Sub-pixel defect detection and judgment method and system based on dual-camera cooperation and multi-channel color segmentation

Abstract

The invention discloses a sub-pixel defect detection and judgment method and system based on dual-camera collaboration and multi-channel color segmentation. The method comprises the steps of dual-camera collaboration collection, defect positioning and sub-pixel-level background modeling, multi-channel color point color judgment, black and white image defect extraction and brightness measurement and dynamic defect judgment. According to the method, the problems of sensitivity, positioning precision, color judgment accuracy and judgment consistency of sub-pixel defect detection under the low-gray-scale (L0 level) working condition are solved, the color point recognition accuracy rate reaches 97.2%, the human vision correlation coefficient reaches 0.92, and the method is suitable for factory detection of self-luminous displays such as OLEDs and Micro-LEDs.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing technology for display panel defects. Specifically, it relates to a method for detecting and judging sub-pixel defects based on dual-camera collaboration and multi-channel color segmentation. It is particularly suitable for detecting micron-level sub-pixel bright spot defects (including white spots and colored spots) in self-emissive displays such as OLED and Micro-LED under low grayscale (L0 level) conditions, and can be widely used in the factory quality inspection process of display panels. Background Technology

[0002] In the display panel manufacturing industry, especially in the production of self-emissive displays such as OLED and Micro-LED, the detection of sub-pixel-level bright spot defects (such as white spots and colored spots) is a critical step in ensuring product quality. These defects are typically on the micrometer scale and have extremely low brightness under low grayscale (L0 level) conditions, posing a severe challenge to detection technology.

[0003] The current mainstream testing technologies in the industry have the following significant limitations: The physical contradictions of a single-camera acquisition system: Although a color camera can acquire RGB / HSV color information, the Bayer filter structure reduces photon capture efficiency by more than 60%, resulting in a brightness detection signal-to-noise ratio (SNR) of less than 5dB at the L0 grayscale and extremely low sensitivity in low light. On the other hand, while removing the filter from a monochrome camera improves low light sensitivity, it completely loses its color classification ability and cannot distinguish between red, green, and blue defect types. As a result, a single-camera solution cannot simultaneously meet the requirements of color recognition and low light sensitivity.

[0004] Insufficient defect location accuracy: Traditional algorithms are mostly based on global thresholds or screen brightness differences, without considering the unevenness of weak light emission of sub-pixels in L0 state, resulting in defect position deviation of more than 10μm, making it impossible to achieve precise sub-pixel-level location.

[0005] Poor accuracy in color determination of colored dots: High-brightness colored dots are prone to single-channel saturation under high exposure, resulting in a central whitening phenomenon (e.g., when the R value of a bright red sub-pixel is equal to 255, it is recorded as (255,255,255) white). Traditional methods only determine the defect type based on the central white area, which easily leads to color misjudgment. At the same time, after overexposure, the light spot will expand to 3-5 times the actual area, which seriously interferes with the counting of defect numbers, resulting in a colored dot recognition accuracy of only 78% under L0 grayscale.

[0006] The judgment threshold design is unreasonable: existing algorithms mostly use a single global brightness threshold (such as a fixed gray value of 120) for defect judgment, without considering the differences in brightness perception of different colors such as red, green and blue by the human eye, which leads to systematic misjudgment and poor consistency between the detection results and human vision.

[0007] To address the aforementioned technical bottlenecks, this invention proposes a method and system for detecting and judging sub-pixel defects based on dual-camera collaboration and multi-channel color segmentation. Through hardware collaboration, algorithm optimization, and innovative judgment logic, it achieves high-precision and high-reliability detection of sub-pixel defects under low grayscale conditions. Summary of the Invention

[0008] The core objective of this invention is to solve the problems of existing display panel defect detection technologies in low grayscale (L0 level) conditions, such as the inability to balance color and sensitivity, low positioning accuracy, high color misjudgment rate, and unreasonable judgment logic. Specifically, it aims to achieve the following goals: overcome the physical limitations of a single camera to achieve synchronous and accurate acquisition of brightness and color signals; establish a sub-pixel-level background model to improve defect positioning accuracy to the micrometer level. A multi-channel fusion color determination method is proposed to eliminate misjudgments caused by overexposure and whitening; a dynamic determination system that conforms to human visual perception is constructed to improve the consistency between detection results and actual visual effects.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting and determining sub-pixel defects based on dual-camera collaboration and multi-channel color segmentation includes the following steps: Step S1: Dual-camera collaborative acquisition: The acquisition system, consisting of a main camera (high-sensitivity monochrome camera) and an auxiliary camera (high-resolution color camera), synchronously acquires images of the display panel at low grayscale (L0 level). The main camera uses a long exposure mode to capture the brightness signal, while the auxiliary camera uses a short exposure mode to freeze the original color information. The hardware synchronization timing error of the two cameras is ≤1ms, and the optical path consistency is achieved by sharing the same optical axis through a beam splitter prism, with a beam splitting ratio of 50:50. Step S2: Defect localization and sub-pixel-level background modeling: The color image acquired by the auxiliary camera is converted to HSV space. In the V channel, candidate defect regions are initially extracted through local contrast enhancement and adaptive threshold segmentation. In the H channel, red, green and blue sub-pixel space templates are constructed by analyzing the periodic distribution of hue values. Combined with the RGB channel and HSV space multi-channel fusion segmentation results of the defect-free screen, a background sub-pixel template with an accuracy of 1.5μm is generated to achieve sub-pixel-level defect localization. Step S3: Multi-channel color determination: Perform RGB differential enhancement processing on the color image acquired by the auxiliary camera to obtain enhancement masks corresponding to the red, green and blue channels. Perform spatial verification on the enhancement masks and the sub-pixel space template of the H channel. Combine the hue features of the HSV channel to jointly determine the defect color type. The color type includes red (R), green (G), blue (B) and white (W). Step S4: Defect extraction and brightness measurement of black and white image: First, perform median filtering or bilateral filtering on the black and white image captured by the main camera, then extract the defect area, and calculate the average brightness value (MeanValue) and the average gray value (GrayMax100) of the brightest 100 pixels of the defect area. Step S5: Dynamic Defect Judgment: Based on the defect color type determined in Step S3, call the pre-stored physical limit sample calibration threshold database (to obtain the average grayscale reference value T of the limit sample corresponding to that color). cmean The maximum grayscale reference value T for the brightest 100 pixels. cmax Then, combined with the average brightness value of the defect area and the average gray value of the brightest 100 pixels calculated in step S4, a comprehensive judgment score is calculated using a preset formula; the comprehensive judgment score is compared with a preset judgment threshold (a pre-set quantization threshold, such as 0.2). If the score is ≥ the preset judgment threshold, an unqualified judgment result (NG) is output; if the score is < the preset judgment threshold, a qualified judgment result (OK) is output.

[0010] Furthermore, the main camera is configured with the following parameters: resolution 4096×3000 (12 Megapixels), field of view (FOV) 14.1mm×10.3mm, imaging accuracy 3μm / pixel, exposure time 5ms, and output bit depth 12bit; the auxiliary camera is configured with the following parameters: resolution 4096×3000 (12 Megapixels), pixel size 3.45μm×3.45μm, equipped with an RK-TC5M230-10150 telecentric lens, magnification 1×, field of view (FOV) 14.1mm×10.3mm, imaging accuracy 3μm / pixel, exposure time 0.5ms, and output bit depth 12bit.

[0011] Furthermore, the specific formula for converting the color image to HSV space in step S2 is as follows: Calculate the maximum value C max =max(R,G,B), minimum value C min =min(R,G,B) and the difference Δ=C max -C min ; Calculate hue H: ; Calculate the saturation S: ; Calculate the lightness V: V=C max Where R, G, and B are the pixel brightness values ​​of the red, green, and blue channels of the color image, respectively.

[0012] Furthermore, the specific formula for RGB differential enhancement in step S3 is as follows: Red defect enhancement mask: ; where I R (x,y), I G (x,y), I B (x, y) represent the pixel brightness values ​​of the red, green, and blue channels at image coordinates (x, y), respectively. and is the empirical threshold, and ∧ is the logical AND operation.

[0013] Furthermore, the specific logic for the multi-channel joint determination in step S3 is as follows: Spatial verification: D R =M R ∩T R D G =M G ∩T G D B =M B ∩T B M R M G M B These are the masking results after RGB differential enhancement; T R ,T G T B These are the red, blue, and green sub-pixel space templates on the hue map, with ∩ representing the intersection; Color confirmation: If the mask result of RGB differential enhancement is consistent with the main color corresponding to the H channel, then the defect color is confirmed to be R, G, or B; If three consecutive colored dots of different colors are detected, then it is determined to be a white dot.

[0014] Furthermore, the formula for filtering in step S4 is: I f (x,y)=medianfiter(I C (x,y),k), where I c (x, y) represents the pixel values ​​of the original black and white image, k is the filter kernel size, and I f (x,y) represents the pixel values ​​of the filtered image.

[0015] Furthermore, the formula for calculating the average brightness value in step S4 is as follows: , where Ω is the defect area, N is the number of pixels in the area, and I(x,y) is the pixel brightness value of the defect area.

[0016] Furthermore, the formula for calculating the average grayscale value of the brightest 100 pixels in step S4 is as follows: The Top100(Ω) represents the 100 brightest pixels in the defect area, sorted by brightness in descending order.

[0017] Furthermore, the formula for calculating the comprehensive judgment score in step S5 is as follows: Where W1 and W2 are weighting coefficients; T C,meanThis is the average grayscale reference value for the limit sample of the corresponding color type; T C,max The grayscale reference value is the brightest 100 pixels in the limit sample for the corresponding color type; the judgment criterion is: , To determine the threshold, products exceeding the threshold are considered NG (Not Good) products, and products below the threshold are considered OK (Good) products.

[0018] Furthermore, the method is applicable to the factory quality inspection of OLED and Micro-LED self-emissive displays, focusing on the detection and judgment of micron-level sub-pixel bright spot defects under low grayscale conditions.

[0019] The present invention also provides a sub-pixel defect detection and judgment system based on dual-camera collaboration and multi-channel color segmentation, comprising: a dual-camera collaborative acquisition system, consisting of a main camera (black and white) and an auxiliary camera (color); An image processing unit, configured as follows: Defect localization: The color image from the auxiliary camera is converted to HSV space, defects are initially extracted in the V channel, and a sub-pixel template is constructed in the H channel to achieve sub-pixel level defect localization; Color determination: By combining the brightness comparison of the RGB channels with the hue information of the H channel, the defect color type (R / G / B / W) is determined. Brightness measurement: Utilizing the high-sensitivity black and white image of the main camera, the gray value of the defect area is measured; Defect determination: Based on the measured feature values ​​and the physical limit sample, a dynamic comparison is performed to output the final defect determination result.

[0020] Furthermore, in the defect localization step, the spatial arrangement of red, green, and blue sub-pixels is identified and located by analyzing the periodic distribution of H (hue) channel values; in the color determination step, if the color initially determined by the RGB channel is consistent with the main color of the H channel, the color type is confirmed; if three consecutive colored dots are detected, they are determined to be white dots (W); brightness measurement includes calculating the average brightness value (MeanValue) of the defect area and the average grayscale value (GrayMax100) of the brightest 100 pixels; defect determination is achieved by calculating a comprehensive determination score.

[0021] The present invention has the following beneficial effects: Complementary acquisition of luminance and color signals: Through the collaboration of dual cameras, the parallel acquisition of luminance and color signals is realized, which effectively overcomes the detection blind zone of a single camera under low brightness conditions, solves the physical contradiction of low sensitivity of color cameras in low light and lack of color recognition capability of black and white cameras, and improves the signal-to-noise ratio of luminance detection at L0 grayscale to more than 15dB.

[0022] Sub-pixel-level high-precision positioning: The background sub-pixel template generated by multi-channel fusion segmentation has a resolution of 1.5μm, realizing sub-pixel-level brightness difference modeling. The defect positioning deviation is controlled within 2μm, which improves the positioning accuracy by 80% compared with the traditional method (deviation of more than 10μm).

[0023] High-accuracy color determination: Through the RGB difference enhancement-spatial template verification-multi-feature fusion technology path, the accuracy of color dot recognition at L0 low grayscale is improved from 78% to 97.2% of the traditional method, and the overexposure whitening misjudgment elimination rate reaches 92.5%, which completely solves the problems of color misjudgment of bright color dots and light spot expansion interference.

[0024] Judgment results highly consistent with human visual perception: A dynamic threshold database based on physical limit samples was established, taking into account the differences in human visual perception of brightness of different colors. The detection results were consistent with human visual perception (Pearson correlation coefficient) of 0.92, which significantly reduced the systematic misjudgment caused by traditional single threshold.

[0025] High value for engineering applications: The detection method is applicable to self-emissive displays such as OLED and Micro-LED, with detection accuracy down to the micrometer level, which can meet the strict requirements of factory quality inspection. The detection efficiency is more than 30% higher than that of traditional methods, and it has broad prospects for industrial application.

[0026] To more clearly illustrate the structural features and effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0027] Figure 1 This is a flowchart of the algorithm of the present invention; Figure 2 This is a partial image taken by the black and white camera in this invention; Figure 3 This is a partial image of an image captured by the color camera in this invention; Figure 4 This is a schematic diagram illustrating the decomposition of an RGB color image into an HSV image in this invention; Figure 5 This is a schematic diagram of RGB sub-pixel region segmentation in this invention; Figure 6 This is a schematic diagram illustrating the determination of the color of the colored dots in this invention; Figure 7 This is a schematic diagram of defect extraction from black and white images in this invention. Detailed Implementation

[0028] The present invention will now be further described in conjunction with the accompanying drawings and relevant knowledge, and will be described clearly and completely. Obviously, the described applications are only some embodiments of the present invention, and not all embodiments.

[0029] This invention provides a method for detecting and judging sub-pixel defects based on dual-camera collaboration and multi-channel color segmentation. Addressing the bottleneck in identifying bright spot defects in sub-pixels at low grayscale (L0) levels of displays, it overcomes industry limitations through four core technologies: First, the physical contradictions of the single-camera solution are resolved through a dual-camera collaborative acquisition mechanism. The high-resolution color camera is synchronously controlled to use short exposure to freeze the original form of defects, while the high-sensitivity monochrome camera uses long exposure to capture the true brightness signal. The hardware synchronization timing error is ≤1ms, which enables the color camera to make up for the color recognition deficiency of the monochrome camera, and the monochrome camera to compensate for the insufficient low-light sensitivity of the color camera, thus achieving data complementarity. Secondly, precise positioning is achieved by using sub-pixel-level background modeling. Based on the physical characteristics of weak light emission of background sub-pixels under L0 grayscale, multi-channel fusion and segmentation of RGB channel and HSV space are performed on the defect-free screen to generate a background sub-pixel template with an accuracy of 1.5μm, which solves the positional deviation caused by ignoring background light emission in traditional methods. Furthermore, a dual-exposure fusion technology was designed to solve the problem of misjudging overexposure of bright color points—extracting the original color information from the short exposure image of the color camera, combining it with the background sub-pixel template to determine the position of the luminous sub-pixel, and combining the results of the RGB differential enhancement model to eliminate the interference of center whitening caused by channel saturation; Finally, a dynamic judgment logic driven by physical limit samples is constructed—based on the multi-channel segmentation results, the defect color type is classified, and the color-specific brightness threshold calibrated by actual measurement using standard defect samples (limit samples) in the pre-stored database is called, completely replacing the traditional single global threshold scheme and solving the systemic misjudgment caused by differences in human visual perception: This invention uses a dual-camera collaborative acquisition, multi-channel feature reconstruction, and human eye perception alignment as its core chain: First, it overcomes the physical contradiction of a single camera's inability to balance color and sensitivity by achieving dual-camera hardware synchronization with ≤1ms accuracy (short exposure for color to freeze the original color, and long exposure for black and white to capture the true brightness); second, based on the weak luminescence characteristics of L0 background sub-pixels, it integrates RGB / HSV multi-channel segmentation to generate a 1.5μm precision background template, providing a sub-pixel-level positioning benchmark for defects; furthermore, to address the problem of overexposure and whitening, it uses the template to lock the luminescence position and an RGB differential enhancement model to eliminate color misjudgment and dilation interference; finally, it calls a biological perception threshold library calibrated by physical limit samples based on color classification, achieving a leap from engineering experience to visual physiology in judgment; Specifically, refer to Figures 1-7 As shown, a method for sub-pixel defect detection and judgment based on dual-camera collaboration and multi-channel color segmentation includes the following steps: Step S1: Dual-camera collaborative acquisition: An acquisition system consisting of a main camera (a high-sensitivity monochrome camera) and an auxiliary camera (a high-resolution color camera) synchronously acquires images of the display panel at low grayscale (L0 level). The main camera uses a long exposure mode to capture the brightness signal, while the auxiliary camera uses a short exposure mode to freeze the original color information. The hardware synchronization timing error of the two cameras is ≤1ms, and the optical path consistency is achieved by sharing the same optical axis through a beam splitter prism, with a beam splitting ratio of 50:50. The long exposure mode of the monochrome main camera compensates for the low photon capture efficiency of the color camera at low brightness (L0 level), improving the brightness detection signal-to-noise ratio from less than 5dB in traditional color cameras to more than 15dB. The short exposure mode of the color auxiliary camera avoids channel saturation and center whitening caused by high exposure, completely preserving the original color information of defects, achieving a dual breakthrough in brightness sensitivity and color integrity.

[0030] Ensure data spatiotemporal consistency: The hardware synchronization timing error of ≤1ms and the optical path design with a 50:50 split ratio ensure that the defect images acquired by the dual cameras correspond precisely in spatial location (complete field of view matching) and temporal dimension (same detection time), providing an unbiased data foundation for subsequent color judgment and brightness measurement fusion analysis, and avoiding detection errors caused by spatiotemporal misalignment.

[0031] Step S2: Defect localization and sub-pixel-level background modeling: The color image acquired by the auxiliary camera is converted to HSV space. In the V channel, candidate defect regions are initially extracted through local contrast enhancement and adaptive threshold segmentation. In the H channel, red, green and blue sub-pixel space templates are constructed by analyzing the periodic distribution of hue values. Combined with the RGB channel and HSV space multi-channel fusion segmentation results of the defect-free screen, a background sub-pixel template with an accuracy of 1.5μm is generated to achieve sub-pixel-level defect localization. It can eliminate the interference of brightness and color coupling: HSV space conversion separates the brightness (V channel) and color (H / S channel) processing. The local contrast enhancement and adaptive threshold segmentation of the V channel can focus on the abnormal brightness features, avoiding the interference of color difference on defect extraction in the traditional RGB space, and improving the detection rate of weak bright spots of L0 level to more than 99%.

[0032] More specifically, it achieves micron-level precise positioning: Based on the periodic distribution of H-channel hue, a sub-pixel space template is constructed, combined with a 1.5μm precision background template generated by multi-channel fusion segmentation. This can accurately pinpoint the physical positions of red, green, and blue sub-pixels, reducing the defect positioning deviation from more than 10μm in traditional methods to within 2μm. The positioning accuracy is improved by 80%, meeting the micron-level precision requirements for sub-pixel-level detection.

[0033] Step S3: Multi-channel color determination: Perform RGB differential enhancement processing on the color image acquired by the auxiliary camera to obtain enhancement masks corresponding to the red, green and blue channels. Perform spatial verification on the enhancement masks and the sub-pixel space template of the H channel. Combine the hue features of the HSV channel to jointly determine the defect color type. The color type includes red (R), green (G), blue (B) and white (W). It can enhance the recognition of weak color signals: RGB differential enhancement compares the brightness difference between channels (e.g., for red defects, the difference between the red channel and the green / blue channel must be greater than the empirical threshold), separates and amplifies the weak defect signal of a specific color from the complex background, and improves the signal-to-noise ratio of color defects under L0 level low brightness by 3-5 times.

[0034] It can further eliminate overexposure whitening misjudgment: Spatial verification (enhanced mask and sub-pixel template intersection operation) ensures that defects are only judged as signals within the corresponding color sub-pixel area. Combined with the joint judgment of HSV hue features, it completely solves the color misjudgment problem caused by overexposure whitening of bright colored dots in traditional methods. The color dot recognition accuracy is improved from 78% to 97.2%, and the overexposure whitening misjudgment elimination rate reaches 92.5%. At the same time, by using the logic of judging white dots with three consecutive colored dots, white dot defects are accurately identified, avoiding misclassification of multi-color defects.

[0035] Step S4: Defect Extraction and Brightness Measurement in Black and White Images: The black and white image captured by the main camera is first processed by median filtering or bilateral filtering, and then the defect area is extracted. The average brightness value (MeanValue) and the average grayscale value (GrayMax100) of the brightest 100 pixels in the defect area are calculated. This ensures the accuracy of brightness measurement. Median filtering or bilateral filtering can suppress random noise (such as circuit noise and ambient light interference) while preserving the details of the defect edges, avoiding the brightness distortion caused by traditional filtering, and keeping the brightness measurement error within 3%. It also comprehensively characterizes the brightness characteristics of the defect. The average brightness value (MeanValue) reflects the overall brightness level of the defect, and the average grayscale value (GrayMax100) of the brightest 100 pixels captures the peak brightness characteristics of the defect. The combination of the two can completely describe the brightness distribution of the defect, providing richer and more accurate brightness basis for subsequent judgment, and avoiding judgment bias caused by a single brightness index (such as using only average brightness, which may easily overlook defects with excessive peak brightness).

[0036] Step S5: Dynamic Defect Judgment: Based on the defect color type determined in Step S3, the pre-stored physical limit sample calibration threshold database is called, and a comprehensive judgment score is calculated using a preset formula. The defect judgment result (OK / NG) is output based on the comparison between the score and the preset judgment threshold (τ). This method solves the systematic misjudgment inherent in traditional single-threshold methods. The physical limit sample calibration threshold database sets specific brightness thresholds for red, green, blue, and white (e.g., the human eye is more sensitive to blue light, so the blue defect threshold is lower), fully considering the differences in human eye perception of brightness for different colors. This results in a Pearson correlation coefficient of 0.92, improving the consistency of judgment by more than 60% compared to traditional single-threshold methods (e.g., a fixed 120 grayscale value). It also achieves objective and standardized judgment. The comprehensive judgment score is calculated through weighted average (integrating average brightness and peak brightness deviation), avoiding subjective errors from manual judgment. The preset judgment threshold (τ) can be flexibly adjusted according to product quality standards to meet the testing needs of different manufacturers, achieving standardized and engineered application of the testing process. The judgment efficiency is more than 5 times higher than manual judgment.

[0037] In a preferred embodiment of the present invention, the main camera is configured with the following parameters: resolution 4096×3000 (12 Megapixels), field of view (FOV) 14.1mm×10.3mm, imaging accuracy 3μm / pixel, exposure time 5ms, and output bit depth 12bit; the auxiliary camera is configured with the following parameters: resolution 4096×3000 (12 Megapixels), pixel size 3.45μm×3.45μm, equipped with an RK-TC5M230-10150 telecentric lens, magnification 1×, field of view (FOV) 14.1mm×10.3mm, imaging accuracy 3μm / pixel, exposure time 0.5ms, and output bit depth 12bit. To ensure a balance between imaging quality and inspection efficiency, the high resolution of 12 megapixels and the imaging accuracy of 3μm / pixel can clearly capture micron-level sub-pixel defects; the combination of 5ms long exposure of the main camera and 0.5ms short exposure of the auxiliary camera ensures brightness sensitivity and color integrity while controlling the single field of view inspection time to within 10ms, meeting the high-speed requirements of display panel factory inspection (≥360 panels can be inspected per hour).

[0038] Telecentric lenses eliminate imaging distortion. The RK-TC5M230-10150 telecentric lens, with its 1× magnification and low distortion characteristics (distortion rate <0.1%), ensures consistent imaging accuracy between the image edges and the center, avoiding edge sub-pixel defect positioning deviations caused by lens distortion, and further improving detection uniformity.

[0039] In a preferred embodiment of the present invention, in step S2, to avoid interference from color changes on brightness detection, the color image I is...rgb The formula for converting a color image to HSV space using the formula [R,G,B] is as follows: Calculate the maximum value C. max =max(R,G,B), minimum value C min =min(R,G,B) and the difference Δ=C max -C min ; Calculate hue H: ; The unit of hue H is usually angle (°), and the value range is [0, 360]. Calculate the saturation S: ; Calculate the brightness V: V = Cmax, where R, G, and B are the pixel brightness values ​​of the red, green, and blue channels of the color image, respectively. By accurately quantifying color and brightness characteristics, the differential calculation of hue H (e.g., calculated as (GB) / Δ when Cmax = R) can accurately distinguish the periodic distribution of red, green, and blue, providing a quantitative basis for sub-pixel template construction; saturation S reflects color purity and can help eliminate grayscale noise (S = 0 indicates a grayscale signal); brightness V directly corresponds to pixel brightness, providing a precise quantitative indicator for defect brightness extraction, enabling feature extraction accuracy in the HSV space to reach over 98%.

[0040] In a preferred embodiment of the present invention, step S3 first performs RGB differential enhancement on the image captured by the color camera. The purpose is to separate and enhance weak, specific color defect signals from the complex background in the color image. For red defect enhancement, the specific formula for RGB differential enhancement is: Red defect enhancement mask: Among them, I R (x,y), I G (x,y), I B (x, y) represent the pixel brightness values ​​of the red, green, and blue channels at the image coordinates (x, y), after being captured and decomposed by the color camera. and This is an empirical threshold; only when the difference is greater than this threshold is the red component considered significantly stronger than other components. ∧ represents a logical AND operation, requiring... and Both conditions must be met simultaneously; similarly, we obtain the green M. G (x,y) and blue M B Enhancement mask for (x,y), Green defect enhancement mask M G (x,y)=[I G (x,y)-I R (x,y)>τ gr ]∧[I G(x,y)-I B (x,y)>τ gb Blue defect enhancement mask M B (x,y)=[I B (x,y)-I R (x,y)>τ br ]∧[I B (x,y)-I G (x,y)>τ bg ], where τ gr τ gb τ br τ bg All thresholds are empirical. This method enables precise separation of color defects. Through dual-channel difference comparison and logical AND operation, it retains only areas where the target color channel brightness is significantly higher than the other two channels, effectively eliminating cross-color interference (e.g., red defects will not be falsely triggered by weak signals from the green / blue channels). This results in a single color defect extraction purity of over 95%, providing high-purity candidate regions for subsequent color determination.

[0041] Furthermore, the specific logic of the multi-channel joint determination in step S3 is as follows: the mask result after RGB difference enhancement is combined with the sub-pixel template for spatial verification; Spatial verification: D R =M R ∩T R D G =M G ∩T G D B =M B ∩T B M R M G M B These are the masking results after RGB differential enhancement; T R ,T G T B These are the red, blue, and green sub-pixel space templates on the hue map, which precisely mark the physical positions of all red, blue, and green sub-pixels on the screen. ∩ represents the intersection. Color confirmation requires that a point must meet two conditions simultaneously to be output to the final result: if the RGB difference enhancement mask result is consistent with the main color corresponding to the H channel, then the defect color is confirmed as R, G, or B; if three consecutive colored dots of different colors are detected, then it is determined to be a white dot.

[0042] Dual verification enhances the reliability of color determination. Spatial verification ensures that defects are located only within the physical area of ​​the corresponding color sub-pixel, while H-channel primary color verification further confirms color attributes. Dual verification reduces the color determination error rate to below 2.8%. The logic for determining white points using three consecutive color points can accurately identify joint defects of multiple sub-pixels (such as red, green, and blue sub-pixels being adjacent and all abnormal), avoiding white points being misjudged as single color defects. The white point recognition accuracy rate reaches over 96%.

[0043] Furthermore, in step S4, to ensure the stability and robustness of defect detection, median filtering or bilateral filtering is performed on the original image before it enters the detection algorithm to suppress random noise while preserving edge details. The filtering formula is: I f (x,y)=medianfiter(I C (x,y),k), where I c (x, y) represents the pixel values ​​of the original black and white image, k is the filter kernel size, and I f (x,y) represents the pixel values ​​of the filtered image. Median filtering achieves a suppression rate of over 90% for impulse noise (such as random bright spot noise), while bilateral filtering can further preserve the details of defect edges (edge ​​sharpness loss <5%), avoiding the defect blurring caused by traditional mean filtering. This provides low-noise, high-detail image data for subsequent brightness measurements, keeping the brightness measurement error within 3%.

[0044] Furthermore, the formula for calculating the average brightness value in step S4 is as follows: , where Ω is the defect area, N is the number of pixels in the area, and I(x,y) is the pixel brightness value of the defect area.

[0045] Furthermore, the formula for calculating the average grayscale value of the brightest 100 pixels in step S4 is as follows: The Top100 (Ω) represents the 100 brightest pixels in the defect area, sorted by brightness in descending order. This comprehensively reflects the brightness state of the defect. The average brightness value avoids misjudging the overall defect by local bright spots (e.g., if the edge of the defect is low in brightness but the center is bright, using only the peak brightness can easily overestimate the severity of the defect). The average grayscale value of the 100 brightest pixels can capture the peak brightness characteristics of the defect (avoiding the problem of the average brightness masking the problem of peak brightness exceeding the standard). The combination of the two makes the integrity of the brightness feature representation reach more than 99%, providing a more comprehensive brightness basis for judgment.

[0046] Furthermore, in step S5, the dynamic threshold Tc calibrated by the limit sample is called according to the color type (R / G / B / W), and the deviation score (comprehensive judgment score) is calculated: The formula for calculating the comprehensive judgment score is: Where W1 and W2 are weighting coefficients; TC,mean This is the reference value for the average grayscale of the limit sample corresponding to the color type (c=R / G / B / W); T C,max The grayscale reference value is the brightest 100 pixels in the limit sample for the corresponding color type; the judgment criterion is: , To determine the threshold, products exceeding the threshold are considered NG (Not Good) products, and products below the threshold are considered OK (Good) products.

[0047] It can achieve personalized and precise judgment. The weight coefficients W1 and W2 can be adjusted according to product needs (such as increasing W2 for products sensitive to peak brightness) to meet the detection standards of different scenarios. The threshold based on the limit sample ensures that the judgment benchmark is fully matched with the actual product quality requirements, avoiding the disconnect between theoretical thresholds and engineering practice.

[0048] Furthermore, the method is applicable to the factory quality inspection of OLED and Micro-LED self-emissive displays, focusing on the detection and judgment of micron-level sub-pixel bright spot defects under low grayscale conditions.

[0049] Compared with existing display panel defect detection methods, this invention, under low grayscale (L0 level) conditions, forms a complete high-sensitivity detection chain through dual-camera collaboration, multi-channel segmentation, and dynamic judgment, achieving simultaneous and accurate determination of brightness and color, and achieving the following significant technical effects: Complementary Acquisition of Brightness and Color Signals: Traditional single-camera systems are limited by their physical structure. Color cameras suffer from low signal-to-noise ratios in low light due to the filtering loss of the Bayer filter array; while monochrome cameras are sensitive, they lack color dimension. This invention achieves parallel acquisition of brightness and color signals through a dual-camera collaborative acquisition mechanism, employing a time-complementary mode of short color exposure and long monochrome exposure. Through hardware synchronization of ≤1ms and sub-pixel-level registration, it effectively overcomes the detection blind spot of single cameras under low-light conditions, achieving physical complementarity between color and brightness.

[0050] High-precision positioning of sub-pixel-level background modeling: Traditional algorithms are usually based on global thresholds or screen-wide brightness differences, which cannot consider the non-uniformity of weak emission of sub-pixels in the L0 state, resulting in positional deviations of more than 10μm. This invention establishes a background template by fusing statistical features of RGB and HSV six channels, achieving a resolution of 1.5μm, realizing sub-pixel-level brightness difference modeling, and making subsequent threshold determination highly stable.

[0051] Multi-channel color dot determination: Compared with existing display panel defect detection technologies, the multi-channel color dot determination method proposed in this invention achieves a significant improvement in color recognition accuracy and reliability under low grayscale (L0) conditions through the technical path of "RGB differential enhancement - spatial template verification - multi-feature fusion". Under L0 low grayscale conditions, the color dot recognition accuracy is increased from 78% of the traditional method to 97.2%, and the overexposure whitening misjudgment elimination rate reaches 92.5%.

[0052] Limit Sample-Driven Dynamic Threshold Determination: Traditional detection methods use a uniform brightness threshold (e.g., 120 grayscale) for determination, which has drawbacks, as it does not consider the differences in human eye sensitivity to brightness of different colors, leading to systematic bias. This invention establishes a dynamic threshold database based on physical limit samples, where the brightness perception threshold of each color sub-pixel is independently calibrated. The detection results show a consistency of 0.92 with human visual perception (Pearson correlation coefficient).

[0053] The present invention also provides a sub-pixel defect detection and judgment system based on dual-camera collaboration and multi-channel color segmentation, comprising: a dual-camera collaborative acquisition system, consisting of a main camera (black and white) and an auxiliary camera (color); An image processing unit, configured as follows: Defect localization: The color image from the auxiliary camera is converted to HSV space, defects are initially extracted in the V channel, and a sub-pixel template is constructed in the H channel to achieve sub-pixel level defect localization; Color determination: By combining the brightness comparison of the RGB channels with the hue information of the H channel, the defect color type (R / G / B / W) is determined. Brightness measurement: Utilizing the high-sensitivity black and white image of the main camera, the gray value of the defect area is measured; Defect determination: Based on the measured feature values ​​and the physical limit sample, a dynamic comparison is performed to output the final defect determination result.

[0054] Furthermore, in the defect localization step, the spatial arrangement of red, green, and blue sub-pixels is identified and located by analyzing the periodic distribution of H (hue) channel values; in the color determination step, if the color initially determined by the RGB channel is consistent with the main color of the H channel, the color type is confirmed; if three consecutive colored dots are detected, they are determined to be white dots (W); brightness measurement includes calculating the average brightness value (MeanValue) of the defect area and the average grayscale value (GrayMax100) of the brightest 100 pixels; defect determination is achieved by calculating a comprehensive determination score.

[0055] This invention achieves accurate color determination of colored dots and eliminates misjudgments caused by overexposure and whitening through a technical path combining RGB differential enhancement, spatial template verification, and HSV feature joint analysis. Specifically, for RGB differential enhancement, differential enhancement algorithms are designed for red, green, and blue color defects. By comparing the brightness values ​​between channels, defect signals of specific colors are separated and enhanced from complex backgrounds. For example, red defect enhancement calculates the brightness difference between the red channel and the green and blue channels. When both differences are greater than the corresponding empirical thresholds, a candidate region for a red defect is identified. Similarly, enhancement masks for green and blue defects are obtained. Spatial template verification ensures that the determined defect region is accurately located at the sub-pixel position of the corresponding color, eliminating background interference. HSV feature joint determination: if the results of RGB differential enhancement and spatial template verification are consistent and match the primary color corresponding to the H channel, the defect color is confirmed as R, G, or B. If three consecutive colored dots of different colors are detected clustered together, it is determined to be a white dot (W), completely solving the color misjudgment problem caused by overexposure and whitening of high-brightness colored dots.

[0056] This invention describes a dynamic defect judgment process that utilizes a physical limit sample-driven dynamic judgment system to achieve defect judgment aligned with human visual perception. Specifically, this involves establishing a threshold database: Through actual measurement and calibration of standard defect samples (limit samples), a dedicated threshold database for four colors—red, green, blue, and white—is established, storing the average grayscale reference value and the grayscale reference value of the brightest 100 pixels for each color, replacing the traditional single global threshold. A comprehensive judgment score is calculated by calling the corresponding threshold parameters based on the defect color type. This comprehensive judgment score can be adjusted according to engineering experience or visual perception characteristics to match brightness features with visual perception. Finally, a judgment threshold τ is set. When Score > τ, the product is judged as non-compliant (NG); when Score ≤ τ, the product is judged as compliant (OK), ensuring the objectivity and accuracy of the judgment results.

[0057] Example 1, see reference 2- Figure 3 As shown, for capturing high-definition images, this invention employs a dual-camera collaborative acquisition system, consisting of a main camera (high-sensitivity monochrome channel) and an auxiliary camera (high-resolution color channel). The two cameras share the same optical axis through a beam splitter to achieve optical path consistency, with a beam splitting ratio of 50:50.

[0058] (1) Main Camera Model: MV-CH120-10GM (Black and White Camera) Resolution: 4096×3000 (12 Megapixels) Field of view (FOV): 14.1mm × 10.3mm Imaging accuracy: 3μm / pixel Exposure time: 5ms (long exposure, used to enhance details in low light) Output bit depth: 12 bits Functional positioning: Responsible for collecting weak brightness signals of the display panel at the L0 gray level to achieve high-sensitivity brightness capture.

[0059] (2) Auxiliary camera (ColorCamera) Model: MV-CH120-10GC (Color Camera) Resolution: 4096×3000 (12 Megapixels) Pixel size: 3.45μm × 3.45μm Lens Model: RK-TC5M230-10150 Telecentric Lens Magnification: 1× Field of view (FOV): 14.1mm × 10.3mm Imaging accuracy: 3μm / pixel Exposure time: 0.5ms (short exposure, used to freeze true color information) Output bit depth: 12 bits Functional role: Responsible for capturing the original color features of defective areas and assisting the main camera in color reconstruction.

[0060] Example 2: Refer to Figure 4 , Figure 5 As shown, this embodiment establishes a high-precision background modeling method for sub-pixel-level defect localization based on multi-channel feature decomposition and reconstruction of color images. This method fully utilizes the characteristic of brightness and color separation in the HSV color space to achieve highly robust defect extraction and sub-pixel position calibration at low grayscale levels.

[0061] First, the original RGB image captured by the color camera in the dual-camera system is converted to a color space to obtain the corresponding HSV image. In the HSV model, Hue reflects the dominant color distribution of the pixel, Saturation represents the color purity, and Value describes the light intensity information of the pixel.

[0062] Preliminary extraction of candidate defect regions: Preliminary defect extraction is performed on the converted V (luminance) channel image. Since the HSV color space effectively separates luminance and color information, V channel processing can focus on luminance anomalies, thus avoiding detection interference caused by color differences or sub-pixel arrangement structures. The system uses local contrast enhancement and adaptive threshold segmentation algorithms to accurately extract faint bright spots or dark patches from the V channel, obtaining preliminary defect candidate regions.

[0063] Sub-pixel space template construction based on hue channel: In the Hue channel, sub-pixel structure segmentation is performed on the image. By analyzing the periodic distribution of Hue values, the spatial arrangement of red, green, and blue sub-pixels on the screen can be identified. Combined with the sub-pixel distribution template established in the previous defect-free sample, the luminous units of each channel can be accurately located and numbered in the current detection image.

[0064] Finally, the defect region extracted by the V channel is superimposed and compared with the sub-pixel localization results of the Hue channel to determine the specific sub-pixel type (R, G, or B) where the defect is located, and to generate defect coordinate information with micrometer-level accuracy. This method effectively solves the false detection problem caused by brightness and color coupling in the traditional RGB space, and achieves stable and accurate sub-pixel-level defect localization of high-resolution displays in low grayscale conditions.

[0065] Example 3: Reference Figure 6 As shown, this embodiment provides a multi-channel segmentation and spatial localization method based on color images for accurately determining the color type of red (R), green (G), blue (B), and white (W) sub-pixel highlights. This method achieves efficient and accurate highlight classification by fusing the periodic structure information of the hue (H) channel with the relative intensity relationship of the RGB channels.

[0066] The steps are as follows: 1. RGB Differential Enhancement: This technique decomposes the raw image captured by a color camera into three independent channels: R, G, and B. Ideally, a single-color point region will have a channel value significantly higher than the others. By differentiating these three channels, the defect regions corresponding to the red, green, and blue dots in the color image can be identified.

[0067] For each candidate defect region, calculate the brightness value of each channel, and find the channel with the highest brightness among the R / G / B components, which will serve as the preliminary color determination basis for the colored dots.

[0068] 2. HSV channel-assisted determination: Based on the preliminary extraction of candidate defect areas obtained in Example 2 and the sub-pixel space template constructed in the hue map.

[0069] 3. Multi-channel joint judgment: The color point results of the RGB channel are combined with the HSV hue features to form the color point judgment rule; if the RGB color point color is consistent with the main color corresponding to hue H, the color point color type is confirmed to be R, G or B; if three consecutive color points appear, the color is determined to be white.

[0070] Example 4: Reference Figure 7As shown, this embodiment provides a method for accurate measurement of defect brightness based on high-sensitivity images from a black and white camera. This method leverages the high signal-to-noise ratio of a black and white camera under low-light conditions to quantize the grayscale values ​​of the located defect areas, providing accurate brightness data for subsequent defect determination.

[0071] Example 5: Feature calculation and spec determination. Based on the aforementioned sub-pixel level defect extraction and color point determination results, this example performs quantitative feature calculation on each candidate defect and makes a determination in conjunction with the standard limit sample (Spec) to achieve the engineering determination and classification of micron-level defects.

[0072] Defect feature calculation: The following key features are calculated for each sub-pixel defect region: Mean Value: Using the results of Example 4, the average brightness of the effective pixels around the defect center is obtained; GrayMax100: Using the results of Example 4, calculate the average grayscale value of the 100 brightest pixels in the defect area; Color type: Based on the color determination of the colored dots in Example 3, the defective sub-pixel type (R / G / B / W) is determined. For specification determination, this embodiment, based on the aforementioned color dot determination using a color camera and feature calculation using a monochrome camera, combined with physical limit samples (red, green, blue, and white dot limit sample products), proposes a dynamic determination method driven by physical limit samples. This method achieves accurate classification and color recognition of micron-level sub-pixel defects. The determination logic shifts from a traditional fixed threshold to a dynamic, visual physiological perception alignment method, significantly improving detection accuracy and reliability.

[0073] The characteristic values ​​of the defects are compared with the corresponding limit samples, and the comprehensive judgment score is calculated. The calculation formula is as follows: ; Where W1 and W2 are weighting coefficients, adjusted based on engineering experience or visual perception; T C,mean This is the limit sample average grayscale reference value for this color type (c=R / G / B / W); T C,max This is the reference grayscale value for the brightest 100 pixels in the limit sample for this color type.

[0074] The technical principles of the present invention have been described above with reference to specific embodiments, which are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection scope. Those skilled in the art can conceive of other specific embodiments of the present invention without creative effort, and these embodiments will all fall within the protection scope of the present invention.

Claims

1. A method for detecting and judging sub-pixel defects based on dual-camera collaboration and multi-channel color segmentation, characterized in that, Includes the following steps: Step S1: The acquisition system consisting of a main camera and an auxiliary camera synchronously acquires images of the display panel at low grayscale. The main camera uses a long exposure mode to capture the brightness signal, and the auxiliary camera uses a short exposure mode to freeze the original color information. Step S2: Convert the color image acquired by the auxiliary camera to the HSV space. In the V channel, extract the candidate defect region through local contrast enhancement and adaptive threshold segmentation. In the H channel, construct red, green and blue sub-pixel space templates by analyzing the periodic distribution of hue values. Step S3: Perform RGB differential enhancement processing on the color image acquired by the auxiliary camera to obtain enhancement masks corresponding to the red, green and blue channels. Perform spatial verification on the enhancement masks and the sub-pixel space template of the H channel. Combine the hue features of the HSV channel to jointly determine the defect color type. Step S4: First, perform median filtering or bilateral filtering on the black and white image captured by the main camera, then extract the defect area, and calculate the average brightness value and the average gray value of the brightest 100 pixels of the defect area. Step S5: Based on the determined defect color type, call the pre-stored physical limit sample calibration threshold database to obtain the average grayscale reference value of the limit sample corresponding to the defect color type, the grayscale reference value of the brightest 100 pixels in the limit sample, and calculate the comprehensive judgment score according to the average brightness value and the average grayscale value of the brightest 100 pixels in the defect area using a preset formula. Based on the comparison result between the comprehensive judgment score and the preset judgment threshold, output the defect judgment result.

2. The method according to claim 1, characterized in that, The specific formula for converting the color image to HSV space in step S2 is as follows: Calculate the maximum value C max =max(R,G,B), minimum value C min =min(R,G,B) and the difference Δ=C max -C min ; Calculate hue H: ; Calculate the saturation S: ; Calculate the lightness V: V=C max Where R, G, and B are the pixel brightness values ​​of the red, green, and blue channels of the color image, respectively.

3. The method according to claim 1, characterized in that, The specific formula for RGB differential enhancement in step S3 is as follows: Red defect enhancement mask: Among them, I R (x,y), I G (x,y), I B (x, y) represent the pixel brightness values ​​of the red, green, and blue channels at image coordinates (x, y), respectively. and is the empirical threshold, and ∧ is the logical AND operation.

4. The method according to claim 1, characterized in that, The spatial verification in step S3 specifically involves: D R =M R ∩T R D G =M G ∩T G D B =M B ∩T B M R M G M B These are the masking results after RGB differential enhancement; T R ,T G T B These are the red, blue, and green sub-pixel space templates on the hue map, with ∩ representing the intersection; Color confirmation: If the mask result of RGB differential enhancement is consistent with the main color corresponding to the H channel, then the defect color is confirmed to be R, G, or B; If three consecutive colored dots of different colors are detected, then it is determined to be a white dot.

5. The method according to claim 1, characterized in that, The median filtering formula in step S4 is: I f (x,y)=medianfiter(I C (x,y),k), where I c (x, y) represents the pixel values ​​of the original black and white image, k is the filter kernel size, and I f (x,y) represents the pixel values ​​of the filtered image.

6. The method according to claim 1, characterized in that, The formula for calculating the average brightness value in step S4 is: , where Ω is the defect area, N is the number of pixels in the area, and I(x,y) is the pixel brightness value of the defect area.

7. The method according to claim 6, characterized in that, The formula for calculating the average grayscale value of the brightest 100 pixels in step S4 is: The Top100(Ω) represents the 100 brightest pixels in the defect area, sorted by brightness in descending order.

8. The method according to claim 7, characterized in that, The formula for calculating the comprehensive judgment score in step S5 is as follows: Where W1 and W2 are weighting coefficients; T C,mean This is the average grayscale reference value for the limit sample of the corresponding color type; T C,max The grayscale reference value is the brightest 100 pixels in the limit sample for the corresponding color type; the judgment criterion is: , To determine the threshold, products exceeding the threshold are considered unqualified, while those below the threshold are considered qualified. NG indicates unqualified, and OK indicates qualified.

9. A sub-pixel defect detection and judgment system based on dual-camera collaboration and multi-channel color segmentation, characterized in that, The system is used to perform the sub-pixel defect detection and judgment method based on dual-camera collaboration and multi-channel color segmentation as described in any one of claims 1-8, the system comprising: a dual-camera collaborative acquisition system, consisting of a main camera and an auxiliary camera; An image processing unit is configured to: convert the color image from the auxiliary camera to the HSV space, initially extract defects in the V channel, and construct a sub-pixel template in the H channel to achieve sub-pixel-level defect localization; fuse the brightness comparison of the RGB channels with the hue information of the H channel to determine the color type of the defect; use the high-sensitivity black and white image of the main camera to measure the gray value of the defect area; and dynamically compare the measured feature values ​​with the physical limit sample to output the final defect judgment result.

Images