Industrial visual inspection method based on edge light scanning

Abstract

The invention discloses an industrial visual inspection method based on edge light scanning, which comprises the following steps of: initializing projector parameter setting, adjusting shooting frequency and exposure time of an industrial camera, acquiring defect images and forming an image sequence according to the size and the material of a detected piece and an expected defect type to be detected and based on an edge light scanning method; carrying out fusion operation on the image sequence, enhancing defect signals of the images through a weighted stacking algorithm by utilizing high-contrast characteristics of defect areas in multiple frames of images, and carrying out defect positioning and classification on the enhanced images by adopting an image segmentation technology, so as to realize visual detection on defects of the detected piece; by optimizing the design of an optical imaging system, edge light scanning of the quartz wafer is realized, defects are highlighted, surface defects can be quickly and accurately identified, the detection efficiency is greatly improved, defect characteristics are directly enhanced in an image acquisition stage, and quick and accurate defect detection is realized in combination with a sequence image analysis technology.

Description

Technical Field

[0001] This invention relates to the field of precision optical inspection technology, and in particular to an industrial vision inspection method based on edge light scanning, which is especially suitable for automated inspection of surface defects of high-precision quartz wafers in semiconductor and optical device manufacturing. Background Technology

[0002] In the production process of quartz wafers, surface defect detection is crucial for product quality control. Current vision inspection systems often face the following problems:

[0003] 1. Low-contrast defect imaging: Under traditional uniform light sources, the difference between tiny defects and the reflected light from the background is small, resulting in insufficient image contrast and making them difficult to identify.

[0004] 2. Complex background interference: stray light from the environment or the reflective properties of the material itself can easily mask defect features.

[0005] 3. Conflict between detection efficiency and accuracy: High-resolution imaging requires long scanning time, but is inefficient; while rapid scanning is prone to missing tiny defects.

[0006] Existing solutions largely rely on post-processing image algorithms to enhance contrast, but these algorithms are complex and computationally intensive, making them difficult to apply in real time. Therefore, a technical solution is needed to improve defect contrast from the optical imaging stage, in order to reduce the difficulty of subsequent processing and achieve efficient and high-precision automated detection. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide an industrial visual inspection method based on edge light scanning, which aims to solve the problems of low imaging contrast and insufficient detection efficiency of quartz wafer surface defects in the prior art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] Firstly, an industrial visual inspection method based on edge light scanning includes:

[0010] Step 1: Check the hardware connection. The hardware includes an adjustable projector, an industrial camera, and a multi-degree-of-freedom stage. Based on the size and material of the part under test and the type of defect to be detected, initialize the projector parameter settings based on the edge light scanning method.

[0011] Step 2: According to the projector's parameter settings, adjust the industrial camera's shooting frequency and exposure time to acquire defect images and form an image sequence;

[0012] Step 3: Perform a fusion operation on the image sequence obtained in Step 2. Utilize the high contrast features of the defect region in the multi-frame images and enhance the defect signal of the image through a weighted superposition algorithm.

[0013] Step 4: For the image enhanced in Step 3, image segmentation technology is used to locate and classify defects in the image, thereby realizing visual detection of defects in the tested part.

[0014] Furthermore, in step 1, initializing the projector parameter settings based on the edge light scanning method includes:

[0015] Based on the size and material of the quartz crystal and the type of defects to be detected, the initial light source intensity parameter of the projector is set to a range of 50-200 mW / cm². 2 The grating modulation frequency parameter ranges from 100 to 500 Hz, the stripe width parameter ranges from 10 to 30 μm, and the spacing parameter ranges from 20 to 50 μm.

[0016] Furthermore, in step 1, initializing the projector parameter settings based on the edge light scanning method also includes:

[0017] Based on the expected defect size type to be detected in the test piece, the frequency and amplitude of the projector grating are periodically modulated. The modulation range of the grating frequency f is set to 100-500Hz, and the range of the grating amplitude A is set to 5-30μm or 30°-90°.

[0018] Furthermore, during the modulation process of the grating, the tiny scribe lines on the grating will produce a diffraction effect on the incident light. The diffracted light in different directions will form a striped pattern of alternating light and dark. Therefore, it is necessary to precisely control the parameters of the grating and generate a static striped light pattern with a specific spatial periodicity based on the scribe line spacing, scribe line shape and modulation depth. In order to realize the dynamic scanning of the striped light on the wafer surface, the mechanical movement of the stage or grating is controlled in a coordinated manner so that the striped light moves periodically along a preset path, exciting the temporal response characteristics of the defects. In controlling the mechanical movement of the stage or grating, the scanning speed v is 1-10 mm / s and the directional accuracy is ±0.1°.

[0019] Furthermore, in step 1, initializing the projector parameter settings based on the edge light scanning method includes:

[0020] By utilizing the rotation and tilt functions of the multi-degree-of-freedom stage, the placement angle of the quartz wafer is precisely adjusted so that the wafer surface forms a specific incident angle with the striped light projected by the projector. The projector is then moved laterally or longitudinally along a preset path at a scanning speed of 10–20 mm / s, where the incident angle is 30°–60°. This incident angle is optimized based on the optical characteristics of the wafer and the expected direction of defects.

[0021] Furthermore, in step 2, the industrial camera and projector are controlled synchronously. While the projector projects striped light and scans the surface of the quartz wafer, the industrial camera continuously acquires image sequences according to the preset frame rate and exposure time.

[0022] Furthermore, in step 3, the acquired image sequence is converted to grayscale, and the color image is converted to a grayscale image. The grayscale image is then normalized, mapping the pixel values ​​to the [0,1] interval. Since the quartz crystal may have slight displacement or deformation during the projector scanning process, the image sequence needs to be aligned. The first image is selected as the reference image, and the feature points in each image are extracted using the SIFT feature matching algorithm. Based on the matched feature point pairs, the transformation matrix is ​​obtained, and the subsequently acquired images are transformed according to the transformation matrix so that all images are in the same coordinate system, achieving accurate registration.

[0023] Furthermore, in step 3, enhancing the image defect signal using a weighted superposition algorithm includes:

[0024] First, determine the weighting coefficients and analyze the contrast of defect features in each image of the image sequence. Images with obvious defect features are assigned higher weights, while images with blurred defect features are assigned lower weights. The weighting coefficients are calculated as follows:

[0025]

[0026] Where C i The contrast of the defect features in the i-th image is represented by n, where n is the total number of images in the image sequence.

[0027] The aligned image sequences are then weighted and superimposed according to determined weight coefficients to obtain the fused image, as shown in the following formula:

[0028]

[0029] Among them I i (x,y) represents the gray value of the i-th image at pixel (x,y), w i These are the corresponding weighting coefficients.

[0030] Furthermore, step 3 includes performing histogram equalization contrast enhancement processing on the fused image to further improve the contrast between the defect signal and the background, making the defects more obvious. Histogram equalization makes the image histogram more uniform by redistributing the gray values ​​of image pixels, thereby enhancing the contrast of the image. Then, Gaussian filtering is used for noise filtering to remove noise that may be introduced during the fusion process.

[0031] Furthermore, in step 4, the defect localization and classification of the image using image segmentation technology includes:

[0032] 4.1 Threshold Segmentation: Adaptive threshold segmentation is performed on the defect region. First, Otsu's method is used to traverse all possible thresholds to find the threshold that maximizes the inter-class variance between the target region and the background region, thus achieving automatic threshold determination. Then, the image is segmented into foreground and background based on the determined threshold. Specifically, pixels with gray values ​​greater than the threshold are classified as foreground, and the remaining pixels are classified as background.

[0033] 4.2 Post-processing of segmentation results: For some regions with small areas due to noise or missegmentation, they are removed by setting an area threshold;

[0034] 4.3 Defect Category Determination: The processed image is trained using a convolutional neural network to automatically extract features from the image. The parameters of the classifier are adjusted through an optimization algorithm to enable it to correctly classify the training samples to the greatest extent possible. Then, the classifier determines the defect category according to its internal decision rules and outputs the determination result.

[0035] The convolutional neural network uses the ResNet-34 architecture and is fine-tuned for 50 epochs on a dataset containing 3000 labeled defects. The optimizer uses the Adam algorithm with an initial learning rate of 0.001, which decays by a factor of 0.95 every 10 epochs. The classifier obtains the defect category probability through the Softmax function.

[0036]

[0037] Where P(y=c|x) is the probability that the input feature x belongs to category c, and z c Let P be the raw score of category c output by the fully connected layer, where C is the total number of defect categories, and P be the probability vector output by the Softmax layer, where P = [p1, p2, ..., p2]. c ], where p c =P(y=c|x);

[0038] The decision rule is to select the category with the highest probability.

[0039]

[0040] when When determined as a category Otherwise, mark it as an unknown defect.

[0041] Secondly, the present invention also provides an industrial vision inspection system based on edge light scanning, including an adjustable projector, a high-resolution industrial camera, a multi-degree-of-freedom stage, and a computer control module. The adjustable projector and the high-resolution industrial camera are both connected to the multi-degree-of-freedom stage, which is controlled by the computer control module. The projector uses a narrow-band LED light source and generates dynamic stripe light through a grating modulator, which is projected onto the surface of a quartz crystal at a specific incident angle. The projector and the industrial camera are synchronously controlled by the computer to achieve timing matching between light source scanning and image acquisition.

[0042] The technical solution adopted in this invention has the following beneficial effects:

[0043] 1. In this invention, by optimizing the design of the optical imaging system, edge light scanning of quartz wafers is achieved, directly enhancing defect features during the image acquisition stage. Combined with sequential image analysis technology, rapid and accurate defect detection is realized.

[0044] 2. Improved detection efficiency: By highlighting defects, surface defects can be identified quickly and accurately, greatly improving detection efficiency.

[0045] 3. Improve product quality: Accurate detection of surface defects allows for timely identification and resolution of problems, thereby improving the overall quality of the product.

[0046] 4. Reduced maintenance costs: Since this method can accurately detect surface defects, it can reduce unnecessary maintenance and repair work, thereby reducing maintenance costs. Attached Figure Description

[0047] Figure 1 A flowchart of an industrial vision inspection method based on edge light scanning provided by the present invention;

[0048] Figure 2 An image acquisition design diagram for an industrial vision inspection system based on edge light scanning provided by the present invention;

[0049] Figure 3 This is a defect classification result image obtained by an industrial vision inspection method based on edge light scanning in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] This invention provides an industrial vision inspection method based on edge light scanning, the inspection process of which is as follows: Figure 1 As shown, the details are as follows:

[0052] Step 1: Check the hardware connections. The hardware includes an adjustable projector, an industrial camera, and a multi-DOF stage. Based on the size and material of the workpiece under test and the expected type of defect to be detected, initialize the projector parameter settings using the edge light scanning method.

[0053] 1. Hardware system setup:

[0054] (1) Optical imaging module: including adjustable projector, high-resolution industrial camera, multi-degree-of-freedom stage and computer control system.

[0055] (2) Light source design: The projector uses a narrow-band LED light source, which generates dynamic stripe light through grating modulation and projects it onto the surface of the quartz crystal at a specific incident angle.

[0056] (3) Collaborative control: The projector and industrial camera are controlled synchronously by a computer to achieve timing matching of light source scanning and image acquisition.

[0057] 2. Edge light scanning method:

[0058] (1) Hardware preparation and parameter setting:

[0059] 1) Connect the adjustable projector to the computer control system and ensure normal communication. The projector uses a narrow-band LED light source, which has high luminous efficiency and a narrow spectral width, effectively reducing ambient light interference and improving image contrast.

[0060] 2) Based on the size and material of the quartz crystal and the type of defect to be detected, the initial light source intensity parameter of the projector is set to a range of 50-200 mW / cm² in the computer control system. 2 The selection criteria are shown in Table 1; the grating modulation frequency parameter range is 100-500Hz, and the selection criteria are shown in Table 2; the stripe width parameter range is 10-30μm, and the spacing parameter range is 20-50μm, and the selection criteria are shown in Table 3.

[0061] Table 1. Basis for Selecting Initial Light Source Intensity

[0062]

[0063] Table 2. Basis for selecting grating modulation frequency

[0064]

[0065]

[0066] Table 3. Basis for Selecting Stripe Width and Spacing

[0067]

[0068] In the actual test, the initial value of the light source intensity was 100mW / cm². 2 The grating modulation frequency is 200Hz, the stripe width is 20±5μm, the spacing is 30±5μm, and it is dynamically adjusted according to the wafer thickness, surface roughness and target defect size.

[0069] (2) Grating modulation and stripe generation:

[0070] 1) The projector has a built-in grating modulator. Upon receiving a control signal from the computer, the grating begins periodic modulation at a set frequency and amplitude. The function of grating modulation is to convert the continuous light emitted by the narrow-band LED light source into striped light with a specific pattern. The grating modulator receives the computer control signal and performs periodic modulation at a set frequency f (100-500Hz) and amplitude A (5-30μm or 30°-90°). The value of amplitude A is optimized according to the target defect size: micron-level linear defects: A = 5-10μm; deep volume defects: A = 15-25μm; highly reflective areas: dynamically adjusted A ±40%.

[0071] 2) The modulation process of the grating is achieved by changing the diffraction angle and intensity distribution of the grating. Specifically, the tiny scribe lines on the grating produce a diffraction effect on the incident light, and the diffracted light in different directions forms alternating bright and dark stripe patterns. By precisely controlling the parameters of the grating, such as the scribe line spacing, scribe line shape, and modulation depth, a static stripe light pattern with specific spatial periodicity is generated. To achieve dynamic scanning of the stripe light on the wafer surface, the mechanical movement of the stage or grating needs to be controlled in a coordinated manner to make the stripe light move periodically along a preset path (scanning speed v = 1-10 mm / s, directional accuracy ±0.1°), thereby exciting the temporal response characteristics of defects.

[0072] (3) Incident angle adjustment and projection control:

[0073] 1) Utilizing the rotation and tilting functions of the multi-degree-of-freedom stage, the placement angle of the quartz wafer is precisely adjusted so that the wafer surface forms a specific incident angle with the striped light projected by the projector. This incident angle is optimized between 30° and 60° based on the optical characteristics of the wafer and the expected direction of the defects, to ensure that the striped light can produce a significant difference in scattering and reflection in the defect area.

[0074] 2) The projector moves along a preset path (horizontal / vertical) at a scanning speed of 10–20 mm / s. Simultaneously, the computer system acquires the projector's pose data in real time through a high-precision encoder (such as a magnetic ruler) and a MEMS tilt sensor, and adjusts the light source intensity and grating modulation parameters in a closed loop accordingly. This dynamic control ensures that the spatial position deviation of the stripe light is less than 3 μm and the illuminance uniformity is better than 95%, providing a stable input for defect feature extraction.

[0075] (4) Dynamic scanning and defect induction:

[0076] 1) When striped light shines on the surface of a quartz wafer, defect-free areas will process the striped light according to the laws of light reflection and refraction, exhibiting a relatively regular light intensity distribution. However, areas with defects, such as cracks, pits, or scratches, will produce different optical effects on the striped light, such as scattering, diffraction, or absorption, due to the discontinuity of their surface microstructure and changes in their optical properties.

[0077] 2) These optical effects cause significant changes in local light intensity in the defect area, resulting in a characteristic region with a clear contrast to the surrounding background in the striped light image. As the projector continues to scan, striped light from different angles and directions will sequentially illuminate the wafer surface, causing the defect to appear with different contrasts and shapes in multiple image frames.

[0078] Step 2: Based on the projector's parameter settings, adjust the industrial camera's shooting frequency and exposure time to acquire defect images and form an image sequence:

[0079] Image acquisition and synchronization control:

[0080] 1) The high-resolution industrial camera and projector are synchronously controlled by a computer control system. While the projector projects striped light and scans the surface of the quartz wafer, the industrial camera continuously acquires image sequences according to a preset frame rate and exposure time.

[0081] 2) The computer system precisely controls the timing of the projector and camera, ensuring that the projection of each stripe of light is accurately matched with the camera's exposure operation, avoiding blurring or missing stripes during image acquisition. The acquired image sequence will be transmitted to the computer in real time for use by subsequent image processing and analysis modules.

[0082] Step 3: Perform a fusion operation on the image sequence obtained in Step 2. Utilize the high contrast features of the defect region in multiple frames of images to enhance the defect signal of the image through a weighted superposition algorithm.

[0083] Defect detection process:

[0084] (1) Image sequence fusion: The defect signal is enhanced by weighted superposition algorithm by utilizing the high contrast features of the defect area in multiple frames of images.

[0085] 1) Image preprocessing: The acquired image sequence is converted to grayscale to reduce the amount of data and computational complexity; the grayscale image is normalized to map the pixel values ​​to the [0,1] interval for subsequent algorithm processing.

[0086] 2) Image Alignment and Registration: Since the quartz crystal may have slight displacement or deformation during the projector scanning process, it is necessary to align the image sequence. Here, the first image is selected as the reference image, and the feature points in each image are extracted using the SIFT feature matching algorithm. Based on the matched feature point pairs, the transformation matrix is ​​calculated, and the subsequently acquired images are transformed according to the transformation matrix so that all images are in the same coordinate system, achieving accurate registration.

[0087] 3) Weighted superposition algorithm to enhance defect signals: First, determine the weight coefficients, analyze the contrast of defect features in each image of the image sequence, assign higher weights to images with obvious defect features, and lower weights to images with blurred defect features. The weight calculation coefficients are shown in Formula 1, where C i The image represents the contrast of the defect features in the i-th image, and n is the total number of images in the image sequence.

[0088]

[0089] Then, the aligned image sequences are weighted and superimposed according to determined weight coefficients to obtain the fused image, as shown in Formula 2, where I i (x,y) represents the gray value of the i-th image at pixel (x,y), w i These are the corresponding weighting coefficients.

[0090]

[0091] 4) Post-processing of the fused image: Histogram equalization and contrast enhancement are performed on the fused image to further improve the contrast between the defect signal and the background, making the defects more obvious. Histogram equalization redistributes the gray values ​​of image pixels, making the image histogram more uniform, thereby enhancing the image contrast. Then, Gaussian filtering is used for noise filtering to remove noise that may have been introduced during the fusion process. Gaussian filtering uses a Gaussian kernel function to perform a weighted average on the image, which can effectively remove Gaussian noise while maintaining the relative integrity of the image's edge information.

[0092] Step 4: For the image enhanced in Step 3, image segmentation technology is used to locate and classify defects, thereby achieving visual inspection of defects in the tested part.

[0093] 1) Threshold segmentation: Adaptive threshold segmentation is performed on the defect region. First, the Otsu method is used to traverse all possible thresholds to find the threshold that maximizes the inter-class variance between the target region and the background region, thereby achieving automatic threshold determination. Then, the image is segmented into two parts, foreground (defect region) and background (no defect region), based on the determined threshold. Specifically, pixels with gray values ​​greater than the threshold are classified as foreground, and the remaining pixels are classified as background.

[0094] 2) Post-processing of segmentation results: For some regions with small areas due to noise or missegmentation, they can be removed by setting an area threshold to reduce interference in subsequent processing.

[0095] 3) Defect Category Determination: A convolutional neural network is used to train the processed images, automatically extracting features from the images. The classifier parameters are adjusted through an optimization algorithm to maximize the accuracy of classifying training samples. The classifier then determines the defect category based on its internal decision rules and outputs the result. The convolutional neural network uses the ResNet-34 architecture and is fine-tuned for 50 epochs on a dataset containing 3000 labeled defects. The optimizer uses the Adam algorithm with an initial learning rate of 0.001, decaying by a factor of 0.95 every 10 epochs. The classifier calculates the defect category probability using the Softmax function (Equation 3), where P(y=c|x) is the probability that the input feature x belongs to category c, and z... c Let P be the raw score of category c output by the fully connected layer, where C is the total number of defect categories, and P be the probability vector output by the Softmax layer, where P = [p1, p2, ..., p2]. c (where p) c =P(y=c|x)); The decision rule is to take the category with the highest probability (Formula 4), when When determined as a category Otherwise, mark it as an unknown defect.

[0096]

[0097] Results and Applications:

[0098] After the image processing and analysis are completed, the system can generate a report that details all detected defects and their locations. This information can be used for quality control, product classification, or other related applications.

[0099] In general, the visual inspection method of the present invention has the following beneficial effects:

[0100] (1) The projector moves and scans periodically.

[0101] A projector is used to illuminate the surface of the quartz wafer. By moving the scan periodically, it is ensured that the entire surface of the quartz wafer is uniformly illuminated, which is crucial for subsequent image acquisition and analysis.

[0102] (2) Taking pictures with an industrial camera

[0103] As the projector illuminates the sample, the industrial camera begins to capture images. Due to the movement of the projector, the camera is able to photograph the entire surface of the quartz wafer, thus obtaining a sequence of images.

[0104] (3) Detecting surface defects

[0105] If a quartz wafer has surface defects, such as cracks, stains, or other irregular shapes, then the way light is reflected or scattered will differ in these defective areas from that in defect-free areas. This difference will create high-contrast areas in at least one image, making the defects obvious and easy to detect.

[0106] (4) Image sequence analysis

[0107] The resulting image sequences are then fed into a computer for analysis. Image processing techniques, such as edge detection, thresholding, or more sophisticated machine learning methods, can automatically identify and locate high-contrast regions in the images, thereby determining the location and type of defects.

[0108] This invention also provides an industrial vision inspection system based on edge light scanning. The image acquisition design diagram of the system described in this invention is shown below. Figure 2 As shown, it includes an adjustable projector, a high-resolution industrial camera, a multi-degree-of-freedom stage, and a computer control module. Both the adjustable projector and the high-resolution industrial camera are connected to the multi-degree-of-freedom stage, which is controlled by the computer control module. The projector uses a narrow-band LED light source and generates dynamic stripe light through a grating modulator, which is projected onto the surface of a quartz crystal at a specific incident angle. The projector and the industrial camera are synchronously controlled by the computer to achieve timing matching between light source scanning and image acquisition.

[0109] In this process, a projector periodically scans the entire surface of a quartz wafer sample, while an industrial camera takes pictures, resulting in an image sequence of the inspected surface. If defects are present, at least one image in the sequence must contain a high-contrast image. Finally, the images are sent to a computer for analysis, and the defect classification results are shown in the figure below. Figure 3 As shown.

Claims

1. An industrial visual inspection method based on edge light scanning, characterized in that, include: Step 1: Check the hardware connection. The hardware includes an adjustable projector, an industrial camera, and a multi-degree-of-freedom stage. Based on the size and material of the part under test and the type of defect to be detected, initialize the projector parameter settings based on the edge light scanning method. Step 2: According to the projector's parameter settings, adjust the industrial camera's shooting frequency and exposure time to acquire defect images and form an image sequence; Step 3: Perform a fusion operation on the image sequence obtained in Step 2. Utilize the high contrast features of the defect region in the multi-frame images and enhance the defect signal of the image through a weighted superposition algorithm. Step 4: For the image enhanced in Step 3, image segmentation technology is used to locate and classify defects in the image, thereby achieving visual detection of defects in the tested part.

2. The industrial visual inspection method based on edge light scanning according to claim 1, characterized in that, Step 1, which initializes the projector parameter settings based on the edge light scanning method, includes: Based on the size and material of the quartz crystal and the type of defects to be detected, the initial light source intensity parameter of the projector is set to a range of 50-200 mW / cm². 2 The grating modulation frequency parameter ranges from 100 to 500 Hz, the stripe width parameter ranges from 10 to 30 μm, and the spacing parameter ranges from 20 to 50 μm.

3. The industrial visual inspection method based on edge light scanning according to claim 1, characterized in that, Step 1, which initializes the projector parameter settings based on the edge light scanning method, also includes: Based on the expected defect size type to be detected in the test piece, the frequency and amplitude of the projector grating are periodically modulated. The modulation range of the grating frequency f is set to 100-500Hz, and the range of the grating amplitude A is set to 5-30μm or 30°-90°.

4. The industrial visual inspection method based on edge light scanning according to claim 3, characterized in that, During the modulation process of the grating, the tiny grating lines will produce a diffraction effect on the incident light. The diffracted light in different directions will form a striped pattern of alternating light and dark. Therefore, it is necessary to precisely control the parameters of the grating and generate a static striped light pattern with specific spatial periodicity based on the grating spacing, grating shape and modulation depth. To achieve dynamic scanning of stripe light on the wafer surface, the mechanical movement of the stage or grating is controlled in a coordinated manner to make the stripe light move periodically along a preset path, thereby exciting the temporal response characteristics of defects. In controlling the mechanical movement of the stage or grating, the scanning speed v is 1-10 mm / s, and the directional accuracy is ±0.1°.

5. The industrial visual inspection method based on edge light scanning according to claim 1, characterized in that, Step 1, which initializes the projector parameter settings based on the edge light scanning method, includes: By utilizing the rotation and tilt functions of the multi-degree-of-freedom stage, the placement angle of the quartz wafer is precisely adjusted so that the wafer surface forms a specific incident angle with the striped light projected by the projector. The projector is then moved laterally or longitudinally along a preset path at a scanning speed of 10–20 mm / s, where the incident angle is 30°–60°. This incident angle is optimized based on the optical characteristics of the wafer and the expected direction of defects.

6. The industrial visual inspection method based on edge light scanning according to claim 1, characterized in that, In step 2, the industrial camera and projector are controlled synchronously. While the projector projects striped light and scans the surface of the quartz wafer, the industrial camera continuously acquires image sequences according to the preset frame rate and exposure time.

7. The industrial visual inspection method based on edge light scanning according to claim 1, characterized in that, In step 3, the acquired image sequence is converted to grayscale, and the color image is converted to a grayscale image. The grayscale image is then normalized, and the pixel values ​​are mapped to the [0,1] interval. Since the quartz crystal may have slight displacement or deformation during the projector scanning process, the image sequence needs to be aligned. The first image is selected as the reference image, and the feature points in each image are extracted using the SIFT feature matching algorithm. Based on the matched feature point pairs, the transformation matrix is ​​obtained, and the subsequently acquired images are transformed according to the transformation matrix so that all images are in the same coordinate system, achieving accurate registration.

8. The industrial visual inspection method based on edge light scanning according to claim 7, characterized in that, In step 3, enhancing the defect signal of the image using a weighted superposition algorithm includes: First, determine the weighting coefficients and analyze the contrast of defect features in each image of the image sequence. Images with obvious defect features are assigned higher weights, while images with blurred defect features are assigned lower weights. The weighting coefficients are calculated as follows: Where C i The contrast of the defect features in the i-th image is represented, and n is the total number of images in the image sequence. The aligned image sequences are then weighted and superimposed according to determined weight coefficients to obtain the fused image, as shown in the following formula: Among them I i (x,y) represents the gray value of the i-th image at pixel (x,y), w i These are the corresponding weighting coefficients.

9. The industrial visual inspection method based on edge light scanning according to claim 8, characterized in that, Step 3 includes performing histogram equalization contrast enhancement processing on the fused image to further improve the contrast between the defect signal and the background, making the defects more obvious. Histogram equalization makes the image histogram more uniform by redistributing the gray values ​​of image pixels, thereby enhancing the contrast of the image. Then, Gaussian filtering is used to perform noise filtering processing to remove noise that may be introduced during the fusion process.

10. The industrial visual inspection method based on edge light scanning according to claim 1, characterized in that, In step 4, image segmentation technology is used to locate and classify defects in the image, including: 4.1 Threshold Segmentation: Adaptive threshold segmentation is performed on the defect region. First, Otsu's method is used to traverse all possible thresholds to find the threshold that maximizes the inter-class variance between the target region and the background region, thus achieving automatic threshold determination. Then, the image is segmented into foreground and background based on the determined threshold. Specifically, pixels with gray values ​​greater than the threshold are classified as foreground, and the remaining pixels are classified as background. 4.2 Post-processing of segmentation results: For some regions with small areas due to noise or missegmentation, they are removed by setting an area threshold; 4.3 Defect Category Determination: The processed image is trained using a convolutional neural network to automatically extract features from the image. The parameters of the classifier are adjusted through an optimization algorithm to enable it to correctly classify the training samples to the greatest extent possible. Then, the classifier determines the defect category according to its internal decision rules and outputs the determination result. The convolutional neural network uses the ResNet-34 architecture and is fine-tuned for 50 epochs on a dataset containing 3000 labeled defects. The optimizer uses the Adam algorithm with an initial learning rate of 0.001, which decays by a factor of 0.95 every 10 epochs. The classifier obtains the defect category probability through the Softmax function. Where P(y=c|x) is the probability that the input feature x belongs to category c, and z c Let P be the raw score of category c output by the fully connected layer, where C is the total number of defect categories, and P be the probability vector output by the Softmax layer, where P = [p1, p2, ..., p2]. c ], where p c =P(y=c|x); The decision rule is to select the category with the highest probability. when When determined as a category Otherwise, mark it as an unknown defect.

Images