge arrester state detection method based on image semantic segmentation according to claim 1, characterized in that: In step S5, the assembly status detection of the AC surge arrester includes: extracting the outlines of the metal components and the insulation components according to the semantic segmentation results, calculating the minimum Euclidean distance between the outlines of the metal components and the outlines of the insulation components at the preset assembly position as the actual assembly gap, comparing the actual assembly gap with the preset standard assembly gap threshold, and if the actual assembly gap exceeds the allowable deviation range of the standard assembly gap threshold, the AC surge arrester is determined to be unqualified in assembly.
6. The AC surge arrester state detection method based on image semantic segmentation according to claim 5, characterized in that: Metal components include equalizing rings and flanges, while insulating components include porcelain bushings and skirts. In step S5, the actual assembly gap includes the first assembly gap between the equalizing ring and the porcelain bushing skirt, and the second assembly gap between the flange and the porcelain bushing skirt. The first assembly gap and the second assembly gap are compared with the corresponding standard assembly gap thresholds. If either assembly gap exceeds the allowable deviation range, the assembly is deemed unqualified.
7. The AC surge arrester state detection method based on image semantic segmentation according to claim 1, characterized in that: Step S5, the inspection of surface defects in the AC surge arrester includes: Pixel features of the insulating component region are extracted based on the semantic segmentation results. The pixel features are the gray values of each pixel in the insulating component region. Calculate the local mean and local standard deviation of pixel grayscale values within the insulating component area. Mark continuous pixel areas with grayscale values lower than the local mean minus 3 times the local standard deviation as abnormal dark areas. When the continuous area of abnormal dark areas is greater than a preset area threshold, the area is determined to be a crack defect.
8. The AC surge arrester state detection method based on image semantic segmentation according to claim 1, characterized in that: In step S1, the standard industrial high-intensity lighting is an LED array light source with an illuminance of 1000 to 2000 lux and a color temperature of 5000 to 6500K; The illuminance range and color temperature range are configured to simultaneously satisfy: generating specular highlights on the surface of the metal component sufficient to trigger the specular contamination determination in step S3, and maintaining the visibility of the umbrella skirt texture details on the surface of the insulating component.
Citation Information
Patent Citations
Building lightning arrester quality detection and evaluation method based on artificial intelligence
CN116612118A
Self-adaptive mirror surface highlight detection method based on polarization-like characteristics
CN118898617A
Image highlight removal method and system, storage medium and program product
CN121213406A
