Method for detecting surface defects of backlight LED screen based on machine vision
By combining multimodal collaborative imaging and deep learning methods with transmission and dark-field reflection imaging, and synchronously controlling the LED dimming signal, the problems of low efficiency, high false negative rate and insufficient detection accuracy in the detection of surface defects in backlit LED screens are solved, and efficient and accurate defect identification is achieved.
Patent Information
- Application Number
- CN202511088707.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-11-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing technologies, the detection of surface defects in backlit LED screens relies on manual visual inspection, which is inefficient and has a high rate of missed detection. The single-modal optical imaging mode cannot simultaneously capture brightness abnormalities and physical damage, and the traditional Mura algorithm has insufficient defect recognition capabilities, which limits the improvement of detection accuracy.
A multimodal collaborative imaging method is adopted, combining transmission and dark-field reflection imaging. Defects are identified through a deep learning dual-branch neural network, and LED dimming signals are synchronously controlled to eliminate flicker noise. Dynamic background correction and defect area quantization are used to improve detection accuracy.
It improves the efficiency and accuracy of defect detection in backlit LED screens, reduces the false negative rate, achieves sensitive identification of low-contrast defects and simultaneous detection of transmission/reflection features, and enhances the robustness of the detection algorithm.
Smart Images
Figure CN120908207A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of machine vision detection, and particularly relates to a method for detecting surface defects of a backlight LED screen based on machine vision. BACKGROUND
[0002] An LED screen (Light Emitting Diode panel) is a display screen composed of a light-emitting diode array and used to display various information such as text, graphics, images, animations, market information, videos and video signals. Scratches, depressions, bubbles, foreign matter, color spots, bright spots / dark spots and other defects on the surface of a backlight LED screen will directly affect the display effect and structural stability of the screen. Through detection, unqualified products can be screened out in time to avoid affecting user experience after entering the market.
[0003] Current surface defect detection of a backlight LED screen mainly relies on manual inspection and single-mode optical imaging, and has the following problems: 1. Manual detection is low in efficiency and high in missed detection rate, and is significantly affected by personnel fatigue and subjectivity; 2. Single transmission / reflection imaging mode cannot simultaneously capture brightness abnormalities and physical damage, resulting in missed detection of dark field defects and low-contrast defects; 3. Image stripe noise is caused by PWM high-frequency dimming of the LED, and traditional algorithms have insufficient Mura defect recognition capability, which restricts the improvement of detection accuracy. SUMMARY
[0004] The technical problem to be solved by the application is to overcome the shortcomings of the prior art and provide a method for detecting surface defects of a backlight LED screen based on machine vision.
[0005] The technical solution adopted to solve the above technical problem is a method for detecting surface defects of a backlight LED screen based on machine vision, comprising the following specific steps: Step 1: Multi-modal collaborative imaging, in the on state of the backlight source, a brightness distribution image of the screen is collected by a transmission imaging module, and in the off state of the backlight source, a surface texture image of the screen is collected by a dark field reflection imaging module, wherein the image collection process is hardware-synchronized with the PWM dimming signal of the LED, so that the camera exposure time window is strictly within the LED bright state duration, and the phase-locked deviation is not more than 1 microsecond; Step 2: Dynamic background correction, non-uniformity compensation is performed on the transmission image, based on the local background area gray mean value and standard deviation of the current image, combined with the statistical characteristics of the non-defective template image, the illumination field non-uniformity is eliminated through linear transformation; Step three: deep learning defect recognition, input the corrected transmission image and reflection image into a double-branch neural network model, wherein the first branch extracts the brightness abnormality feature of the transmission image, the second branch extracts the texture defect feature of the reflection image, and the double-branch features are dynamically fused through a channel attention mechanism, and finally a pixel-level defect probability heat map is output; Step four: defect area quantification, based on the defect probability heat map, the pixel area with a probability value greater than 0.8 is binarized, and the percentage of the binarized area in the total area of the effective display area of the screen is calculated as the quantitative indicator of Mura defects.
[0006] Through the above technical solution, the detection efficiency is improved through multi-modal collaborative imaging and deep learning fusion decision, the recognition sensitivity for low-contrast defects is greatly improved, the PWM stripe noise is completely eliminated through hardware-level stroboscopic synchronous control, the stroboscopic is completely eliminated, the image noise is zeroed, the background is corrected, the perspective relationship error is avoided, the data accuracy during detection is improved, the non-uniformity of the projection image gray scale is greatly reduced, the transmission / reflection collaborative imaging realizes synchronous detection of scratches and Mura, the low-contrast Mura missing detection problem is solved by combining the deep learning confidence, and the robustness of the algorithm is improved.
[0007] Further, the transmission imaging module includes a continuous spectrum area light source with a wavelength range of 450-650nm, and the illumination is greater than 10,000lux, and the dark-field reflection imaging module adopts a ring light source structure, and the incident light line and the screen normal form an angle of 30° to 60°, which is used to enhance the scattering signal of the surface micro scratch.
[0008] Through the above technical solution, by setting the angle between the incident light line and the screen normal, the lens glare can be avoided while capturing micro defect scattering, the illumination is greatly improved in the detection process, the defect is difficult to detect due to the problem of illumination, and the detection performance of the micro scratch is improved by limiting the wavelength range, avoiding the problem that a single wavelength light source cannot accurately detect a small scratch.
[0009] Further, the double-branch neural network includes a rule enhancement layer, which performs the following logic: when the local 5x5 pixel neighborhood contrast in the transmission image is greater than 0.25 and the deep learning branch confidence is less than 0.5, it is forced to determine that the area has defects, and the local contrast is defined as the difference between the maximum gray value and the minimum gray value in the neighborhood divided by the maximum gray value.
[0010] Through the above technical solution, the Mura detection rate of low contrast can be improved through the rule bottoming, and the recognition degree of defects can be effectively improved through the complementation of local contrast and deep learning, avoiding the problem that defects are ignored due to algorithm problems.
[0011] Further, the hardware synchronization in the step one includes generating a trigger pulse signal with the same frequency as the PWM dimming frequency by using the FPGA, and the time deviation between the camera exposure starting moment and the rising edge of the LED bright state is controlled within ±1 microsecond, and the exposure duration is less than or equal to the LED bright state duty cycle.
[0012] Through the above technical solution, the image stripe noise caused by the PWM high-frequency dimming of the LED is avoided by hardware synchronization, and the transmission / reflection features are complementary by the double-branch neural network, so that the scratch and Mura are synchronously detected.
[0013] Further, the double-branch neural network includes that the transmission branch extracts multi-scale brightness features by using a ResNet-34 backbone network, the reflection branch extracts texture features by using a lightweight network containing 4 convolutional layers, and the feature fusion layer calculates the channel attention weight by using a Squeeze-and-Excitation module to realize adaptive weighted fusion of features.
[0014] Through the above technical solution, the model volume is greatly reduced, the operation resource utilization rate is improved, and the anti-electromagnetic interference ability is improved to avoid interference from other components.
[0015] Further, the camera exposure time window satisfies the following formula: wherein is the PWM rising edge synchronization deviation, is the LED dimming period, is the LED bright state duty cycle; The dynamic background correction adopts the following specific formula: wherein, and are the average and standard deviation of the gray scale of the 5*5 background window in the current image, and are the corresponding values of the standard flawless template.
[0016] Through the above technical solution, the frequency flash can be eradicated, the image noise can be zeroed, the background can be corrected, the perspective relationship error can be avoided, the data accuracy during detection can be improved, and the non-uniformity of the projection image gray scale can be greatly reduced.
[0017] Further, the flaw area quantification is calculated by using the following specific formula: wherein, is an indicator function, is the area of the effective screen area. The rule for the forced determination of flaws is: wherein is a 5x5 pixel neighborhood centered at (x, y).
[0018] By the above technical solution, combined with deep learning confidence, the low-contrast Mura missing detection problem is solved, and the robustness of the algorithm is improved.
[0019] The beneficial effects of the present application are as follows: the present application improves detection efficiency through multi-modal collaborative imaging and deep learning fusion decision, greatly improves the recognition sensitivity of low-contrast flaws, completely eliminates PWM stripe noise through hardware-level stroboscopic synchronous control, can eliminate stroboscopic, realizes image noise zero, corrects the background at the same time, avoids the error of perspective relationship, improves the accuracy of data during detection, greatly reduces the non-uniformity of projection image gray scale, realizes synchronous detection of scratches and Mura through transmission / reflection collaborative imaging, combined with deep learning confidence, solves the low-contrast Mura missing detection problem, and improves the robustness of the algorithm. BRIEF DESCRIPTION OF DRAWINGS
[0020] Figure 1 is a method flowchart of the present application. DETAILED DESCRIPTION
[0021] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.
[0022] As Figure 1 shown, the method for detecting surface flaws of a backlight LED screen based on machine vision of the present embodiment includes the following specific steps: Step one: multi-modal collaborative imaging, in the on state of the backlight source, the brightness distribution image of the screen is collected through the transmission imaging module, in the off state of the backlight source, the surface texture image of the screen is collected through the dark field reflection imaging module, wherein the image collection process is kept hardware synchronous with the PWM dimming signal of the LED, so that the camera exposure time window is strictly within the LED bright state duration, and the phase lock deviation is not more than 1 microsecond; Step two: dynamic background correction, non-uniformity compensation is performed on the transmission image, based on the local background area gray mean value and standard deviation of the current image, combined with the statistical characteristics of the non-flaw template image, the illumination field non-uniformity is eliminated through linear transformation; Step three: deep learning defect recognition, input the corrected transmission image and reflection image into a double-branch neural network model, wherein the first branch extracts the brightness abnormality feature of the transmission image, the second branch extracts the texture defect feature of the reflection image, and the double-branch features are dynamically fused through a channel attention mechanism, and finally a pixel-level defect probability heat map is output; Step four: defect area quantification, based on the defect probability heat map, the pixel area with a probability value greater than 0.8 is binarized, and the percentage of the binarized area in the total area of the effective display area of the screen is calculated as the quantitative indicator of Mura defect. The detection efficiency is improved through the fusion decision of multi-modal collaborative imaging and deep learning. The recognition sensitivity for low-contrast defects is greatly improved. The PWM stripe noise is completely eliminated through hardware-level stroboscopic synchronous control. The transmission / reflection collaborative imaging realizes synchronous detection of scratches and Mura. Combined with the confidence of deep learning, the low-contrast Mura missing detection problem is solved, and the robustness of the algorithm is improved.
[0023] The transmission imaging module includes a continuous spectrum area light source with a wavelength range of 450-650 nm, and the illumination is greater than 10,000 lux. The dark-field reflection imaging module adopts a ring light source structure, and the incident light is at an angle of 30° to 60° with the screen normal, which is used to enhance the scattering signal of surface micro scratches. Through the angle setting between the incident light and the screen normal, the lens glare can be avoided while capturing micro defect scattering, greatly improving the comprehensiveness and accuracy of illumination during detection, and avoiding defects caused by lighting problems.
[0024] The double-branch neural network includes a rule enhancement layer, which performs the following logic: when the local 5x5 pixel neighborhood contrast in the transmission image is greater than 0.25 and the confidence of the deep learning branch is less than 0.5, it is forced to determine that the area has defects. The local contrast is defined as the difference between the maximum and minimum gray values in the neighborhood divided by the maximum gray value. Through the rule bottom, the Mura detection rate of low contrast can be improved, and through the complement of local contrast and deep learning, the recognition of defects can be effectively improved, avoiding defects caused by algorithm problems.
[0025] The hardware synchronization in step one includes generating a trigger pulse signal with the same frequency as the PWM dimming frequency using FPGA, and the time deviation between the camera exposure start time and the LED bright state rising edge is controlled within ±1 microsecond. The exposure duration is less than or equal to the LED bright state duty cycle. Through hardware synchronization, the PWM high-frequency dimming of LED can avoid image stripe noise. Combined with the double-branch neural network, transmission / reflection feature complementarity is realized, and synchronous detection of scratches and Mura is achieved.
[0026] The double-branch neural network comprises a transmission branch adopting a ResNet-34 backbone network to extract multi-scale brightness features, and a reflection branch adopting a lightweight network comprising four layers of convolution to extract texture features; a feature fusion layer calculates channel attention weights through a Squeeze-and-Excitation module to realize adaptive weighted fusion of features, greatly reduces the model volume, improves the utilization rate of computing resources, and improves the anti-electromagnetic interference capability to avoid interference from other components.
[0027] The camera exposure time window satisfies the following formula: wherein is a PWM rising edge synchronization deviation, is an LED dimming period, is an LED bright state duty cycle; The dynamic background correction adopts the following specific formula: wherein, and are the average and standard deviation of the gray scale of a 5*5 background window in the current image, and are the corresponding values of a standard flawless template, which can eliminate flicker, realize image noise zeroing, correct the background, avoid perspective relationship errors, improve the accuracy of data during detection, and greatly reduce the non-uniformity of the projection image gray scale.
[0028] The flaw area quantification is calculated by the following specific formula: wherein, is an indicator function, is the area of the effective screen area; The rule for forcibly determining the flaw is: wherein is a 5*5 pixel neighborhood centered on (x, y), which, in combination with the deep learning confidence, solves the low-contrast Mura detection problem and improves the robustness of the algorithm.
[0029] The optical system adopts the following configuration: Transmission channel: 5000K color temperature surface light source + 0.8MP global shutter CMOS, pixel size 3.45um; Reflection channel: annular LED dark field light source + 2k line array CCD, scanning speed 0.5m / s; Optical resolution: 12 μm / pixel.
[0030] The above description is only the preferred embodiment of the application, not intended to limit the protection scope of the application.
Claims
1. A method for detecting surface defects of a backlit LED screen based on machine vision, characterized in that, The method comprises the following specific steps: Step one: multi-modal collaborative imaging, under the condition of backlight source opening, the brightness distribution image of the screen is collected by the transmission imaging module, under the condition of backlight source closing, the surface texture image of the screen is collected by the dark-field reflection imaging module, wherein the image collection process is kept hardware synchronization with the PWM dimming signal of the LED, so that the camera exposure time window is strictly within the duration of the LED bright state, and the phase locking deviation is not more than 1 microsecond; Step two: dynamic background correction, non-uniformity compensation is performed on the transmission image, based on the local background area gray mean value and standard deviation of the current image, combined with the statistical characteristics of the flawless template image, the illumination field non-uniformity is eliminated through linear transformation; Step three: deep learning defect identification, the corrected transmission image and the reflection image are input into a double-branch neural network model, wherein the first branch extracts the brightness abnormality features of the transmission image, the second branch extracts the texture defect features of the reflection image, and the double-branch features are dynamically fused through the channel attention mechanism, and finally a pixel-level defect probability heat map is output; Step four: defect area quantification, based on the defect probability heat map, the pixel area with a probability value greater than 0.8 is binarized, the area percentage of the binarized area in the total area of the effective display area of the screen is calculated as the quantitative index of Mura defect.
2. The method for detecting surface defects of a backlit LED screen based on machine vision according to claim 1, wherein, The transmission imaging module comprises a continuous spectrum surface light source with a wavelength range of 450-650 nm, and the illumination is greater than 10,000 lux, and the dark-field reflection imaging module adopts a ring light source structure, and the incident light is at an angle of 30° to 60° with the screen normal, which is used to enhance the scattering signal of the surface micro scratch.
3. The method of claim 2, wherein the method further comprises: The double-branch neural network comprises a rule enhancement layer, which performs the following logic: when the local 5x5 pixel neighborhood contrast in the transmission image is greater than 0.25 and the confidence of the deep learning branch is less than 0.5, the area is forced to be determined as defective, and the local contrast is defined as the difference between the maximum gray value and the minimum gray value in the neighborhood divided by the maximum gray value.
4. The method for detecting surface defects of a backlit LED screen based on machine vision according to claim 3, wherein, The hardware synchronization in step one comprises generating a trigger pulse signal with the same frequency as the PWM dimming frequency using FPGA, and the time deviation between the camera exposure starting time and the rising edge of the LED bright state is controlled within ±1 microsecond, and the exposure duration is less than or equal to the duty cycle of the LED bright state.
5. The method for detecting surface defects of a backlit LED screen based on machine vision according to claim 4, wherein, The double-branch neural network comprises a ResNet-34 backbone network for the transmission branch to extract multi-scale brightness features, a lightweight network comprising 4 convolutional layers for the reflection branch to extract texture features, and a Squeeze-and-Excitation module for the feature fusion layer to calculate the channel attention weight, realizing adaptive weighted fusion of features.
6. The method of claim 1, wherein the method further comprises: The camera exposure time window satisfies the following formula: wherein is a PWM rising edge synchronization deviation, is an LED dimming period, is an LED bright state duty cycle time; The dynamic background correction adopts the following specific formula: wherein, and is the mean and standard deviation of the gray level of the 5x5 background window in the current image, and are the corresponding values of the standard flawless template.
7. The method of claim 1, wherein the method further comprises: The defect area quantification is calculated by the following specific formula: wherein is an indicator function, is the area of the screen active area; The rule for forced judgment of defects is: wherein is a 5x5 pixel neighborhood centered at (x, y).