A method and system for defect image detection in smart LCD screens
By using a multi-angle ring light source and orthogonal polarization imaging technology, combined with a deep convolutional network and a deformation convolutional attention module, a three-dimensional topography model is constructed, which solves the problems of specular reflection interference and misjudgment of defect classification in LCD screen inspection, and achieves high-precision and high-speed defect detection and classification.
Patent Information
- Application Number
- CN202611037420.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-13
- Publication Date
- 2026-11-10
AI Technical Summary
Existing LCD screen defect detection technologies suffer from problems such as specular reflection interference, insufficient detection accuracy, speed that cannot meet mass production requirements, and misjudgment of defect classification. They are unable to effectively identify micron-level defects and adapt to different types of defects.
A multi-angle ring light source combined with orthogonal polarization imaging is used to remove specular reflections through polarization difference and adaptive gamma correction, and a three-dimensional morphology model is constructed. Defect features are extracted by combining a deep convolutional network and a deformation convolutional attention module. A multi-task detection network is used to output defect segmentation, detection and classification in parallel, and defect grading is performed by combining three-dimensional morphology parameters.
It effectively removes specular reflection interference, improves the clarity of defect features and detection accuracy, adapts to the detection needs of different types of defects, reduces the probability of misjudgment, and improves the stability and applicability of the detection system.
Smart Images

Figure CN123071702A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of LCD screen defect image detection technology, and in particular to a defect image detection method and system for smart LCD screens. Background Technology
[0002] In the current manufacturing process of smart LCD screens, defect detection is a core step in controlling product yield. With the widespread adoption of high-resolution, high-refresh-rate LCD screens, the permissible defect size threshold continues to decrease, leading to increased demands on detection accuracy and efficiency. Traditional manual visual inspection relies on the experience and judgment of inspectors, which can easily cause visual fatigue during prolonged operation. This results in poor consistency, high false negative rates, and the detection speed cannot keep up with the pace of mass production lines. Therefore, it has been gradually replaced by machine vision inspection solutions. Existing ordinary 2D machine vision inspection solutions mostly use a single-angle light source combined with a common imaging module to acquire images. Due to specular reflection interference from the polarizing layer on the LCD screen surface, large areas of highlighting are easily generated during imaging, obscuring defect features. The acquisition effect is poor for low-contrast defects such as micron-level dents, scratches, and slight mura, making it highly prone to false negatives.
[0003] Some existing solutions attempt to introduce 3D shape detection technology to improve defect recognition capabilities. Commonly used laser triangulation scanning methods require scanning the screen line by line, resulting in slow detection speeds that cannot meet the demands of large-screen, high-rate mass production inspections. While structured light-based Fourier profilometry detection schemes offer faster imaging speeds, they are susceptible to localized high reflectivity on the screen, leading to phase unwrapping errors and artifacts in the generated shape data. This results in insufficient accuracy of the reconstructed 3D shape, failing to accurately reflect the actual depth information of the defects. At the defect recognition algorithm level, traditional convolutional neural networks have poor adaptability to irregularly shaped defects. For example, the boundary features of defects such as irregular scratches and edge chipping cannot be effectively aligned, easily leading to feature loss. Furthermore, general-purpose detection heads cannot simultaneously address the detection needs of defects with significant scale differences (points, lines, and surfaces), resulting in both false positives and false negatives.
[0004] Existing defect classification methods often rely solely on the planar geometry of defects, neglecting to consider their depth and surface roughness. However, in actual production, defects of the same planar size can have significantly different impacts on LCD screen display quality depending on their depth. Classifying defects based solely on planar size is prone to misjudgment, such as classifying acceptable shallow dents as defective or deep scratches that negatively affect display as acceptable. Furthermore, most existing detection systems lack dynamic compensation mechanisms for the light source and stage. With increased usage, issues such as light source aging and stage repetitive positioning misalignment lead to a gradual decrease in detection accuracy, requiring frequent manual calibration, increasing maintenance costs, and impacting the operational efficiency of mass production lines. Summary of the Invention
[0005] This invention proposes a defect image detection method and system for smart LCD screens to solve the problems mentioned in the prior art.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a defect image detection method for smart LCD screens, comprising the following steps: S1. A multi-angle ring light source sequentially projects striped structured light onto the LCD screen under test from a preset elevation angle, and simultaneously uses an orthogonal polarization imaging module to acquire orthogonal image pairs of polarization states at each angle to establish a multi-dimensional image dataset. S2. Perform polarization difference operation on each pair of polarization state orthogonal images to obtain a difference image with specular reflection removed, and then perform dark channel scattering suppression and adaptive gamma correction to obtain preprocessed difference images at each angle. S3. Register the preprocessed differential image at the same angle with the stripe projection pattern, extract the subpixel-level height topography map of the LCD screen surface using Fourier profilometry, and construct a three-dimensional topography model of the LCD screen surface by weighted fusion of the height topography maps at each angle. S4. The preprocessed differential images from each angle are spliced along the angle dimension and input into the pre-trained deep convolutional backbone network. The feature pyramid is used to extract point defect texture, line defect geometry and surface defect semantic features from multiple stages to construct a multi-scale defect feature map set. S5. Input the defect feature maps of each layer into the deformable convolutional attention module and the cross-scale fusion module set in series. The deformable convolutional attention module obtains the defect geometric offset through deformable convolution kernels. The cross-scale fusion module upsamples layer by layer and performs weighted summation, fusing high-level semantics and low-level textures to output the defect enhancement feature map. S6. Input the defect enhancement feature map into the multi-task defect detection network. The network has a point defect segmentation branch, a line defect detection branch, and a surface defect classification branch. Each branch shares the backbone weight and uses independent anchor points and detection heads. It outputs point defect segmentation masks, line defect vector trajectories, and surface defect bounding boxes and confidence scores in parallel. S7. Perform defect severity classification based on three-dimensional topography model for each branch output, calculate the defect hazard index by combining defect planar size and depth topography parameters, classify defects into three levels of qualified, acceptable and rejected according to the preset LCD screen quality standard, and generate a full-screen defect distribution quality map.
[0007] Furthermore, it also includes: in step S1, a spectroradiometer is used to pre-calibrate the spectral power distribution and spatial illuminance uniformity corresponding to each angle of the multi-angle ring light source, and a light source calibration parameter lookup table is established. When acquiring each batch of LCD screen images, radiometric correction is performed on each angle polarization image pair according to the lookup table. A laser displacement sensor is used to measure the spatial coordinates of the four corner reference marks of the LCD screen positioning stage, and the spatial tilt angle deviation between the stage plane and the imaging optical axis is calculated. When the tilt angle deviation exceeds the preset tolerance, the stage adjustment mechanism is driven to complete the attitude compensation and then re-acquire the image.
[0008] Furthermore, it also includes: in step S3, performing local weighted regression outlier filtering on each angle topography map, marking points where the residual in the local neighborhood of each height data point exceeds the threshold as measurement outliers, replacing outliers with the median height value of the neighborhood, performing iterative nearest point algorithm multi-view point cloud fine registration on each filtered angle topography map, using the first angle point cloud as the reference, aligning the remaining angle point clouds to the reference coordinate system through rigid body transformation, determining the weights based on the vertical distance of each point to the nearest neighbor patch of the reference point cloud after registration, and obtaining a three-dimensional topography model through weighted voxel fusion.
[0009] Furthermore, in step S5, the deformation convolutional attention module superimposes the two-dimensional deformation offset predicted by the additional convolutional layer onto the standard convolutional sampling grid for each sampling position of the input feature map. The sampling position of the convolutional kernel generates a spatial offset based on the defect morphology. Each offset sampling position generates attention weights based on the cosine similarity between the local features and the corresponding defect category prototype feature vector through nonlinear activation mapping. The additional convolutional layer, offset prediction branch, and attention branch adopt an end-to-end approach, and are jointly optimized based on a weighted combination of alignment loss and detection loss. The deformation offset learning process introduces defect segmentation ground value labeling constraints.
[0010] Furthermore, in step S6, the point defect segmentation branch performs defect morphology classification based on principal curvature analysis for each connected defect region: the height distribution surface equation is obtained by fitting a quadratic surface to the coordinates of all pixels within the defect region; the maximum and minimum principal curvatures of each pixel are calculated from this surface equation; and the ratio of the average maximum principal curvature to the average minimum principal curvature of all pixels within the region is used as the concavity / convexity discrimination index of the defect region. The concavity / convexity discrimination index is quantitatively calculated using the following formula: ; In the formula, Indicates the first The unevenness discrimination index of a point defect area This represents the total number of pixels within the defect area at that point. For pixel index; and They represent the first The absolute values of the maximum and minimum principal curvature of each pixel. This is a preset regularization constant.
[0011] Furthermore, the defect hazard index mentioned in step S7 is calculated by combining planar geometric parameters and depth morphology parameters. The planar geometric parameters include the diameter and aspect ratio of the defect's equivalent circle, and the depth morphology parameters include the maximum indentation depth or maximum protrusion height of the defect relative to the surrounding normal surface reference plane in the three-dimensional morphology model. For each defect output in step S6, the planar geometric parameters are extracted from the mask or bounding box, and the depth morphology parameters are extracted from the corresponding spatial coordinates indexed in the three-dimensional morphology model. Both are input into a multi-dimensional threshold classifier pre-trained based on the LCD screen defect acceptance criteria, and the qualified, acceptable, or rejected quality level corresponding to the defect is output. When the number of rejection-level defects is zero and the number of acceptable-level defects is less than the preset maximum acceptable number, the entire LCD screen is determined to be qualified.
[0012] Furthermore, the weighted fusion weights of the height topography images at each angle mentioned in step S3 are determined according to the following angle confidence weighting formula: ; In the formula, Indicates the first Each illumination angle corresponds to a normalized weight in the height topography map during fusion, and satisfies... , This is the attenuation control coefficient. Indicates the first The root mean square error of phase unwrapping at each angle. Indicates the first The angle between the principal direction of the light source and the surface normal. and They represent the first The root mean square error of each angle and the angle between the rays, The total number of lighting angles in the preset elevation angle set and .
[0013] Furthermore, a defect image detection system for smart LCD screens includes the following modules: The image acquisition subsystem is equipped with a multi-angle ring structure light source, an orthogonal polarization imaging module and an LCD screen positioning stage, which is used to acquire multi-dimensional polarization state orthogonal image pairs of the LCD screen under test by multi-angle structure light projection and orthogonal polarization synchronous acquisition. The image preprocessing subsystem is equipped with a polarization difference operation module, a dark channel scattering suppression module, and an adaptive gamma correction module, which are used to perform specular reflection removal and enhancement processing on polarization image pairs at various angles to generate preprocessed difference images. The three-dimensional topography reconstruction subsystem is equipped with a Fourier contour phase extraction module, an outlier filtering module, a multi-view point cloud registration module, and a weighted voxel fusion module, which are used to reconstruct the three-dimensional topography model of the LCD screen surface from structured light patterns and differential images. The defect feature extraction subsystem is equipped with a deep convolutional backbone network, a feature pyramid, and a deformation convolutional attention module, which is used to extract multi-scale defect features from preprocessed difference images and align the deformation features of irregular defects. The multi-task defect detection subsystem is configured with point defect segmentation branches, line defect detection branches, and surface defect classification branches, which are used to output point defect segmentation masks, line defect vector trajectories, and surface defect bounding boxes and confidence scores in parallel. The defect classification and assessment subsystem is equipped with a defect hazard index calculation module, a multidimensional threshold classifier, and a full-screen quality map generation module. It is used to classify defect quality levels by comprehensively considering planar geometric parameters and depth morphology parameters and generate a full-screen defect distribution quality map.
[0014] Furthermore, the multi-angle ring light source in the image acquisition subsystem consists of narrow-band LED beads arranged along the circumference. The pitch angle of each group of LED beads is independently adjusted within a preset pitch angle range by a stepper motor. The orthogonal polarization imaging module includes a linear polarizer and a quarter-wave plate, with the fast axis angle between the two fixed at 45 degrees. Laser reflection reference marks are embedded at the four corners of the LCD screen positioning platform. After the LCD screen under test is replaced, the laser displacement sensor detects the spatial position of the reference marks.
[0015] Furthermore, the multidimensional threshold classifier of the defect grading and evaluation subsystem consists of three cascaded binary support vector machines. The first-level support vector machine uses the equivalent circle diameter and aspect ratio of the defect as input to distinguish between acceptable and unacceptable defects. The second-level support vector machine uses the ratio of the maximum indentation depth of the defect to the equivalent circle diameter, the ratio of the maximum protrusion height of the defect to the equivalent circle diameter, and the concavity / convexity discrimination index as input to distinguish between acceptable and unacceptable defects. The third-level support vector machine uses the defect hazard index and the position coordinates of the defect within the LCD display area as input to determine the rework feasibility of unacceptable defects.
[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention employs a multi-angle ring structured light combined with orthogonal polarization imaging acquisition method, which can effectively remove specular reflection interference from the LCD screen surface. The acquired multi-dimensional images can cover various defects with different orientations, avoiding the defect omission problem caused by single-angle imaging. Through preprocessing steps of polarization difference, dark channel scattering suppression, and adaptive gamma correction, the feature clarity of low-contrast defects can be improved, and the influence of ambient scattered light on image quality can be reduced, providing more reliable input data for subsequent morphology reconstruction and defect feature extraction.
[0017] The present invention constructs a three-dimensional topography model through weighted fusion of multi-view height topography maps, which can reduce the influence of phase unwrapping artifacts caused by local high reflection on modeling accuracy. The reconstructed three-dimensional topography can accurately reflect the height distribution on the surface of the screen, providing a topography-level basis for subsequent attribute judgment and classification of defects. The feature extraction structure combining feature pyramid with deformable convolution attention and cross-scale fusion can adapt to the morphology of irregular defects, achieve effective alignment of defect features, and simultaneously take into account the texture features of small-sized point defects and the semantic features of large-sized surface defects, so as to improve the accuracy of feature extraction for different types of defects.
[0018] The present invention adopts a multi-task parallel defect detection network structure. The shared backbone weights can reduce the overall calculation amount, and the independent detection heads can adapt to the detection requirements of three different types of defects: point, line and surface. The classification of the concave-convex attribute of point defects can distinguish structural defects from surface dirt defects, and reduce the probability of misjudgment of different types of defects. The defect classification method combined with three-dimensional topography parameters can better fit the quality requirements in the actual use of liquid crystal screens, reduce the misjudgment problem caused by classification only based on plane parameters. The generated full-screen defect distribution quality map can visually present the defect distribution of the screen, providing data support for subsequent adjustment of production processes.
[0019] The supporting detection system of the present invention is provided with a light source calibration and stage posture compensation mechanism, which can reduce the influence of light source aging and stage repeated positioning deviation on detection accuracy, improve the stability of long-term operation of the system, eliminate the need for frequent manual calibration, adapt to the detection requirements of liquid crystal screens of different sizes and different batches, and has a wider application range. Brief Description of the Drawings
[0020] Figure 1 is the overall flow chart of the defect image detection for smart liquid crystal screens proposed by the present invention; Figure 2 is the flow chart of image acquisition and preprocessing of the present invention; Figure 3 is the flow chart of three-dimensional topography reconstruction on the surface of the liquid crystal screen of the present invention; Figure 4 is the flow chart of defect feature extraction and multi-task detection of the present invention; Figure 5 is the flow chart of defect severity classification and evaluation of the present invention. Detailed Description of the Embodiments
[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by a person of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.
[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.
[0024] Reference Figures 1 to 5 A defect image detection method for smart LCD screens includes the following steps: S1. A multi-angle ring light source sequentially projects striped structured light onto the LCD screen under test from a preset elevation angle, and simultaneously uses an orthogonal polarization imaging module to acquire orthogonal image pairs of polarization states at each angle, and establishes a multi-dimensional image dataset with angle-polarization joint index. S2. Perform polarization difference operation on each pair of polarization state orthogonal images to obtain a difference image with specular reflection removed, and then perform dark channel scattering suppression and adaptive gamma correction to obtain preprocessed difference images at each angle. S3. Register the preprocessed differential image at the same angle with the stripe projection pattern, extract the subpixel-level height topography map of the LCD screen surface using Fourier profilometry, and construct a three-dimensional topography model of the LCD screen surface by weighted fusion of the height topography maps at each angle. S4. The preprocessed differential images from each angle are spliced along the angle dimension and input into the pre-trained deep convolutional backbone network. The feature pyramid is used to extract point defect texture, line defect geometry and surface defect semantic features from multiple stages to construct a multi-scale defect feature map set. S5. Input the defect feature maps of each layer into the concatenated deformation convolutional attention module and cross-scale fusion module. The deformation convolutional attention module learns the defect geometric offset through deformable convolutional kernels to align the irregular defect features. The cross-scale fusion module aggregates high-level semantics and low-level textures across scales through layer-by-layer upsampling and weighted summation, and outputs a defect enhancement feature map. S6. Input the defect enhancement feature map into the multi-task defect detection network. The network has a point defect segmentation branch, a line defect detection branch, and a surface defect classification branch. Each branch shares the backbone weight and uses independent anchor points and detection heads. It outputs point defect segmentation masks, line defect vector trajectories, and surface defect bounding boxes and confidence scores in parallel. S7. Perform defect severity classification based on three-dimensional topography model for each branch output, calculate the defect hazard index by combining defect planar size and depth topography parameters, classify defects into three levels of qualified, acceptable and rejected according to the preset LCD screen quality standard, and generate a full-screen defect distribution quality map.
[0025] The present invention further includes: in step S1, a multi-angle ring light source is pre-calibrated using a spectroradiometer to determine the spectral power distribution and spatial illuminance uniformity at each angle to establish a light source calibration parameter lookup table; when acquiring images of each batch of LCD screens, radiometric correction is performed on polarized images at each angle according to the lookup table to eliminate brightness deviations caused by light source aging and uneven illuminance between angles; simultaneously, a laser displacement sensor is used to measure the spatial coordinates of the four corner reference marks of the LCD screen positioning stage and calculate the spatial tilt angle deviation between the stage plane and the imaging optical axis; when the tilt angle deviation exceeds the preset tolerance, the stage adjustment mechanism is driven to perform attitude compensation and then the image is re-acquired.
[0026] The present invention further includes: in step S3, performing outlier filtering based on robust local weighted regression on the topography images of each angle, marking points whose residuals exceed the threshold in the local neighborhood of each height data point as measurement outliers and replacing them with the median height value of the neighborhood to eliminate phase unwrapping artifacts caused by local high reflectivity of the LCD screen; performing multi-view point cloud fine registration based on the iterative nearest point algorithm on the filtered topography images of each angle, aligning the subsequent angle point clouds to the reference coordinate system through rigid body transformation with the first angle point cloud as the reference, and performing weighted voxel fusion to generate a three-dimensional topography model after registration with the vertical distance from each point to the nearest neighbor patch of the reference point cloud as the weight.
[0027] In this invention, the deformation convolutional attention module in step S5 superimposes a two-dimensional deformation offset predicted by an additional convolutional layer on each sampling position of the input feature map based on the standard convolutional sampling grid. This causes the sampling position of the convolutional kernel to undergo adaptive spatial offset according to the actual shape of the defect. For each sampled position after deformation offset, attention weights are generated by nonlinear activation mapping based on the cosine similarity between its local features and the prototype feature vector of the defect category to which the position belongs. The additional convolutional layer, the offset prediction branch, and the attention branch are jointly optimized end-to-end through a weighted combination of alignment loss and detection loss so that the deformation offset learning is constrained by the ground truth label of defect segmentation, thereby enhancing the alignment effect of irregular defect boundary features.
[0028] In this invention, the point defect segmentation branch in step S6 performs defect morphology classification based on principal curvature analysis for each connected defect region: a height distribution surface equation is obtained by fitting a quadratic surface to the coordinates of all pixels within the defect region; the maximum and minimum principal curvatures of each pixel are calculated from this surface equation; and the ratio of the average maximum principal curvature to the average minimum principal curvature of all pixels within the region is used as the concavity / convexity discrimination index of the defect region; the concavity / convexity discrimination index is quantitatively calculated using the following formula: ; In the formula, Indicates the first The unevenness discrimination index for a single defect area is dimensionless. When the defect exceeds a preset protrusion threshold, the defect is classified as a protrusion-type defect. Defects smaller than the preset indentation threshold are classified as indented defects, while those between the two are classified as flat dirt defects. This represents the total number of pixels within the defect area at that point. For pixel index; and They represent the first The absolute values of the maximum and minimum principal curvature of a pixel are obtained by solving the Hessian matrix eigenvalues after performing second-order partial derivative operations on the height surface equation in the neighborhood of that pixel. The dimension is the reciprocal of the length unit. To predetermine the regularization constant, the dimension is consistent with the principal curvature (the reciprocal of the length unit) and >0 is used to prevent the denominator from being zero.
[0029] In this invention, the defect hazard index in step S7 is calculated by combining planar geometric parameters and depth morphology parameters. The planar geometric parameters include the diameter and aspect ratio of the defect's equivalent circle, and the depth morphology parameters include the maximum indentation depth or maximum protrusion height of the defect relative to the surrounding normal surface reference plane in the three-dimensional morphology model. For each defect output in step S6, the planar geometric parameters are extracted from the mask or bounding box, and the depth morphology parameters are extracted from the corresponding spatial coordinates indexed in the three-dimensional morphology model. Both are input into a multi-dimensional threshold classifier pre-trained based on the LCD screen defect acceptance criteria, and the qualified, acceptable, or rejected quality level corresponding to the defect is output. When the number of rejection-level defects is zero and the number of acceptable-level defects is less than the preset maximum acceptable number, the LCD screen is determined to be qualified.
[0030] In this invention, the weighted fusion weights of the height topography images at each angle in step S3 are determined according to the following angle confidence weighting formula: ; In the formula, Indicates the first Each illumination angle corresponds to a normalized weight in the height topography map during fusion, and satisfies... The smaller the angle between the light and the normal to the LCD screen surface, the higher the fusion weight is assigned, and the lower the interference from specular reflection. This is the attenuation control coefficient, with the dimension in radians. -1 and >0; Indicates the first The root mean square error of phase unwrapping at each angle, with the dimension in radians, is used to reflect the reliability of height measurement. Indicates the first The angle between the principal direction of the light source and the surface normal is measured in radians. and They represent the first The root mean square error of each angle and the angle between rays; The total number of lighting angles in the preset elevation angle set and .
[0031] This invention discloses a defect image detection system for smart LCD screens, comprising the following modules: The image acquisition subsystem is equipped with a multi-angle ring structure light source, an orthogonal polarization imaging module and an LCD screen positioning stage, which is used to acquire multi-dimensional polarization state orthogonal image pairs of the LCD screen under test by multi-angle structure light projection and orthogonal polarization synchronous acquisition. The image preprocessing subsystem is equipped with a polarization difference operation module, a dark channel scattering suppression module, and an adaptive gamma correction module, which are used to perform specular reflection removal and enhancement processing on polarization image pairs at various angles to generate preprocessed difference images. The three-dimensional topography reconstruction subsystem is equipped with a Fourier contour phase extraction module, an outlier filtering module, a multi-view point cloud registration module, and a weighted voxel fusion module, which are used to reconstruct the three-dimensional topography model of the LCD screen surface from structured light patterns and differential images. The defect feature extraction subsystem is equipped with a deep convolutional backbone network, a feature pyramid, and a deformation convolutional attention module, which is used to extract multi-scale defect features from preprocessed difference images and align the deformation features of irregular defects. The multi-task defect detection subsystem is configured with point defect segmentation branches, line defect detection branches, and surface defect classification branches, which are used to output point defect segmentation masks, line defect vector trajectories, and surface defect bounding boxes and confidence scores in parallel. The defect classification and assessment subsystem is equipped with a defect hazard index calculation module, a multidimensional threshold classifier, and a full-screen quality map generation module. It is used to classify defect quality levels by comprehensively considering planar geometric parameters and depth morphology parameters and generate a full-screen defect distribution quality map.
[0032] In this invention, the multi-angle ring structure light source in the image acquisition subsystem is composed of multiple sets of narrow-band LED beads evenly arranged along the circumference, and the pitch angle of each set of beads can be independently adjusted by a stepper motor within a preset pitch angle range. The orthogonal polarization imaging module includes a linear polarizer and a quarter-wave plate, and the fast axis angle between the two is fixed at 45 degrees to form a circular polarization imaging optical path to eliminate specular reflection interference on the surface of the liquid crystal screen. Each of the four corners of the liquid crystal screen positioning stage is embedded with a laser reflection reference mark, and after the liquid crystal screen under test is replaced, the spatial position of the reference mark is automatically detected by a laser displacement sensor to compensate for the stage repetitive positioning error.
[0033] In this invention, the multidimensional threshold classifier in the defect grading and evaluation subsystem consists of three cascaded binary support vector machines. The first-level support vector machine distinguishes between acceptable and unacceptable defects using the equivalent circle diameter and aspect ratio of the defect as input. The second-level support vector machine distinguishes unacceptable defects into acceptable and unacceptable defects using the ratio of the maximum indentation depth or maximum protrusion height of the defect to the equivalent circle diameter and the concavity / convexity discrimination index as input. The third-level support vector machine determines the rework feasibility of unacceptable defects using the defect hazard index and the position coordinates of the defect within the LCD screen display area as input. The classification hyperplane parameters of each level of support vector machine are fixed and deployed after being labeled and trained according to the LCD screen quality standards.
[0034] The specific embodiments of the present invention are further illustrated below: Example 1: This example is applied to the online inspection scenario of a mass production line for small-sized mobile phone smart LCD screens. It is deployed at the final inspection station after the LCD screen module assembly is completed, and the inspection cycle time is matched with the mass production line transmission cycle time. The inspection system consists of six components: an image acquisition subsystem, an image preprocessing subsystem, a 3D topography reconstruction subsystem, a defect feature extraction subsystem, a multi-task defect detection subsystem, and a defect grading and evaluation subsystem. The multi-angle ring structure light source of the image acquisition subsystem has 6 groups of narrow-band LED beads evenly arranged along the circumference. The pitch angle of each group of LED beads can be independently adjusted within the range of 15 degrees to 45 degrees by a stepper motor. The orthogonal polarization imaging module consists of a linear polarizer and a quarter-wave plate, with the fast axis angle between the two fixed at 45 degrees, forming a circular polarization imaging optical path. A laser reflection reference mark is embedded in each of the four corners of the LCD screen positioning stage, and the matching laser displacement sensor can complete the spatial coordinate measurement of the reference mark.
[0035] Before the testing process begins, the light source calibration and stage reference calibration are completed in advance. A spectroradiometer is used to collect data on the spectral power distribution and spatial illuminance uniformity of the light source at different angles, and a lookup table of light source calibration parameters is established. Before each batch of LCD screens begins testing, radiometric correction is performed on the acquired images according to the lookup table to eliminate brightness deviations caused by light source aging and uneven illuminance between angles. After each change of the stage or adjustment of the production line, the spatial coordinates of the four corner reference marks on the stage are measured using a laser displacement sensor. The spatial tilt angle deviation between the stage plane and the imaging optical axis is calculated. When the deviation exceeds the preset tolerance, the stage adjustment mechanism is driven to complete attitude compensation. After compensation is completed, the formal testing process is started.
[0036] During formal testing, the LCD screen to be tested is transferred to the positioning platform along the production line and fixed. A multi-angle ring light source sequentially projects sinusoidal striped structured light onto the LCD screen from six preset elevation angles. After each angle of structured light projection is completed, an orthogonal polarization imaging module is simultaneously triggered to acquire two sets of orthogonal polarization images at that angle. After all angles are acquired, a multi-dimensional image dataset with an angle-polarization joint index is established. The image preprocessing subsystem performs pixel-by-pixel polarization difference operations on each pair of orthogonal polarization images to remove specular reflection components from the LCD screen surface, obtaining a difference image. Then, dark channel scattering suppression processing is performed on the difference image to remove background noise from ambient scattered light. Finally, adaptive gamma correction is performed based on the global average grayscale of the image to improve the feature clarity of low-contrast defects, ultimately obtaining the preprocessed difference image corresponding to each angle.
[0037] The 3D topography reconstruction subsystem performs pixel-level registration between the preprocessed differential images at the same angle and the pre-stored stripe projection pattern, and extracts the sub-pixel-level height topography map of the LCD screen surface using Fourier profilometry. First, outlier filtering is performed on the height topography maps at each angle. A robust local weighted regression method is used to calculate the residual of each height data point in its local neighborhood. Points with residuals exceeding a preset threshold are marked as measurement outliers, and their measured values are replaced with the median height value of the neighborhood, eliminating phase unwrapping artifacts caused by local high reflectivity of the LCD screen. After filtering, the height topography maps at each angle are converted into 3D point cloud data. Using the point cloud at the first angle as a reference, an iterative nearest-point algorithm is used to complete the multi-view fine registration of the point clouds at the remaining angles, aligning all point clouds to the same reference coordinate system. After registration, fusion weights are assigned to each point cloud based on angle confidence. The smaller the root mean square error of phase unwrapping and the smaller the angle between the principal direction of the light source and the normal to the screen surface, the higher the weight of the corresponding point cloud. After weighted voxel fusion, a 3D topography model of the LCD screen surface is generated.
[0038] The defect feature extraction subsystem concatenates preprocessed differential images from all angles along the angular dimension into a multi-channel image, which is then input into a pre-trained deep convolutional backbone network. The backbone network employs a five-level downsampling residual network structure, and feature maps of different resolutions output from each stage are input into a feature pyramid structure to extract point defect texture, line defect geometry, and surface defect semantic features at corresponding scales, constructing a multi-scale defect feature map set. The defect feature maps from each layer are then input into a concatenated deformation convolutional attention module and a cross-scale fusion module. For each sampling position in the input feature map, the deformation convolutional attention module predicts a two-dimensional deformation offset based on a standard 3x3 convolutional sampling grid through two additional convolutional branches. This allows the sampling position of the convolutional kernel to adaptively shift spatially according to the actual shape of the defect. For each offset sampling position, the cosine similarity between its local features and the pre-defined defect category prototype feature vector is calculated, and after nonlinear activation, it is mapped to attention weights, which are then multiplied by the convolutional output to complete feature enhancement. The additional convolutional layers, offset prediction branch, and attention branch are optimized end-to-end during the training phase. The loss function is obtained by a weighted combination of feature alignment loss and detection loss. The learning of deformation offset is constrained by the ground truth annotation of defect segmentation, which improves the feature alignment effect of irregular defect boundaries. The cross-scale fusion module performs upsampling layer by layer starting from the semantic features of the highest layer, and performs weighted summation with the texture features of the next layer to complete the cross-scale aggregation of high-level semantics and low-level texture, and finally outputs a defect enhancement feature map that matches the resolution of the input image.
[0039] The multi-task defect detection subsystem inputs the enhanced feature map of defects into the multi-task detection network. The network includes a point defect segmentation branch, a line defect detection branch, and a surface defect classification branch. These three branches share the backbone network weights but each has its own independent anchor point parameters and detection head. The point defect segmentation branch uses a fully convolutional structure to output a pixel-level point defect segmentation mask. It performs principal curvature analysis on each connected defect region, obtaining the height distribution surface equation of the defect region through quadratic surface fitting. It calculates the principal curvature of each pixel within the region and determines the convexity / concave attribute of the defect based on the ratio of the region's average principal curvature, distinguishing between convex defects, concave defects, and flat contamination defects. The line defect detection branch uses an anchor-frame-based detection head and outputs the vector trajectory coordinates of line defects. The surface defect classification branch outputs the bounding box coordinates and corresponding confidence scores of surface defects.
[0040] The defect grading and evaluation subsystem performs defect severity grading on the detection results output by each branch. For each detected defect, it extracts planar geometric parameters such as the equivalent circle diameter and aspect ratio from the segmentation mask or bounding box, and extracts depth morphology parameters such as the maximum indentation depth or maximum protrusion height of the defect relative to the surrounding normal surface reference plane from the corresponding spatial coordinates of the 3D morphology model. These two types of parameters are input into a pre-trained and fixed multi-dimensional threshold classifier. The classifier consists of three cascaded binary support vector machines. The first-level support vector machine uses planar geometric parameters as input to distinguish between acceptable and unacceptable defects. The second-level support vector machine uses the ratio of defect height to equivalent circle diameter and the concavity / convexity discrimination index as input to distinguish unacceptable defects into acceptable and rejected defects. The third-level support vector machine uses the defect hazard index and the defect's position coordinates within the display area as input to determine the rework feasibility of rejected defects. After all defects are graded, when the number of rejected defects is zero and the number of acceptable defects is less than the preset maximum acceptable number, the LCD screen is deemed acceptable, and a full-screen defect distribution quality map is finally generated.
[0041] This embodiment is adapted to the online inspection requirements of small-size LCD screen mass production lines. It can effectively eliminate the interference of specular reflection on the screen surface. The acquired multi-dimensional images can cover tiny defects in different orientations. The accuracy of the reconstructed three-dimensional morphology model can meet the detection requirements of micron-level defects. The defect feature extraction structure can adapt to the morphology of various irregular defects. Multi-task parallel detection can take into account the detection accuracy of different types of defects. The grading method of morphology parameters can reduce the probability of misjudgment. The system's built-in calibration compensation mechanism can improve the stability of long-term operation. It does not require frequent manual calibration and is adapted to the continuous operation requirements of mass production scenarios.
[0042] Example 2: This example is applied to the offline sampling inspection of large-size smart TV LCD screens. Deployed at the sampling inspection station in the LCD screen production workshop, it can complete full-area defect inspection of LCD screens of 65 inches and above. The inspection system also includes six components: image acquisition subsystem, image preprocessing subsystem, 3D topography reconstruction subsystem, defect feature extraction subsystem, multi-task defect detection subsystem, and defect grading and evaluation subsystem. The multi-angle ring structure light source of the image acquisition subsystem has 12 groups of narrow-band LED beads evenly arranged along the circumference. The pitch angle of each group of LED beads can be independently adjusted within the range of 10 degrees to 50 degrees by a stepper motor to meet the edge area illumination requirements of large-size screens. The imaging resolution of the orthogonal polarization imaging module matches the full-area acquisition requirements of large-size screens. It is also composed of a linear polarizer and a quarter-wave plate, with the fast axis angle between the two fixed at 45 degrees. Each of the four corners of the large-size LCD screen positioning stage is embedded with a laser reflection reference mark, and the range of the matching laser displacement sensor is adapted to the spatial coordinate measurement requirements of the large-size stage.
[0043] Before the testing process begins, light source calibration and stage benchmark calibration are completed for large-scale scenes. A spectroradiometer is used to collect data on the spectral power distribution and spatial illuminance uniformity of the light source at different angles and in different screen areas. A lookup table of light source calibration parameters for each region is established. Before testing each batch of sampled LCD screens, radiometric correction is performed on the collected images of each region according to the lookup table to eliminate illuminance deviations in different areas of the large-scale screen and brightness deviations caused by light source aging. After each stage adjustment or screen size change, the spatial coordinates of the four corner reference marks on the stage are measured using a laser displacement sensor. The spatial tilt angle deviation between the stage plane and the imaging optical axis is calculated. When the deviation exceeds a preset tolerance, the stage adjustment mechanism is activated to complete attitude compensation. After compensation is completed, the formal testing process is started.
[0044] During formal testing, the LCD screen to be tested is placed on the positioning platform and fixed by vacuum adsorption. Due to the large size of the screen, it is divided into four acquisition areas for sequential acquisition. For each acquisition area, a multi-angle ring light source projects sinusoidal striped structured light from eight preset elevation angles. After each angle of structured light projection is completed, the orthogonal polarization imaging module is simultaneously triggered to acquire two sets of orthogonal polarization images at that angle. After all areas and all angles are acquired, the images of each area are stitched together to form a full-screen image, establishing a multi-dimensional image dataset with an angle-polarization joint index. The image preprocessing subsystem performs pixel-by-pixel polarization difference operation on each pair of orthogonal polarization images to remove the specular reflection component of the LCD screen surface to obtain a difference image. Then, dark channel scattering suppression processing is performed on the difference image to remove background noise caused by ambient scattered light. Afterward, regional adaptive gamma correction is performed based on the average gray level of each area to improve the clarity of low-contrast defect features in different brightness areas, finally obtaining the full-screen preprocessed difference image corresponding to each angle.
[0045] The 3D topography reconstruction subsystem performs pixel-level registration between preprocessed differential images at the same angle and pre-stored stripe projection patterns, and extracts sub-pixel-level height topography maps of the LCD screen surface using Fourier profilometry. First, outlier filtering is performed on the height topography maps at each angle. A robust local weighted regression method is used to calculate the residual of each height data point in its local neighborhood. Points with residuals exceeding a preset threshold are marked as measurement outliers, and their measured values are replaced with the median height value of the neighborhood, eliminating phase unwrapping artifacts caused by local high reflectivity of the LCD screen. After filtering, the height topography maps at each angle are converted into 3D point cloud data. Using the full-screen point cloud at the first angle as a reference, an iterative nearest-point algorithm is used to complete multi-view fine registration of the point clouds at the remaining angles, aligning all point clouds to the same reference coordinate system. After registration, fusion weights are assigned to each point cloud based on angle confidence. The smaller the root mean square error of phase unwrapping and the smaller the angle between the principal direction of the light source and the normal to the screen surface, the higher the weight of the corresponding point cloud. After weighted voxel fusion, a 3D topography model of the large-size LCD screen surface is generated.
[0046] The defect feature extraction subsystem stitches preprocessed differential images from all angles along the angular dimension into a multi-channel image, which is then input into a pre-trained deep convolutional backbone network adapted to large-size images. The backbone network employs a five-level downsampling residual network structure, adjusting the convolution stride of each layer to adapt to the large-size input image. Feature maps of different resolutions output from each stage are input into a feature pyramid structure to extract point defect texture, line defect geometry, and surface defect semantic features at corresponding scales, constructing a multi-scale defect feature map set. The defect feature maps from each layer are then input into a concatenated deformation convolutional attention module and a cross-scale fusion module. For each sampling position of the input feature map, the deformation convolutional attention module predicts a two-dimensional deformation offset based on a standard 3x3 convolutional sampling grid by adding two convolutional branches. This allows the sampling position of the convolutional kernel to adaptively offset spatially according to the actual shape of the defect. For each offset sampling position, the cosine similarity between its local features and the pre-defined defect category prototype feature vector is calculated, and after nonlinear activation, it is mapped to attention weights, which are then multiplied with the convolutional output to complete feature enhancement. The additional convolutional layers, offset prediction branch, and attention branch are optimized end-to-end during the training phase. The loss function is obtained by a weighted combination of feature alignment loss and detection loss. The learning of deformation offset is constrained by the ground truth annotation of defect segmentation, which improves the feature alignment effect of irregular defect boundaries. The cross-scale fusion module performs upsampling layer by layer starting from the semantic features of the highest layer, and performs weighted summation with the texture features of the next layer to complete the cross-scale aggregation of high-level semantics and low-level texture, and finally outputs a defect enhancement feature map that matches the resolution of the input image.
[0047] The multi-task defect detection subsystem inputs the enhanced feature map of defects into a multi-task detection network adapted to the defect scale of large-size screens. The network includes a point defect segmentation branch, a line defect detection branch, and a surface defect classification branch. These three branches share backbone network weights and each has its own independent anchor point parameters and detection head adapted to large-size defects. The point defect segmentation branch uses a fully convolutional structure to output a pixel-level point defect segmentation mask. It performs principal curvature analysis on each connected defect region, obtains the height distribution surface equation of the defect region through quadratic surface fitting, calculates the principal curvature of each pixel within the region, and determines the concavity / convexity attribute of the defect based on the ratio of the mean principal curvature of the region, distinguishing between convex defects, concave defects, and flat contamination defects. The line defect detection branch uses an anchor-frame-based detection head and outputs the vector trajectory coordinates of line defects. The surface defect classification branch outputs the bounding box coordinates and corresponding confidence scores of large-size surface defects.
[0048] The defect grading and evaluation subsystem performs defect severity grading on the detection results output by each branch. For each detected defect, it extracts planar geometric parameters such as the equivalent circle diameter and aspect ratio from the segmentation mask or bounding box, and extracts depth morphology parameters such as the maximum indentation depth or maximum protrusion height of the defect relative to the surrounding normal surface reference plane from the corresponding spatial coordinates of the 3D morphology model. These two types of parameters are input into a pre-trained and fixed multi-dimensional threshold classifier adapted to the quality standards of large-size screens. The classifier consists of three cascaded binary support vector machines. The first-level support vector machine distinguishes between acceptable and unacceptable defects using planar geometric parameters as input. The second-level support vector machine uses the ratio of defect height to equivalent circle diameter and the concavity / convexity discrimination index as input to distinguish unacceptable defects into acceptable defects and rejected defects. The third-level support vector machine uses the defect hazard index and the position coordinates of the defect within the display area as input to determine the rework feasibility of rejected defects. After all defects are classified, when the number of rejection-level defects is zero and the number of acceptance-level defects is less than the preset maximum acceptable number, the LCD screen is deemed to be qualified as a whole, and a full-screen defect distribution quality map is finally generated.
[0049] This embodiment is adapted to the offline sampling inspection requirements of large-size LCD screens. The area-based acquisition and stitching method can balance full-area coverage and detection accuracy of large-size screens, effectively eliminating interference from specular reflection and uneven illumination in different areas of the large-size screen. The reconstructed 3D morphology model can accurately reflect the surface height distribution of the large-size screen. The defect feature extraction structure can adapt to various irregular defect morphologies of large-size screens. Multi-task parallel detection can balance the detection accuracy of defects of different scales. The grading method of morphology parameters can match the quality judgment standards of large-size screens. The system's built-in calibration compensation mechanism can reduce the impact of positioning deviation of the large-size stage, meeting the high-precision detection requirements of sampling inspection scenarios. Reference Figure 1This flowchart illustrates the overall architecture and core processing steps for defect image detection in smart LCD screens. The system first utilizes a multi-angle ring light source and orthogonal polarization imaging technology to acquire multi-dimensional image data. Polarization difference calculations and preprocessing are then performed to effectively eliminate specular reflection interference. In the feature processing stage, the process is divided into two parallel branches: one is to reconstruct a high-precision 3D surface topography model using structured light patterns; the other is to extract and enhance defect features using deep convolutional networks and attention mechanisms. Finally, the 2D defect features output by the multi-task detection network are deeply fused with the 3D topography model, and a defect hazard assessment is performed by combining planar dimensions and 3D depth, achieving accurate quality grading of the LCD screen and generating a distribution quality map.
[0050] Reference Figure 2 This flowchart details the equipment self-calibration and de-reflection image preprocessing before image data acquisition. Before acquisition begins, the system uses a spectroradiometer and laser displacement sensor to correct the light source radiance and measure the stage tilt angle. If the detected tilt exceeds the preset tolerance, the compensation mechanism automatically adjusts the attitude to ensure a horizontal reference level. During the acquisition phase, the system performs differential operations on polarized orthogonal image pairs to remove strong specular reflections from the LCD screen surface. Subsequently, to improve the contrast of extremely subtle defects, the system performs dark channel scattering suppression and adaptive gamma correction in parallel, effectively improving the image's signal-to-noise ratio and brightness uniformity, laying a high-quality data foundation for subsequent 3D reconstruction and feature extraction.
[0051] Reference Figure 3 This flowchart illustrates the specific logic of extracting height data and reconstructing the 3D shape of an LCD screen using Fourier profilometry. First, the preprocessed differential image is registered with the stripe projection pattern to extract sub-pixel height information. Then, a robust local weighted regression filtering algorithm is introduced. To address phase unwrapping artifacts caused by local high reflectivity, the system intelligently identifies outliers and uses a median replacement strategy for smoothing correction. After single-angle filtering, an iterative nearest-point algorithm is used to align the point cloud data from different viewpoints to the reference coordinate system using a rigid body transformation. Finally, by combining parameters such as illumination angle and root mean square error of measurement to determine the calculation weights, weighted voxel fusion is performed to construct a complete and high-precision 3D digital model of the LCD screen surface.
[0052] Reference Figure 4This flowchart illustrates the internal operating mechanism of a multi-scale feature extraction and multi-task parallel detection network for complex and irregular defects. After the image is input into the backbone network, feature information at different levels is obtained through a feature pyramid. To address the challenge of diverse defect morphologies in LCD screens, a deformation convolutional attention module is introduced. This module adaptively learns the geometric offset of the defect boundary to achieve irregular feature alignment and, in conjunction with a cross-scale module, fuses high-level semantics with low-level texture to generate an enhanced feature map. Subsequently, this feature map is fed into a parallel multi-task detection network. The point defect segmentation branch, line defect detection branch, and surface defect classification branch each utilize independent anchor points and detection heads to perform parallel computation and accurately output segmentation masks, vector trajectories, and bounding boxes.
[0053] Reference Figure 5 This flowchart reveals the multi-level classification and final quality assessment process for defect severity, integrating 2D parameters and 3D morphology data. The system receives the 2D geometric parameters from the detection network and simultaneously indexes the corresponding depth and height data into the 3D morphology model. Principal curvature analysis based on surface fitting is specifically performed on point defects to accurately calculate their concavity / convexity properties. Subsequently, the extracted comprehensive parameters are input into a cascaded multi-dimensional threshold classifier for progressive evaluation: the first-level network intercepts conventional acceptable defects; the second-level network deeply evaluates morphological indicators to classify acceptance or rejection levels; and the third-level network further assesses the feasibility of repairing rejected defects. Finally, the system integrates the various assessment results with the full-screen acceptance quantity limit, outputting an intuitive defect distribution quality map.
[0054] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for detecting defect images in smart LCD screens, characterized in that, Includes the following steps: S1. A multi-angle ring light source sequentially projects striped structured light onto the LCD screen under test from a preset elevation angle, and simultaneously uses an orthogonal polarization imaging module to acquire orthogonal image pairs of polarization states at each angle to establish a multi-dimensional image dataset. S2. Perform polarization difference operation on each pair of polarization state orthogonal images to obtain a difference image with specular reflection removed, and then perform dark channel scattering suppression and adaptive gamma correction to obtain preprocessed difference images at each angle. S3. Register the preprocessed differential image at the same angle with the stripe projection pattern, extract the subpixel-level height topography map of the LCD screen surface using Fourier profilometry, and construct a three-dimensional topography model of the LCD screen surface by weighted fusion of the height topography maps at each angle. S4. The preprocessed differential images from each angle are spliced along the angle dimension and input into the pre-trained deep convolutional backbone network. The feature pyramid is used to extract point defect texture, line defect geometry and surface defect semantic features from multiple stages to construct a multi-scale defect feature map set. S5. Input the defect feature maps of each layer into the deformable convolutional attention module and the cross-scale fusion module set in series. The deformable convolutional attention module obtains the defect geometric offset through deformable convolution kernels. The cross-scale fusion module upsamples layer by layer and performs weighted summation, fusing high-level semantics and low-level textures to output the defect enhancement feature map. S6. Input the defect enhancement feature map into the multi-task defect detection network. The network has a point defect segmentation branch, a line defect detection branch, and a surface defect classification branch. Each branch shares the backbone weight and uses independent anchor points and detection heads. It outputs point defect segmentation masks, line defect vector trajectories, and surface defect bounding boxes and confidence scores in parallel. S7. Perform defect severity classification based on three-dimensional topography model for each branch output, calculate the defect hazard index by combining defect planar size and depth topography parameters, classify defects into three levels of qualified, acceptable and rejected according to the preset LCD screen quality standard, and generate a full-screen defect distribution quality map.
2. The defect image detection method for a smart LCD screen according to claim 1, characterized in that, It also includes: in step S1, a spectroradiometer is used to pre-calibrate the spectral power distribution and spatial illuminance uniformity of each angle of the multi-angle ring light source, and a light source calibration parameter lookup table is established. When acquiring each batch of LCD screen images, radiometric correction is performed on each angle polarization image pair according to the lookup table. A laser displacement sensor is used to measure the spatial coordinates of the four corner reference marks of the LCD screen positioning stage, and the spatial tilt angle deviation between the stage plane and the imaging optical axis is calculated. When the tilt angle deviation exceeds the preset tolerance, the stage adjustment mechanism is driven to complete the attitude compensation and then the image is re-acquired.
3. The defect image detection method for a smart LCD screen according to claim 1, characterized in that, Also includes: In step S3, local weighted regression outlier filtering is performed on the topography images of each angle. Points whose residuals in the local neighborhood of each height data point exceed the threshold are marked as measurement outliers. The outliers are replaced by the median height value of the neighborhood. The iterative nearest point algorithm is used to perform multi-view point cloud fine registration on the filtered topography images of each angle. Taking the first angle point cloud as the reference, the other angle point clouds are aligned to the reference coordinate system through rigid body transformation. After registration, the weights are determined according to the vertical distance of each point to the nearest neighbor patch of the reference point cloud. The three-dimensional topography model is obtained by weighted voxel fusion.
4. The defect image detection method for a smart LCD screen according to claim 1, characterized in that, In step S5, the deformation convolutional attention module superimposes the two-dimensional deformation offset predicted by the additional convolutional layer onto the standard convolutional sampling grid at each sampling position of the input feature map. The sampling position of the convolutional kernel generates a spatial offset based on the defect morphology. Each offset sampling position generates attention weights based on the cosine similarity between the local features and the corresponding defect category prototype feature vector through nonlinear activation mapping. The additional convolutional layer, offset prediction branch, and attention branch adopt an end-to-end approach, and are jointly optimized based on a weighted combination of alignment loss and detection loss. The deformation offset learning process introduces defect segmentation ground value labeling constraints.
5. The defect image detection method for a smart LCD screen according to claim 1, characterized in that, In step S6, the point defect segmentation branch performs defect morphology classification based on principal curvature analysis for each connected defect region: The height distribution surface equation is obtained by fitting a quadratic surface to the coordinates of all pixels within the defect region. The maximum and minimum principal curvatures of each pixel are calculated from this surface equation. The ratio of the average maximum principal curvature to the average minimum principal curvature of all pixels within the region is used as the concavity / convexity discrimination index of the defect region. This concavity / convexity discrimination index is quantitatively calculated using the following formula: In the formula, Indicates the first The unevenness discrimination index of a point defect area This represents the total number of pixels within the defect area at that point. For pixel index; and They represent the first The absolute values of the maximum and minimum principal curvature of each pixel. This is a preset regularization constant.
6. The defect image detection method for a smart LCD screen according to claim 1, characterized in that, The defect hazard index mentioned in step S7 is calculated by combining planar geometric parameters and depth morphology parameters. The planar geometric parameters include the diameter of the defect equivalent circle and the aspect ratio. The depth morphology parameters include the maximum indentation depth or maximum protrusion height of the defect relative to the surrounding normal surface reference plane in the three-dimensional morphology model. For each defect output in step S6, extract planar geometric parameters from the mask or bounding box, and extract depth morphology parameters from the spatial coordinates corresponding to the three-dimensional morphology model index. Input both into a multi-dimensional threshold classifier pre-trained based on the LCD screen defect acceptance criteria, and output the qualified, acceptable, or rejected quality level corresponding to the defect. When the number of rejected defects is zero and the number of acceptable defects is less than the preset maximum acceptable number, the LCD screen is determined to be qualified.
7. The defect image detection method for a smart LCD screen according to claim 1, characterized in that, The weighted fusion weights of the height topography images at each angle mentioned in step S3 are determined according to the following angle confidence weighting formula: In the formula, Indicates the first Each illumination angle corresponds to a normalized weight in the height topography map during fusion, and satisfies... , This is the attenuation control coefficient. Indicates the first The root mean square error of phase unwrapping at each angle. Indicates the first The angle between the principal direction of the light source and the surface normal. and They represent the first The root mean square error of each angle and the angle between the rays, The total number of lighting angles in the preset elevation angle set and .
8. A defect image detection system for intelligent LCD screens, characterized in that, A defect image detection method for a smart LCD screen according to any one of claims 1 to 7 includes the following modules: The image acquisition subsystem is equipped with a multi-angle ring structure light source, an orthogonal polarization imaging module and an LCD screen positioning stage, which is used to acquire multi-dimensional polarization state orthogonal image pairs of the LCD screen under test by multi-angle structure light projection and orthogonal polarization synchronous acquisition. The image preprocessing subsystem is equipped with a polarization difference operation module, a dark channel scattering suppression module, and an adaptive gamma correction module, which are used to perform specular reflection removal and enhancement processing on polarization image pairs at various angles to generate preprocessed difference images. The three-dimensional topography reconstruction subsystem is equipped with a Fourier contour phase extraction module, an outlier filtering module, a multi-view point cloud registration module, and a weighted voxel fusion module, which are used to reconstruct the three-dimensional topography model of the LCD screen surface from structured light patterns and differential images. The defect feature extraction subsystem is equipped with a deep convolutional backbone network, a feature pyramid, and a deformation convolutional attention module, which is used to extract multi-scale defect features from preprocessed difference images and align the deformation features of irregular defects. The multi-task defect detection subsystem is configured with point defect segmentation branches, line defect detection branches, and surface defect classification branches, which are used to output point defect segmentation masks, line defect vector trajectories, and surface defect bounding boxes and confidence scores in parallel. The defect classification and assessment subsystem is equipped with a defect hazard index calculation module, a multidimensional threshold classifier, and a full-screen quality map generation module. It is used to classify defect quality levels by comprehensively considering planar geometric parameters and depth morphology parameters and generate a full-screen defect distribution quality map.
9. A defect image detection system for a smart LCD screen according to claim 8, characterized in that, The multi-angle ring light source in the image acquisition subsystem consists of narrow-band LED beads arranged along the circumference. The pitch angle of each group of LED beads is independently adjusted within a preset pitch angle range by a stepper motor. The orthogonal polarization imaging module includes a linear polarizer and a quarter-wave plate, with the fast axis angle between the two fixed at 45 degrees. Laser reflection reference marks are embedded at the four corners of the LCD screen positioning platform. After the LCD screen under test is replaced, the laser displacement sensor detects the spatial position of the reference marks.
10. A defect image detection system for a smart LCD screen according to claim 8, characterized in that, The multidimensional threshold classifier of the defect grading and evaluation subsystem consists of three cascaded binary support vector machines. The first-level support vector machine uses the equivalent circle diameter and aspect ratio of the defect as input to distinguish between acceptable and unacceptable defects. The second-level support vector machine uses the ratio of the maximum indentation depth of the defect to the equivalent circle diameter, the ratio of the maximum protrusion height of the defect to the equivalent circle diameter, and the concavity / convexity discrimination index as input to distinguish between acceptable and rejected defects. The third-level support vector machine uses the defect hazard index and the position coordinates of the defect within the LCD display area as input to determine the rework feasibility of rejected defects.
