An AI vision-based special glass surface micro-crack detection system

Abstract

The application discloses a kind of based on AI vision's special glass surface micro crack detection system, specifically related to optical imaging and computer vision field, including optical image acquisition module, initial detection module, defect feature extraction module, defect prediction module, secondary detection module, final determination module, database updating module.The application is combined by using multimodal optical imaging and AI dynamic learning, carries out online detection on special glass production line, and introduces timing rate of change analysis when judging defect based on multi-angle optical feature extraction result, to realize the accurate determination of defect type, and automatically screen hard sample and push expert review, can guarantee the product quality safety to the greatest extent, while better guide production line maintenance and process optimization, reduce potential safety hazard, in addition, efficiency bottleneck and cost rise caused by excessive dependence on artificial review can also be avoided, reduce long-term operation cost.

Description

Technical Field

[0001] This invention relates to the fields of optical imaging and computer vision technology, and more specifically, to a special glass surface microcrack detection system based on AI vision. Background Technology

[0002] During the manufacturing, processing and transportation of specialty glass, its surface is prone to micron-level cracks due to mechanical stress, thermal stress or external impact. Although these microcracks are difficult to detect with the naked eye, they can significantly weaken the mechanical strength and optical properties of the glass, thereby affecting product reliability and causing safety hazards, especially in high-end display, automotive and security applications where the consequences are particularly serious.

[0003] In industry, the detection of defects on glass surfaces still largely relies on manual visual inspection. This method is not only inefficient and labor-intensive, but also susceptible to factors such as the experience and fatigue of personnel, resulting in strong subjectivity and poor consistency. It is difficult to meet the requirements of modern high-speed production lines for detection accuracy and efficiency. In recent years, automatic detection systems based on machine vision have been gradually applied to the detection of surface defects in industry. These systems typically rely on high-resolution cameras and specific light source configurations to identify defects through image processing algorithms.

[0004] However, in practical use, it still has some drawbacks. First, the existing manual inspection of microcracks on special glass surfaces is inefficient, easily affected by subjective factors, and difficult to maintain a high level of focus, leading to an increased risk of missed detections and misjudgments. Machine vision systems based on traditional image processing mostly adopt static, single-angle imaging methods, rely on manually designed features, and have limited ability to distinguish microcracks from pseudo-defects with similar shapes such as scratches, dust, and water stains, resulting in high false detection and missed detection rates. Second, these systems lack the ability to dynamically monitor the changes in defects over time and stress, and cannot assess the crack propagation trend and actual hazards, limiting their practical application in high-end manufacturing quality control. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a special glass surface microcrack detection system based on AI vision. By integrating multimodal optical features and temporal variation features, it achieves extremely high accuracy in identifying microcracks and constructs a continuously evolving defect database with industry adaptability, effectively solving the problems raised in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The optical image acquisition module is used to acquire optical images when the special glass enters the initial inspection area with the conveyor belt, and to generate an image sequence through the imaging module.

[0008] The initial detection module is used to perform an initial detection on the generated image sequence to identify suspected defect areas.

[0009] The defect feature extraction module is used to extract the feature image sequence of the suspected defect area and extract the defect features, including shape features, defect edge angle features and defect refractive features.

[0010] The defect prediction module performs feature fusion based on defect features to obtain the defect confidence of suspected defect areas, thereby predicting the defect type of the suspected defect areas.

[0011] The secondary inspection module is used to perform a secondary inspection on the same special glass after it has passed through the initial inspection area, thereby performing time-series change rate analysis.

[0012] The final judgment module makes a final judgment on defects based on the defect morphology change rate and the defect confidence change rate. When the defect type in the suspected defect area is determined to be a microcrack, the unique ID is retrieved for source tracing.

[0013] The database update module is used to filter fuzzy defects as hard samples, push the hard samples and their data to the expert review interface for manual review, and perform incremental training on the AI ​​model based on the review results.

[0014] The technical effects and advantages of this invention are as follows:

[0015] 1. This invention combines multimodal optical imaging with AI dynamic learning to perform online inspection on special glass production lines. When judging defects based on multi-angle optical feature extraction results, it introduces time-series change rate analysis, thereby achieving accurate determination of defect types. It is not limited to static image recognition. On the one hand, it can maximize the ability to distinguish between microcracks and false defects. On the other hand, it can avoid frequent line stoppages and re-inspections caused by false detection and missed detection, which is conducive to improving the quality control level while ensuring production efficiency.

[0016] 2. This invention calculates the time-series change rate after performing a secondary inspection operation, and makes a final judgment based on the morphological change rate and confidence change rate when an extended crack is identified. It is not limited to a single judgment and can achieve early detection and dynamic risk assessment of high-risk cracks in some cases, thereby maximizing the protection of product quality and safety. At the same time, it can better guide production line maintenance and process optimization, reduce potential safety hazards, and avoid the efficiency bottleneck and cost increase caused by over-reliance on manual re-inspection, thus reducing long-term operation and maintenance costs.

[0017] 3. After completing the defect determination, this invention automatically filters hard samples through a closed-loop database update mechanism and pushes them to experts for review. It uses the results of manual annotation to incrementally train the AI ​​model, thereby achieving continuous evolution and adaptive optimization of system performance. This avoids model failure caused by process changes or new defects, and achieves the goals of improving detection robustness, reducing manual intervention, and enhancing the long-term availability of the system. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0019] Figure 2 This is a flowchart of the time-series change rate analysis and final determination process of the present invention. Detailed Implementation

[0020] 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.

[0021] As attached Figure 1-2 The system shown is a special glass surface microcrack detection system based on AI vision, including an optical image acquisition module, a primary detection module, a defect feature extraction module, a defect prediction module, a secondary detection module, a final judgment module, and a database update module.

[0022] The specific embodiments of the present invention include the following:

[0023] The optical image acquisition module is used to acquire optical images when the special glass enters the initial inspection area with the conveyor belt, and to generate an image sequence through the imaging module.

[0024] It should be explained that when the special glass enters the initial inspection area with the conveyor belt, the photoelectric encoder detects the edge of the glass and sends a trigger signal to the camera. The camera then starts to work at a preset frequency, where the preset frequency can be one set of images every 5mm of the conveyor belt moving.

[0025] It should be added that the imaging module is integrated above the production line conveyor belt or at the end of the robotic arm. Specifically, it includes a main illumination imaging unit, a multi-angle imaging unit, and a high-resolution camera. The main illumination imaging unit is a combination of bright field illumination and dark field illumination. Bright field illumination is used to capture the overall shape of defects and surface contamination, while dark field illumination is used to capture the angular scattered light of tiny scratches and cracks, highlighting their contours. The multi-angle imaging unit uses multiple independently controllable angle light sources arranged around the glass being tested to illuminate from different directions, dynamically capturing the angularity and refraction of defects. The high-resolution camera uses a high-speed and high-resolution CMOS camera to take pictures in the main illumination imaging unit and the multi-angle imaging unit, ensuring that micron-level defects can be resolved.

[0026] In this embodiment, it should be specifically explained that the image sequence generation specifically includes: acquiring optical images of the same piece of special glass, including bright field illumination images, dark field illumination images and multi-angle illumination images, and sorting the optical images in the order of acquisition of bright field illumination images, dark field illumination images and multi-angle illumination images, thereby generating a set of image sequences under different illumination modes and different angle illumination.

[0027] The initial detection module is used to perform initial detection on the generated image sequence to identify suspected defect areas.

[0028] In this embodiment, the initial detection process is described in detail as follows:

[0029] The generated image sequence is stitched and fused, specifically including: fusing the dark field image sequence at the pixel level to generate a synthetic dark field image that highlights the strongest scattered light from all angles; then performing image difference between this synthetic dark field image and the bright field image to generate an enhancement image as the main input for detection.

[0030] It should be added that the maximum value operation retains the brightest value of each pixel in all dark field images. This means that as long as a defect produces scattered light under any angle of illumination, its signal will be preserved and enhanced. This ensures that even very weak or highly directional defects will not be missed, improving the recall rate of detection. At the same time, since bright field images contain a lot of background information, image differencing can effectively suppress the interference of this irrelevant information, while greatly highlighting defects that are only noticeable in dark fields.

[0031] The generated enhanced image is preprocessed, including but not limited to standardization and size scaling, and then the preprocessed image is input into a trained convolutional neural network model to identify defects in the image, including but not limited to bubbles, microcracks, dust, water stains, and scratches.

[0032] It should be added that convolutional neural network models automatically learn and extract features at various levels from images through their internal multi-layer convolutional layers. Shallow networks can identify simple edges and corners, while deep networks can combine these simple features to identify image defects.

[0033] A two-dimensional coordinate system is constructed with the special glass inspection benchmark point as the origin and the special glass boundary as the coordinate axis. Defects are located by coordinates, suspected defect areas are identified, and different suspected defect areas on the same glass are numbered.

[0034] It should be added that when constructing a two-dimensional coordinate system, the lower left corner or center point of the special glass inspection reference point can be used as the origin P0, and another reference point P1 located on the long side of the glass can be identified. A vector between P0 and P1 can be drawn, and the direction of the vector can be used as the direction of the X-axis. Then, the Y-axis can be determined according to the right-hand rule. If the glass is slightly tilted when placed on the conveyor belt, this rotational tilt can be automatically compensated by identifying the coordinate system established by the two reference points, ensuring that the coordinate system is always aligned with the glass and improving the accuracy of identifying suspected defect areas.

[0035] The defect feature extraction module is used to extract the feature image sequence of the suspected defect area and extract the defect features, including shape features, defect edge angle features, and defect refractive features.

[0036] In this embodiment, it should be specifically explained that the feature image sequence is obtained by cropping corresponding image blocks according to the suspected defect area and stitching all image blocks together according to the channel dimension. The morphological features include edge curvature variance and edge gradient mean, the defect edge angle features include end sharpness and scattered light spot direction entropy, and the defect refractive features include defect brightness fluctuation value and refractive sensitivity.

[0037] It needs to be explained that morphological features can reflect the geometric contour of defects. The edge curvature variance is used to describe the discreteness of the edge curvature. By uniformly sampling continuous edges, such as taking 1 sampling point every 5 pixels, an edge point set is obtained. The coordinates of the front and rear edge points are uniformly taken. The curvature is calculated by fitting a circle through three points. The variance of the curvature of all edge points is calculated, thus obtaining the edge curvature variance. For example, the edge curvature variance of microcracks is large, while the edge curvature variance of scratches is small.

[0038] The mean edge gradient is obtained by calculating the gray-level change rate of pixels at the defect edge and averaging the corresponding gray-level change rates of all edge points. For example, the edge of a microcrack is jagged, with large gradient value fluctuations and a high mean, while the edge of a scratch has small gradient fluctuations and a low mean.

[0039] The defect edge angle feature describes the edge structure of the defect surface. End sharpness describes the degree of sharpness at the defect end. Ends are filtered by identifying the direction of the defect's main axis, thereby extracting a set of end edge points. From this set, two edge points are extracted to form a triangle with the end point, and the included angle of the triangle is calculated. The specific calculation formula is as follows:

[0040] ,

[0041] Where θ represents the included angle of the triangle. and Let θ represent the vectors of edge point a and edge point b with the end point, respectively. Then the end sharpness Ga is the reciprocal of the included angle of the triangle. The smaller θ is, the greater the sharpness. When the defect type is a bubble, the sharpness is 0.

[0042] The orientation entropy of scattered light spots is used to quantify the spatial orderliness of their directional distribution. This is achieved by extracting suspected defect regions corresponding to multi-angle illumination images from feature image sequences, and then extracting the centroid coordinates of the scattered light spots for each angle. The orientation angle α of each light spot relative to the defect center is then calculated. i And statistically analyze the probability distribution P(α) of the direction angle. i Then, the directional entropy of the scattered light spot is specifically expressed as:

[0043] ,

[0044] Where Gb represents the directional entropy of the scattered light spot, k represents the number of directions of multi-angle illumination, i=1,2,…,k. The more regular the edges and the more concentrated the directions, the lower the entropy value; the more irregular the edges and the more chaotic the directions, the higher the entropy value.

[0045] Defect refractive characteristics are quantitative indicators that describe the refraction and reflection characteristics of light in a suspected defect area. Among them, the defect brightness fluctuation value represents the degree of dispersion of gray values ​​in the suspected defect area. It is calculated by statistically analyzing the gray values ​​corresponding to the pixels in the suspected defect area and using the variance formula. When the defect type is a microcrack, the brightness fluctuation is large, and the defect brightness fluctuation value is even larger.

[0046] Refractive sensitivity reflects the difference in brightness response of defects under different refractive indices. By extracting the defect brightness at each angle from the suspected defect area corresponding to the multi-angle illumination image, the difference between the defect brightness at adjacent angles is calculated and compared with the average defect brightness at the two angles to obtain the brightness change rate at adjacent angles. Then, the average value of the brightness change rate is calculated to obtain the refractive sensitivity.

[0047] The defect prediction module performs feature fusion based on defect features to obtain the defect confidence of suspected defect areas, thereby predicting the defect type of the suspected defect areas.

[0048] In this embodiment, the specific implementation of feature fusion is as follows:

[0049] A multi-level fusion strategy is adopted, which assigns weights based on defect features and sums the weights of each feature to generate defect feature values, specifically as follows:

[0050] ,

[0051] Where W represents the defect feature value, F, G and S represent the feature factors corresponding to the shape feature, defect edge angle feature and defect refraction feature respectively, w1, w2 and w3 are the weight coefficients corresponding to different feature factors, w1>0, w2>0, w3>0, and w1+w2+w3=1.

[0052] It should be explained that the specific methods for obtaining the feature factor F corresponding to the shape feature include: extracting the edge curvature variance and the edge gradient mean of the shape feature, and then using the formula...

[0053] ,

[0054] Calculate the feature factor corresponding to the shape feature, where Cv and Cm represent the edge curvature variance and the edge gradient mean, respectively, and e represents the natural constant. The larger the edge curvature variance and the larger the edge gradient mean, the larger the feature factor corresponding to the shape feature, and the greater the confidence of the defect and microcrack.

[0055] The specific methods for obtaining the feature factor G corresponding to the defect edge angle feature include: extracting the end sharpness of the defect edge angle feature and the scattering spot direction entropy, and then using the formula...

[0056] ,

[0057] Calculate the characteristic factor corresponding to the defect edge angle feature. Ga and Gb represent the end sharpness and the direction entropy of the scattered light spot, respectively. The larger the end sharpness and the smaller the direction entropy of the scattered light spot, the larger the characteristic factor corresponding to the defect edge angle feature.

[0058] The specific methods for obtaining the feature factor S corresponding to the defect refractive characteristics include: extracting the defect brightness fluctuation value and refractive sensitivity of the defect refractive characteristics, and then using the formula...

[0059] ,

[0060] Calculate the characteristic factor corresponding to the refractive characteristics of the defect. Sa and Sb represent the defect brightness fluctuation value and refractive sensitivity, respectively. The larger the defect brightness fluctuation value and the larger the refractive sensitivity, the larger the characteristic factor corresponding to the defect refractive characteristics.

[0061] It should be added that before inputting the parameters of each feature corresponding to the above feature image sequence into the feature fusion formula, the system first performs data normalization preprocessing on all these parameters to eliminate their dimensional differences. Specifically, a Z-score-based standardization method is adopted: during the model training phase, the mean μ and standard deviation σ of each feature on the training set are calculated; during the model inference phase, these pre-calculated μ and σ are used to transform each feature value, that is: normalized feature = (original feature - μ) / σ. After this processing, all features are converted into values ​​with a mean of 0 and a standard deviation of 1, thus achieving comparability and ensuring that their calculations in the subsequent feature fusion formula have mathematical rationality and stability.

[0062] It should be further explained that the specific operation for predicting the defect type in suspected defect areas is as follows:

[0063] The defect confidence score is calculated using the Sigmoid function on the defect feature values ​​obtained after feature fusion, where the defect confidence score represents the probability that the defect category of the suspected defect area is a microcrack.

[0064] The defect confidence level is compared with the confidence threshold. If the defect confidence level is greater than or equal to the confidence threshold, the defect category of the suspected defect area is predicted to be a microcrack. If the defect confidence level is less than the confidence threshold, the defect category of the suspected defect area is predicted to be a pseudocrack. Pseudocracks include, but are not limited to, bubbles, dust, water stains, scratches, etc.

[0065] It needs to be explained that the Sigmoid function is a mathematical function that can standardize any real number to the interval (0, 1) to solve binary classification problems. Since its output value ranges between 0 and 1 and the curve is smooth, it can interpret any defect feature value as a defect probability. The mathematical formula is generally: f(x) = 1 / (1+e^x), where f(x) represents the output value, x represents the input value, and e represents the natural constant.

[0066] The secondary inspection module is used to perform a secondary inspection on the same special glass after it has passed through the initial inspection area, thereby performing time-series change rate analysis.

[0067] In this embodiment, it should be specifically explained that the secondary detection module includes:

[0068] After the feature glass passes through the initial detection area, the secondary detection is initiated. Specifically, it can be initiated when the glass leaves the edge of the glass by detecting the photoelectric encoder. Before initiating the secondary detection, a unique ID is generated for the suspected defect area, including the detection date, glass number, and suspected defect area number.

[0069] The same piece of glass is photographed again by the optical image acquisition module to generate a new sequence of defect images. This allows us to obtain secondary defect features, secondary defect feature values, and secondary defect confidence scores. The secondary defect features, secondary defect feature values, and secondary defect confidence scores are acquired and calculated using the same methods as the defect features, defect feature values, and defect confidence scores.

[0070] It should be further explained that the time-series change rate analysis includes the defect morphology change rate and the defect confidence change rate, and the specific implementation is as follows:

[0071] Extract the defect image sequences from two separate tests, and extract the defect parameters, including defect length, defect area, and number of defect branches. Calculate the defect morphology change rate O, specifically expressed as: O = [(secondary defect parameter - initial defect parameter) / initial defect parameter] × 100%;

[0072] Extract the defect confidence scores from the two tests, and then calculate the defect confidence score change rate H, which is expressed as: H = [(secondary confidence score - initial confidence score) / initial confidence score] × 100%.

[0073] It should be explained that the magnitude of the defect morphology change rate reflects the activity and severity of the crack. A rapidly propagating crack requires immediate treatment, while a stable or slowly propagating crack can have different treatment priorities. Under stress or time, the propagation behavior of a true crack has typical characteristics (such as rapid propagation and clear directionality), while spurious defects such as scratches and dust remain stable. The defect confidence change rate can reflect the correctness of the defect type matching. A defect with high initial confidence but which then remains stable or even decreases is likely to be a spurious defect. For example, a drop of water may look like a crack when it first appears, but as it begins to evaporate and deform, the model's confidence in it will decrease. Combined with its characteristic of no morphological propagation, it can be excluded. By introducing time and behavioral dimensions for change rate analysis, higher accuracy, lower false positive and false negative rates can be achieved, and the risk of defects can be dynamically assessed.

[0074] The final judgment module makes a final judgment on defects based on the defect morphology change rate and the defect confidence change rate. When the defect type in the suspected defect area is determined to be a microcrack, the unique ID is retrieved for tracing.

[0075] In this embodiment, it should be specifically explained that the final determination module operates as follows:

[0076] An expansion threshold and a growth threshold are set for the defect morphology change rate and the defect confidence change rate, respectively. For example, the expansion threshold is 10% and the growth threshold is 5%.

[0077] When the rate of change of defect morphology and the rate of change of defect confidence are both greater than the expansion threshold and the growth threshold, it indicates that the defect is expanding significantly, and the defect type of the suspected defect area is determined to be a microcrack.

[0078] When the rate of change of defect morphology is greater than the expansion threshold and the rate of change of defect confidence is less than the growth threshold, it indicates that the defect is expanding significantly, but the confidence is decreasing instead. This may be a morphologically specific expansion defect, so the defect type is still determined to be a microcrack.

[0079] When both the defect morphology change rate and the defect confidence change rate are less than the expansion threshold and the growth threshold, it indicates that the defect morphology has shrunk and the defect confidence has decreased. In this case, the defect type of the suspected defect area is determined to be a pseudo-crack.

[0080] The database update module is used to filter fuzzy defects as hard samples, push the hard samples and their data to the expert review interface for manual review, and perform incremental training on the AI ​​model based on the review results.

[0081] In this embodiment, the database update module is implemented as follows:

[0082] Fuzzy defects refer to defect cases where the defect confidence level is within a preset middle range. The system automatically filters fuzzy defects and marks them as hard samples. The hard sample data is pushed to the expert review interface for defect type determination. The hard sample data includes optical images, coordinates of suspected defect areas, defect features, and time-series change rate. The expert review interface is a human-computer interaction software platform used to submit fuzzy defects to human experts for final decision-making and to use the experts' judgment results to provide feedback and optimize the AI ​​model.

[0083] After statistical analysis, hard samples are used to incrementally train the AI ​​model, and the training results are stored in the defect database. This enables the model to continuously learn new defect patterns, adapt to changes in raw materials and processes on the production line, and achieve self-evolution of the model.

[0084] As a preferred embodiment, based on a large amount of experimental data, it was found that cases with a confidence level below 30% are mostly obvious noise or non-defect interference, while the model can be highly confident in cases with a confidence level above 95%. Setting the screening range of hard samples to 30% to 95% can effectively cover the vast majority of boundary cases where the model's judgment is ambiguous and prone to errors. While ensuring that the model captures a sufficient variety of hard samples to promote learning, it avoids invalid verification of extremely obvious cases.

[0085] It should be noted that this confidence interval is not fixed. Those skilled in the art can dynamically adjust and optimize the threshold of this interval based on the quality requirements of specific production lines, the allocation of expert review resources, and the defect characteristics of different glass models.

[0086] It should be added that the defect database not only stores images and labels, but is also a multimodal feature database, which associates the static multimodal feature vector, time-series change data, final identification label and retrieves the unique ID of each defect, providing rich data support for subsequent research and development and optimization.

[0087] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0088] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A special glass surface microcrack detection system based on AI vision, characterized in that, It includes an optical image acquisition module, an initial inspection module, a defect feature extraction module, a defect prediction module, a secondary inspection module, a final judgment module, and a database update module. The optical image acquisition module is used to acquire optical images when the special glass enters the initial inspection area with the conveyor belt, and to generate an image sequence through the imaging module. The initial detection module is used to perform an initial detection on the generated image sequence to identify suspected defect areas. The defect feature extraction module is used to extract the feature image sequence of the suspected defect area and extract the defect features, including shape features, defect edge angle features and defect refractive features. The defect prediction module performs feature fusion based on defect features to obtain the defect confidence of suspected defect areas, thereby predicting the defect type of the suspected defect areas. The secondary inspection module is used to perform a secondary inspection on the same special glass after it has passed through the initial inspection area, thereby performing time-series change rate analysis. The time-series change rate analysis includes defect morphology change rate analysis and defect confidence change rate analysis, which are implemented as follows: Extract the defect image sequences from two separate tests, and extract the defect parameters, including defect length, defect area, and number of defect branches. Calculate the defect morphology change rate O, specifically expressed as: O = [(secondary defect parameter - initial defect parameter) / initial defect parameter] × 100%; Extract the defect confidence scores from the two tests, and then calculate the defect confidence score change rate H, specifically expressed as: H = [(secondary confidence score - initial confidence score) / initial confidence score] × 100%; The final judgment module makes a final judgment on defects based on the defect morphology change rate and the defect confidence change rate. When the defect type in the suspected defect area is determined to be a microcrack, the unique ID is retrieved for source tracing. The database update module is used to filter fuzzy defects as hard samples, push the hard samples and their data to the expert review interface for manual review, and perform incremental training on the AI ​​model based on the review results.

2. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The image sequence generation step specifically includes: acquiring optical images of the same piece of special glass, including bright field illumination images, dark field illumination images, and multi-angle illumination images, and sorting the optical images according to the order in which the bright field illumination images, dark field illumination images, and multi-angle illumination images were acquired, thereby generating a set of image sequences under different illumination modes and different angle illumination.

3. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The initial detection process is as follows: The generated image sequence is stitched and fused, specifically including: fusing the dark field image sequence at the pixel level to generate a synthetic dark field image; then performing image difference between this synthetic dark field image and the bright field image to generate an enhancement map as input for detection; The generated enhanced image is preprocessed, and the preprocessed image is input into a trained convolutional neural network model to identify defects in the image; A two-dimensional coordinate system is constructed with the special glass inspection benchmark point as the origin and the special glass boundary as the coordinate axis. Defects are located by coordinates, suspected defect areas are identified, and different suspected defect areas on the same glass are numbered.

4. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The feature image sequence is obtained by cropping corresponding image blocks from the image sequence according to the suspected defect area and stitching all image blocks together according to the channel dimension. The shape features include edge curvature variance and edge gradient mean, the defect edge angle features include end sharpness and scattering spot direction entropy, and the defect refractive features include defect brightness fluctuation value and refractive sensitivity.

5. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The feature fusion is implemented as follows: A multi-level fusion strategy is adopted, which assigns weights based on defect features and sums the weights of each feature to generate defect feature values, specifically as follows: ， Where W represents the defect feature value, F, G and S represent the feature factors corresponding to the shape feature, defect edge angle feature and defect refraction feature respectively, w1, w2 and w3 are the weight coefficients corresponding to different feature factors, w1>0, w2>0, w3>0, and w1+w2+w3=1.

6. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The specific steps for predicting the defect type in the suspected defect area are as follows: The defect confidence score is calculated using the Sigmoid function on the defect feature values ​​obtained after feature fusion, where the defect confidence score represents the probability that the defect category of the suspected defect area is a microcrack. The defect confidence level is compared with the confidence threshold. If the defect confidence level is greater than or equal to the confidence threshold, the defect category of the suspected defect area is predicted to be a microcrack. If the defect confidence level is less than the confidence threshold, the defect category of the suspected defect area is predicted to be a pseudocrack.

7. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The secondary detection module specifically includes: After the special glass passes through the initial inspection area, a second inspection is initiated. Specifically, the inspection is initiated when the glass leaves the edge of the glass by detecting the photoelectric encoder. Before the second inspection is initiated, a unique ID is generated for the suspected defect area, including the inspection date, glass number, and suspected defect area number. The same piece of glass is photographed again by the optical image acquisition module to generate a new sequence of defect images. This allows us to obtain secondary defect features, secondary defect feature values, and secondary defect confidence scores. The secondary defect features, secondary defect feature values, and secondary defect confidence scores are acquired and calculated using the same methods as the defect features, defect feature values, and defect confidence scores.

8. The special glass surface microcrack detection system based on AI vision according to claim 1, characterized in that: The specific operation of the final determination module is as follows: Set an expansion threshold for the rate of change of defect morphology and an increase threshold for the rate of change of defect confidence. When the rate of change of defect morphology is greater than the expansion threshold and the rate of change of defect confidence is greater than the growth threshold, the defect type of the suspected defect area is determined to be a microcrack. If the rate of change of defect morphology is greater than the expansion threshold and the rate of change of defect confidence is less than the growth threshold, the defect type is still determined to be a microcrack. If the rate of change of defect morphology is less than the expansion threshold and the rate of change of defect confidence is less than the growth threshold, then the defect type of the suspected defect area is determined to be a pseudo-crack.

Images