LED screen surface damage detection method based on machine vision

Abstract

The invention discloses an LED screen surface damage detection method based on machine vision, and belongs to the technical field of machine vision recognized.The LED screen surface damage detection method based on machine vision comprises the following specific steps that firstly, a light source device is arranged, a flexible OLED light source module with the adjustable angle is used and installed around an LED screen, and the flexible OLED light source module with the adjustable angle is arranged around the LED screen; it is ensured that a light source can be evenly adjusted within the range of 0-180 degrees, the angle of the light source is based on the normal of a screen, and five angles including 0 degree, 45 degrees, 90 degrees, 135 degrees and 180 degrees are preset through an embedded controller; and 2, image acquisition control. Through a multi-angle illumination strategy (0-180-degree dynamic adjustment), the problem of missing detection caused by reflection angle difference of defects under a traditional single light source is solved, dynamic gray anomaly analysis is combined with DBSCAN clustering, real defects and environmental noise are effectively distinguished, simultaneous processing and fusion of multi-light-source angle images are supported, and the method is suitable for large-scale popularization and application. And the time accumulation problem of traditional serial detection is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine vision recognition, and in particular relates to a method for detecting surface damage of an LED screen based on machine vision. Background Art

[0002] During production and use, LED screens often suffer surface damage, such as cracks and dead pixels, due to physical impact, environmental factors, or manufacturing process issues. Existing technologies suffer from the following issues: Single illumination angle limitation: Traditional methods rely on a fixed light source angle to capture images, resulting in the same defect appearing differently under different illumination angles, making it difficult to fully capture the defect area; High false positive rate: Manual inspection or simple grayscale thresholding methods are susceptible to environmental interference (such as reflections and dust), resulting in a high false positive rate; Inefficiency: Manual inspection is time-consuming and costly, making it inefficient for large-scale production. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method for detecting surface damage of LED screens based on machine vision.

[0004] The technical solution adopted to solve the above technical problems is: a method for detecting surface damage of LED screens based on machine vision, including the following specific steps: Step 1: Set up the light source device. Use an adjustable flexible OLED light source module and install it around the LED screen. Ensure that the light source can be evenly adjusted within the range of 0° to 180°. The light source angle is based on the screen normal and is preset through the embedded controller to five angles: 0°, 45°, 90°, 135°, and 180°. Step 2: Image acquisition control: A high-resolution CCD camera is paired with an X / Y / Z-axis mobile platform to ensure that the camera is facing the screen surface. At each lighting angle, the camera automatically focuses and captures grayscale images. The control module synchronizes light source angle adjustment with camera triggering to achieve continuous acquisition of multi-angle images. After each angle adjustment, the system waits for the light source to stabilize before shooting, with a waiting time of 0.5 seconds. Step 3: Output the results and obtain a collection of multiple grayscale images of the same LED screen under different lighting angles ; Step 4: Dynamic grayscale anomaly analysis: load the pre-stored normal sample library, which contains several defect-free images, calculate the grayscale mean of each pixel position, and calculate the grayscale difference for each pixel of the image to be inspected; Step 5: Construct an exponential anomaly scoring function based on the difference, then set the domain radius to 5 pixels and the minimum number of points MinPts = 5. Perform density clustering on the abnormal pixels, traverse all pixels, mark the core points that meet the number of abnormal points in the domain ≥ MinPts, trigger from the core points, expand and connect all density-reachable abnormal points to form a candidate defect area; Step 6: Defect classification and output: Use the pre-trained PP-YOLOE model, input the defect area image block at the optimal angle, crop it to 256×256 pixels, input the normalized grayscale image, output the defect category and confidence level. When the confidence level is greater than 0.95, it is determined to be a valid defect, and the defect location is marked with a red box in the visualization interface, and the defect type and size are marked with text labels.

[0005] The above technical solution solves the problem of missed detection of defects due to differences in reflection angles under traditional single light sources. Dynamic grayscale anomaly analysis combined with DBSCAN clustering effectively distinguishes real defects from environmental noise. At the same time, it supports the simultaneous processing and fusion of images from multiple light source angles, avoiding the time accumulation problem of traditional serial detection.

[0006] Furthermore, the light source module has high brightness and uniformity, ensuring consistent light intensity after each angle adjustment; The light source module consists of multiple independently controllable OLED light-emitting units, arranged in a ring or linear array, covering the detection area of ​​the LED screen. The brightness of each light-emitting unit is greater than 1000cd / m², and the color temperature of each light-emitting unit can be adjusted from 3000K to 6500K to adapt to the reflective characteristics of screens of different materials. An integrated absolute encoder provides real-time feedback on the angular position of the light source with an error of ≤±0.1°. The brightness of each OLED light-emitting unit is independently adjusted through a PWM signal, and the duty cycle adjustment range is 0% to 100% with an accuracy of ±0.5%. The OLED light-emitting unit has a built-in current sensor to monitor and compensate the brightness of the OLED in real time.

[0007] Through the above technical solution, external light interference is avoided.

[0008] Furthermore, the grayscale image is collected using the following specific formula: Convert the RGB image captured by the camera into HSV space and extract the saturation S channel: ,Then apply a Gaussian filter to the S channel image with a convolution kernel size of 5×5 and a standard deviation of , the formula is: , and finally calculate the image grayscale mean and standard deviation , normalize the grayscale value of each pixel: .

[0009] Through the above technical solution, the RGB image is converted to the HSV space, the saturation contrast of the defect area is enhanced, the high-frequency noise in the image is suppressed, and the edge information of the defect is retained.

[0010] Furthermore, the flexible OLED light source module controls the angle through a stepping motor, with each step being 1.8°, ensuring an angle accuracy of ±0.5°, and the sample library is updated once a month.

[0011] Through the above technical solution, the accuracy of the illumination angle can be greatly improved, while avoiding errors in data offset caused by the long sample library time.

[0012] Furthermore, the grayscale difference calculation adopts the following specific formula: in, is the grayscale difference value, is the pixel position The detection grayscale, is the mean grayscale value of the pixel position.

[0013] Furthermore, the exponential anomaly scoring function formula is: Set to 1.5 times the normal sample grayscale standard deviation, when When a pixel is detected as abnormal, the candidate defect area and its coordinate position are preliminarily detected; For each candidate defect area at each angle, calculate the sum of its edge gradient strength. The specific formula is as follows: in and are the gradients of the image in the x and y directions, respectively, calculated by the Sobel operator, The candidate defect area detected at the current angle. The higher the gradient intensity and the clearer the edge, the better the imaging quality. Then, by comparing the similarity of the grayscale distribution of the same candidate area at different angles, retain the angle combination with a similarity greater than 0.9 and select the imaging quality score. The largest angle is taken as the best result, and the final defect area position and the corresponding optimal illumination angle image after fusion are output.

[0014] Furthermore, the light source module adopts brightness feedback adjustment: in, As the benchmark brightness, ensure the grayscale consistency of the defect area at different angles. To compensate the value, it is calibrated in real time through the light intensity sensor; In addition to conventional white light, infrared (850nm) and ultraviolet (365nm) light source modules are added. Infrared light is used to penetrate surface stains to detect internal cracks, and ultraviolet light stimulates the fluorescent reaction of the screen coating to enhance the visibility of microscopic scratches on the surface. Multispectral images are integrated through a weighted fusion algorithm to improve the detection rate of complex defects. The light source array has a built-in macro adjustment mechanism that can automatically adjust the light source distance according to the screen size to ensure uniform lighting on the surface of screens of different sizes.

[0015] Furthermore, for curved LED screens, the system calculates the screen curvature radius in real time and dynamically adjusts the light source angle to ensure that the light always enters the detection area vertically. When detecting a curved screen with a radius of R, the light source angle θ is dynamically corrected according to the detection point position to: in, To detect the horizontal offset of the point, the algorithm is executed in real time by the embedded controller, with a compensation accuracy of ±0.1°.

[0016] Through the above technical solution, the screen curvature radius can be calculated in real time, and the light source angle can be dynamically adjusted to ensure that the light is always incident vertically on the detection area. Good defect detection can also be performed on curved screens, which improves the adaptability to different products.

[0017] Furthermore, the collaborative process of hardware and software is as follows: the controller sends instructions to adjust the light source angle, then triggers the camera to shoot, and then transmits the captured image to the industrial computer. If image acquisition at a certain angle fails, it will automatically retry 3 times and then switch to the adjacent angle. When a hardware abnormality is detected, an alarm is triggered and the process is paused.

[0018] The beneficial effects of the present invention are as follows: (1) The present invention solves the problem of missed detection of defects caused by reflection angle differences under traditional single light sources through a multi-angle lighting strategy (dynamic adjustment from 0° to 180°). Dynamic grayscale anomaly analysis combined with DBSCAN clustering effectively distinguishes real defects from environmental noise, and supports simultaneous processing and fusion of images from multiple light source angles, avoiding the time accumulation problem of traditional serial detection; (2) The present invention eliminates the influence of uneven lighting and noise through grayscale normalization, and maintains stable performance in a complex lighting workshop environment. It supports online updates through a normal sample library, adapts to aging or batch differences of LED screens, and avoids performance degradation during long-term use. It supports the detection of various defects such as cracks, scratches, bad pixels, and black spots. Through early and accurate detection, defective screens are prevented from flowing into downstream processes, and the overall yield rate is improved. At the same time, the detection results automatically generate a report containing the defect location, type, and image, supporting production quality traceability and process optimization; (3) The present invention calculates the screen curvature radius in real time and dynamically adjusts the light source angle to ensure that the light is always incident vertically on the detection area. It can also perform good defect detection on curved screens, thereby improving adaptability to different products. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0021] like Figure 1 As shown, a method for detecting surface damage of an LED screen based on machine vision in this embodiment includes the following specific steps: Step 1: Set up the light source device. Use an adjustable flexible OLED light source module and install it around the LED screen. Ensure that the light source can be evenly adjusted within the range of 0° to 180°. The light source angle is based on the screen normal and is preset through the embedded controller to five angles: 0°, 45°, 90°, 135°, and 180°. Step 2: Image acquisition control: A high-resolution CCD camera is paired with an X / Y / Z-axis mobile platform to ensure that the camera is facing the screen surface. At each lighting angle, the camera automatically focuses and captures grayscale images. The control module synchronizes light source angle adjustment with camera triggering to achieve continuous acquisition of multi-angle images. After each angle adjustment, the system waits for the light source to stabilize before shooting, with a waiting time of 0.5 seconds. Step 3: Output the results and obtain a collection of multiple grayscale images of the same LED screen under different lighting angles ; Step 4: Dynamic grayscale anomaly analysis: load the pre-stored normal sample library, which contains several defect-free images, calculate the grayscale mean of each pixel position, and calculate the grayscale difference for each pixel of the image to be inspected; Step 5: Construct an exponential anomaly scoring function based on the difference, then set the domain radius to 5 pixels and the minimum number of points MinPts = 5. Perform density clustering on the abnormal pixels, traverse all pixels, mark the core points that meet the number of abnormal points in the domain ≥ MinPts, trigger from the core points, expand and connect all density-reachable abnormal points to form a candidate defect area; Step 6: Defect classification and output: Use the pre-trained PP-YOLOE model, input the defect area image block at the optimal angle, crop it to 256×256 pixels, input the normalized grayscale image, output the defect category and confidence level. When the confidence level is greater than 0.95, it is determined to be a valid defect, and the defect location is marked with a red box in the visualization interface, and the defect type and size are marked with text labels.

[0022] It solves the problem of missed detection of defects due to differences in reflection angles under traditional single light sources. Dynamic grayscale anomaly analysis combined with DBSCAN clustering effectively distinguishes real defects from environmental noise. At the same time, it supports simultaneous processing and fusion of images from multiple light source angles, avoiding the time accumulation problem of traditional serial detection.

[0023] The light source module has high brightness and uniformity, ensuring consistent light intensity after each angle adjustment and avoiding external light interference; The light source module consists of multiple independently controllable OLED light-emitting units, arranged in a ring or linear array, covering the detection area of ​​the LED screen. The brightness of each light-emitting unit is greater than 1000cd / m², and the color temperature of each light-emitting unit can be adjusted from 3000K to 6500K to adapt to the reflective characteristics of screens of different materials. An integrated absolute encoder provides real-time feedback on the angular position of the light source with an error of ≤±0.1°. The brightness of each OLED light-emitting unit is independently adjusted through a PWM signal, and the duty cycle adjustment range is 0% to 100% with an accuracy of ±0.5%. The OLED light-emitting unit has a built-in current sensor to monitor and compensate the brightness of the OLED in real time.

[0024] The following specific formula is used to collect grayscale images: Convert the RGB image captured by the camera into HSV space and extract the saturation S channel: ,Then apply a Gaussian filter to the S channel image with a convolution kernel size of 5×5 and a standard deviation of , the formula is: , and finally calculate the image grayscale mean and standard deviation , normalize the grayscale value of each pixel: , convert the RGB image to HSV space, enhance the saturation contrast of the defect area, suppress the high-frequency noise in the image, and retain the edge information of the defect.

[0025] The flexible OLED light source module uses a stepper motor to control the angle, with each step being 1.8°, ensuring an angle accuracy of ±0.5°. The sample library is updated once a month, which can greatly improve the accuracy of the lighting angle and avoid errors in data offset caused by excessive sample library time.

[0026] The grayscale difference calculation adopts the following specific formula: in, is the grayscale difference value, is the pixel position The detection grayscale, is the mean grayscale value of the pixel position.

[0027] The formula for the exponential anomaly scoring function is: Set to 1.5 times the normal sample grayscale standard deviation, when When a pixel is detected as abnormal, the candidate defect area and its coordinate position are preliminarily detected; For each candidate defect area at each angle, calculate the sum of its edge gradient strength. The specific formula is as follows: in and are the gradients of the image in the x and y directions, respectively, calculated by the Sobel operator, The candidate defect area detected at the current angle. The higher the gradient intensity and the clearer the edge, the better the imaging quality. Then, by comparing the similarity of the grayscale distribution of the same candidate area at different angles, retain the angle combination with a similarity greater than 0.9 and select the imaging quality score. The largest angle is taken as the best result, and the final defect area position and the corresponding optimal illumination angle image after fusion are output.

[0028] The light source module adopts brightness feedback adjustment: in, As the benchmark brightness, ensure the grayscale consistency of the defect area at different angles. To compensate the value, it is calibrated in real time through the light intensity sensor; In addition to conventional white light, infrared (850nm) and ultraviolet (365nm) light source modules are added. Infrared light is used to penetrate surface stains to detect internal cracks, and ultraviolet light stimulates the fluorescent reaction of the screen coating to enhance the visibility of microscopic scratches on the surface. Multispectral images are integrated through a weighted fusion algorithm to improve the detection rate of complex defects. The light source array has a built-in macro adjustment mechanism that can automatically adjust the light source distance according to the screen size to ensure uniform lighting on the surface of screens of different sizes.

[0029] For curved LED screens, the system calculates the screen curvature radius in real time and dynamically adjusts the light source angle to ensure that the light always enters the detection area vertically. When detecting a curved screen with a radius of R, the light source angle θ is dynamically corrected according to the detection point position to: in, To detect the horizontal offset of the detection point, the algorithm is executed in real time through the embedded controller with a compensation accuracy of ±0.1°. It can calculate the screen curvature radius in real time and dynamically adjust the light source angle to ensure that the light is always incident vertically on the detection area. It can also perform good defect detection on curved screens, improving adaptability to different products.

[0030] The hardware and software collaboration process is as follows: the controller sends instructions to adjust the light source angle, then triggers the camera to shoot, and then transmits the captured image to the industrial computer. If image acquisition fails at a certain angle, it automatically retries three times and then switches to the adjacent angle. When a hardware abnormality is detected, an alarm is triggered and the process is suspended.

[0031] Customized training of deep learning classification models Small sample data augmentation strategies include: Physical simulation defect generation: A ray tracing engine was used to simulate defect morphology under different lighting angles, generating over 100,000 synthetic images. Simulation parameters included: crack depth: 10-50μm; scratch width: 5-20μm; bad spot diameter: 0.1-0.5mm; Synthetic data and real data are mixed in a 1:1 ratio for training to solve the problem of scarcity of labeled data in industrial scenarios.

[0032] Then deploy the model in a lightweight way: Reducing model size through knowledge distillation and quantization compression: The knowledge of the PP-YOLOE model was distilled into a lightweight MobileNetV3, reducing the model size to 1 / 5 while maintaining 95% accuracy. 8-bit integer quantization of weights increased inference speed by 3 times, making it suitable for embedded devices. The final model achieved a mean average precision (mAP) of 98.2% on the test set, with single-frame inference time ≤ 50ms.

[0033] System integration and industrial deployment solutions: Hardware architecture design: Modular detection unit: Optical module: integrated multi-spectral light source, high-resolution camera with 50 million pixels, frame rate of 30fps, and polarization filter to eliminate specular reflection.

[0034] Motion control module: The six-axis robotic arm is equipped with a detection unit, supporting fully automatic scanning and detection of large screens.

[0035] Edge computing nodes: Equipped with NVIDIA Jetson AGX Xavier for localized real-time processing.

[0036] Software platform function expansion: Cloud-based quality monitoring dashboard: The test data is uploaded to the cloud via 5G to generate a multi-dimensional quality analysis report: Real-time defect distribution heat map: locate high-frequency defect areas on the production line.

[0037] Process parameter correlation analysis: Statistically analyze the relationship between defect rate and process parameters such as temperature, humidity, and pressure to guide production optimization.

[0038] Predictive maintenance reminders: predict equipment failure cycles based on historical data.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A method for detecting surface damage of LED screens based on machine vision, characterized in that: The specific steps include: Step 1: Set up the light source device. Use an adjustable flexible OLED light source module and install it around the LED screen. Ensure that the light source can be evenly adjusted within the range of 0° to 180°. The light source angle is based on the screen normal and is preset through the embedded controller to five angles: 0°, 45°, 90°, 135°, and 180°. Step 2: Image acquisition control: A high-resolution CCD camera is paired with an X / Y / Z-axis mobile platform to ensure that the camera is facing the screen surface. At each lighting angle, the camera automatically focuses and captures grayscale images. The control module synchronizes light source angle adjustment with camera triggering to achieve continuous acquisition of multi-angle images. After each angle adjustment, the system waits for the light source to stabilize before shooting, with a waiting time of 0.5 seconds. Step 3: Output the results and obtain a collection of multiple grayscale images of the same LED screen under different lighting angles ; Step 4: Dynamic grayscale anomaly analysis: load the pre-stored normal sample library, which contains several defect-free images, calculate the grayscale mean of each pixel position, and calculate the grayscale difference for each pixel of the image to be inspected; Step 5: Construct an exponential anomaly scoring function based on the difference, then set the domain radius to 5 pixels and the minimum number of points MinPts = 5. Perform density clustering on the abnormal pixels, traverse all pixels, mark the core points that meet the number of abnormal points in the domain ≥ MinPts, trigger from the core points, expand and connect all density-reachable abnormal points to form a candidate defect area; Step 6: Defect classification and output: Use the pre-trained PP-YOLOE model, input the defect area image block at the optimal angle, crop it to 256×256 pixels, input the normalized grayscale image, output the defect category and confidence level. When the confidence level is greater than 0.95, it is determined to be a valid defect, and the defect location is marked with a red box in the visualization interface, and the defect type and size are marked with text labels.

2. The method for detecting surface damage of an LED screen based on machine vision according to claim 1, characterized in that: The light source module has high brightness and uniformity, avoiding external light interference and ensuring consistent light intensity after each angle adjustment; The light source module consists of multiple independently controllable OLED light-emitting units, arranged in a ring or linear array, covering the detection area of ​​the LED screen. The brightness of each light-emitting unit is greater than 1000cd / m², and the color temperature of each light-emitting unit can be adjusted from 3000K to 6500K to adapt to the reflective characteristics of screens of different materials. An integrated absolute encoder provides real-time feedback on the angular position of the light source with an error of ≤±0.1°. The brightness of each OLED light-emitting unit is independently adjusted through a PWM signal, and the duty cycle adjustment range is 0% to 100% with an accuracy of ±0.5%. The OLED light-emitting unit has a built-in current sensor to monitor and compensate the brightness of the OLED in real time.

3. The method for detecting surface damage of an LED screen based on machine vision according to claim 2, characterized in that: The following specific formula is used to collect grayscale images: Convert the RGB image captured by the camera into HSV space and extract the saturation S channel: ,Then apply a Gaussian filter to the S channel image with a convolution kernel size of 5×5 and a standard deviation of , the formula is: , and finally calculate the image grayscale mean and standard deviation , normalize the grayscale value of each pixel: .

4. The method for detecting surface damage of an LED screen based on machine vision according to claim 3, characterized in that: The flexible OLED light source module is angle-controlled by a stepper motor, with each step being 1.8°, ensuring an angle accuracy of ±0.5°. The bracket is made of lightweight carbon fiber, with a rotation speed of ≥30° / second. The sample library is updated once a month.

5. The method for detecting surface damage of an LED screen based on machine vision according to claim 4, characterized in that: The grayscale difference calculation adopts the following specific formula: in, is the grayscale difference value, is the pixel position The detection grayscale, is the mean grayscale value of the pixel position.

6. The method for detecting surface damage of an LED screen based on machine vision according to claim 5, characterized in that: The formula for the exponential anomaly scoring function is: Set to 1.5 times the normal sample grayscale standard deviation, when When a pixel is detected as abnormal, the candidate defect area and its coordinate position are preliminarily detected; For each candidate defect area at each angle, calculate the sum of its edge gradient strength. The specific formula is as follows: in and are the gradients of the image in the x and y directions, respectively, calculated by the Sobel operator, The candidate defect area detected at the current angle. The higher the gradient intensity and the clearer the edge, the better the imaging quality. Then, by comparing the similarity of the grayscale distribution of the same candidate area at different angles, retain the angle combination with a similarity greater than 0.9 and select the imaging quality score. The largest angle is taken as the best result, and the final defect area position and the corresponding optimal illumination angle image after fusion are output.

7. The method for detecting surface damage of an LED screen based on machine vision according to claim 6, characterized in that: The light source module adopts brightness feedback adjustment: in, As the benchmark brightness, ensure the grayscale consistency of the defect area at different angles. To compensate the value, it is calibrated in real time through the light intensity sensor; In addition to conventional white light, infrared (850nm) and ultraviolet (365nm) light source modules are added. Infrared light is used to penetrate surface stains to detect internal cracks, and ultraviolet light stimulates the fluorescent reaction of the screen coating to enhance the visibility of microscopic scratches on the surface. Multispectral images are integrated through a weighted fusion algorithm to improve the detection rate of complex defects. The light source array has a built-in macro adjustment mechanism that can automatically adjust the light source distance according to the screen size to ensure uniform lighting on the surface of screens of different sizes.

8. The method for detecting surface damage of an LED screen based on machine vision according to claim 7, characterized in that: For curved LED screens, the system calculates the screen curvature radius in real time and dynamically adjusts the light source angle to ensure that the light always enters the detection area vertically. When detecting a curved screen with a radius of R, the light source angle θ is dynamically corrected according to the detection point position to: in, To detect the horizontal offset of the point, the algorithm is executed in real time by the embedded controller, with a compensation accuracy of ±0.1°.

9. The method for detecting surface damage of an LED screen based on machine vision according to claim 8, characterized in that: The hardware and software collaboration process is as follows: the controller sends instructions to adjust the light source angle, then triggers the camera to shoot, and then transmits the captured image to the industrial computer. If image acquisition fails at a certain angle, it automatically retries three times and then switches to the adjacent angle. When a hardware abnormality is detected, an alarm is triggered and the process is suspended.

Images