Method and system for diagnosing foreign matter defects of mobile phone screen backlight

Abstract

The application discloses a mobile phone screen backlight foreign matter defect diagnosis method and system, relates to the display quality detection technical field, and comprises the following steps: constructing a diffraction interference model based on electromagnetic time domain solution, generating a stripe prior library according to angle parameters and temperature parameters; under the constraint of the stripe prior library, angle scanning and polarization multiplexing are performed, and narrowband spectrum acquisition is combined to obtain a space-time polarization three-dimensional data body and complete parameter calibration. The application constructs a diffraction interference model by electromagnetic time domain solution, generates a stripe prior library and combines multi-dimensional data for acquisition and calibration. Based on this, a physical guide discrimination network is trained to separate stripe and defect boundaries, and an importance map is output. Subsequently, an optimal action set is generated through virtual relighting simulation, closed-loop adaptive regulation is realized, imaging parameters are updated in real time and resampling is triggered, the contrast of the defect boundary is enhanced, and accurate defect identification is ensured. The method solves the problem of artifact influence and improves the quality control precision of the backlight module.

Description

Technical Field

[0001] This invention relates to the field of display quality testing technology, specifically to a method and system for diagnosing foreign object defects in the backlight of mobile phone screens. Background Technology

[0002] "Foreign Object Defect Diagnosis for Mobile Phone Screen Backlight" refers to the detection and analysis of foreign object problems (such as dust particles, bubbles, scratches, fibers, etc.) that may occur during the manufacturing or use of the backlight module inside the mobile phone screen. This is achieved through image processing technology to identify and classify defects. Specifically, under screen illumination or specific lighting conditions, a high-resolution acquisition device is used to acquire images of the brightness distribution of the backlight area. Then, image preprocessing (denoising, enhancement, contrast adjustment), feature extraction (brightness anomalies, texture disturbances, edge mutations), and defect recognition algorithms (such as pattern matching and deep learning convolutional neural networks) are used to analyze the acquired images, ultimately determining the presence, type, and location of foreign object defects. This diagnostic method not only improves detection accuracy and efficiency but also avoids subjective errors inherent in manual inspection, which is of great significance for ensuring screen display quality.

[0003] The existing technology has the following shortcomings:

[0004] In existing technologies, when a foreign object with a size close to the wavelength of the light source is present inside the backlight module, the incident light will produce significant diffraction and interference effects at the boundary of the foreign object, thus forming a complex stripe structure in the detection image. Because these stripes are highly sensitive to external conditions, their morphology will continuously and dynamically adjust with minute changes in the detection angle and temperature, resulting in flicker-like artifact signals in the image sequence. These artifacts not only cause nonlinear interference in the brightness distribution but also severely affect the stable extraction of edge features by image processing algorithms, causing the algorithms to confuse the artifacts with real defect boundaries. Furthermore, during dynamic detection, key defect areas may be obscured, distorted, or incompletely segmented, leading to a significant decrease in the accuracy of defect diagnosis results and seriously affecting backlight quality control.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for diagnosing foreign object defects in the backlight of mobile phone screens, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for diagnosing foreign object defects in mobile phone screen backlight, comprising the following steps:

[0008] S001, Construct a diffraction interference model based on electromagnetic time-domain solution, and generate a fringe prior library based on angle parameters and temperature parameters to quantify the nonlinear influence of fringes on brightness distribution;

[0009] S002, under the constraint of the fringe prior library, performs angle scanning and polarization multiplexing, and combines narrowband spectral acquisition to obtain a three-dimensional spatiotemporal polarization data volume and complete parameter calibration, so that the acquisition results are consistent with the fringe prior library;

[0010] S003, a physical-guided generation and discrimination network is trained based on the calibrated spatiotemporal polarization three-dimensional data volume to separate the stripe driving term from the defect boundary and output the boundary salience map;

[0011] S004, based on the boundary saliency map, stripe phase unwrapping and structural sparse decomposition are performed to generate the foreign object influence field and extract the perturbation spectrum of edge features;

[0012] S005, reinject the foreign object influence field into the optical model, run virtual re-illumination simulation, generate exposure strategy, acquisition trajectory and temperature control trajectory, and obtain the optimal action set;

[0013] S006 implements closed-loop adaptive optical field polarization temperature control based on the optimal action set, updates imaging parameters in real time using a Bayesian design mechanism, and triggers resampling in key areas to enhance defect boundary contrast and complete diagnostic closure.

[0014] Preferably, step S001 includes:

[0015] A diffraction interference model based on electromagnetic time-domain solution is constructed to simulate the propagation and interaction of incident light inside and at the boundary of a foreign object structure, and a fringe prior library is generated to quantify the nonlinear influence of fringes on brightness distribution.

[0016] Based on the simulation model, angle and temperature control variables are introduced to simulate the influence of different detection angles and working temperatures on the stripe structure and obtain stripe feature data under different conditions.

[0017] The electromagnetic field intensity data obtained from the simulation is converted into a two-dimensional brightness image, and the actual influence of fringe perturbation on the brightness distribution is extracted, including features such as fringe frequency, direction, and contrast.

[0018] A stripe prior library is constructed based on image data, and the stripe feature quantities in the brightness image are dually mapped with the corresponding parameters to generate a complete stripe feature data record, supporting efficient data retrieval and retrieval.

[0019] Preferably, step S002 includes:

[0020] Based on the prediction results of the angle and temperature sensitive range in the stripe prior library, a detailed acquisition plan was formulated to perform angle scanning, polarization multiplexing and narrowband spectral acquisition to generate a spatiotemporal polarization three-dimensional data volume.

[0021] Polarization multiplexing acquisition is performed under angle scanning and temperature conditions, and narrowband spectroscopy is combined with acquisition to obtain a raw image data sequence with complete angle, polarization and wavelength information;

[0022] Flat field correction and dark field subtraction are performed on the acquired image data, and angle consistency, polarization consistency, spectral consistency, radiometric consistency and temperature consistency are calibrated to ensure that the image data is consistent with the fringe prior library;

[0023] The calibrated image data is used to generate a spatiotemporal polarization three-dimensional data volume, which is then matched with the fringe prior library to ensure accurate alignment of the data acquisition results.

[0024] Preferably, step S003 includes:

[0025] Using the spatiotemporal polarization three-dimensional data volume that has completed consistency calibration as training input, training samples are constructed, and four types of quantitative descriptions, namely fringe frequency, fringe direction, fringe contrast and fringe phase, are extracted to generate a fringe reference image.

[0026] Guided by the fringe reference map, a physical-guided generation and discrimination network is established, and the fringe driving term is separated from the defect boundary through the network to output the boundary salience map.

[0027] By training the generated network, stripe phase unwrapping and structural sparse decomposition are performed on the boundary saliency map to generate the foreign object influence field and extract the perturbation spectrum of edge features.

[0028] Based on the generated foreign object influence field back-injection optical model, a virtual re-illumination simulation is performed to optimize the exposure strategy, acquisition trajectory, and temperature control trajectory. Resampling is triggered in key areas to enhance the contrast of defect boundaries, and finally, the diagnostic loop is completed.

[0029] Preferably, during training, the local details of the samples are further optimized by introducing resampling of high-brightness and low-brightness regions in each training sample to ensure comprehensive capture of the stripes. At the same time, by coordinating angle, polarization and center wavelength labels, the spatial and conditional dimensions of the samples are ensured to be strictly aligned with the stripe prior library, thereby improving the accuracy and robustness of network training.

[0030] Preferably, step S004 includes:

[0031] Based on the spatiotemporal polarization three-dimensional data volume that has been uniformly calibrated, the fringe phase is unwrapped based on the boundary saliency map, and uniform phase coordinates across angles, polarizations, and wavelengths are established, while avoiding interference from the real geometric boundary on the fringe phase estimation.

[0032] Based on the unwrapped phase, structural sparse decomposition is performed on the brightness map of each frame to separate the periodic stripe component, geometric boundary component and slowly varying background component, ensuring that the stripe component does not interfere with the extraction of the true boundary.

[0033] The foreign object influence field is constructed using the fringe component and the boundary component. The influence of the fringes on the boundary contrast and position is quantified, and the influence information is recorded with angle, polarization and wavelength labels.

[0034] Based on the foreign object influence field, the perturbation spectrum most unfavorable to edge features is extracted and perturbation bands are generated to ensure that all results are consistent with the measured data in the playback simulation, providing input for subsequent simulation and execution strategy generation.

[0035] Preferably, step S005 includes:

[0036] The foreign object influence field is embedded into the optical propagation link, defining the optical structure of the LED surface source, light guide plate, diffuser, brightness enhancement film and emission surface, and writing phase gradient and contrast reduction amplitude into the propagation model;

[0037] Perform virtual relighting simulations, traversing different angles, polarizations, spectral and exposure combinations, and calculate the boundary contrast enhancement, stripe retention ratio and geometric stability of each frame of image to form a scoring matrix;

[0038] Based on the scoring matrix, a false exposure suppression strategy, acquisition trajectory and temperature control trajectory are generated, and time and energy constraints are introduced for executability verification. After the verification is passed, a candidate action sequence is generated.

[0039] The candidate action sequences are virtually executed and robustness tested. The optimal action set is selected and an execution list that can be directly issued is output, ensuring that the preset contrast, stripe residue and geometric stability indicators are achieved under all conditions.

[0040] Preferably, boundary saliency is performed based on each frame of the candidate action sequence. Figure 1 Consistent fusion is achieved by adjusting the weights based on the stability of cross-frame alignment and by backtracking the entries for adjacent angles and polarizations in the stripe prior library to optimize boundary contrast.

[0041] Preferably, step S006 includes:

[0042] Based on the action set generated by the virtual relighting simulation, a dynamic feedback control mechanism is constructed to adjust the exposure time, angle, polarization, spectrum and temperature combination in real time to ensure maximum contrast at the defect boundary.

[0043] By employing a Bayesian design mechanism, optical parameters are dynamically adjusted based on each image feedback, and the optimal adjustment scheme is calculated in real time to improve defect recognition accuracy and reduce acquisition errors.

[0044] In key areas, online resampling is triggered, and stripe contrast is optimized by adjusting local brightness, polarization direction, and exposure time to reduce artifact interference and ensure the clarity of defect boundaries.

[0045] The closed-loop adaptive control is executed. Based on the image quality feedback after each adjustment, the boundary contrast, stripe residue and geometric stability indicators are verified to meet the preset requirements. The control parameters are then further optimized to complete the diagnostic closed loop.

[0046] A mobile phone screen backlight foreign object defect diagnosis system includes a stripe modeling module, a data acquisition module, a feature separation module, an interference analysis module, a simulation optimization module, and a closed-loop control module.

[0047] The fringe modeling module constructs a diffraction interference model based on electromagnetic time-domain solutions and generates a fringe prior library based on angle and temperature parameters to quantify the nonlinear influence of fringes on brightness distribution.

[0048] The data acquisition module performs angle scanning and polarization multiplexing under the constraints of the fringe prior library, and combines narrowband spectral acquisition to obtain a three-dimensional spatiotemporal polarization data volume and complete parameter calibration, so that the acquisition results are consistent with the fringe prior library.

[0049] The feature separation module trains a physically guided generation and discrimination network based on the calibrated spatiotemporal polarization 3D data volume to separate the stripe driving term from the defect boundary and output a boundary salience map.

[0050] The interference analysis module performs stripe phase unwrapping and structural sparse decomposition based on the boundary saliency map, generates the foreign object influence field, and extracts the perturbation spectrum of edge features.

[0051] The simulation optimization module injects the foreign object influence field back into the optical model, runs a virtual re-illumination simulation, generates an exposure strategy, acquisition trajectory, and temperature control trajectory, and obtains the optimal action set.

[0052] The closed-loop control module implements closed-loop adaptive optical field polarization temperature control based on the optimal action set, updates imaging parameters in real time using a Bayesian design mechanism, and triggers resampling in key areas to enhance the contrast of defect boundaries and complete the diagnostic closed loop.

[0053] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0054] This invention constructs a diffraction interference model based on electromagnetic time-domain solutions to generate a fringe prior library, accurately quantifying the influence of fringes on brightness distribution. It also incorporates multi-dimensional data such as angle, polarization, and temperature for acquisition and calibration, ensuring consistency between the acquired results and the fringe prior library. Based on this, a physically guided discriminant network is trained to effectively separate fringes from defect boundaries, outputting a boundary saliency map to provide precise support for subsequent phase unwrapping and sparse structural decomposition. Finally, the optimal action set generated by virtual re-illumination simulation is used for closed-loop adaptive control, updating imaging parameters in real time and triggering resampling in key areas, thereby enhancing defect boundary contrast and ensuring accurate defect identification and location. This method solves the problems of artifact effects and inaccurate defect identification in traditional detection methods, improving the quality control level of backlight modules. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0056] Figure 1 This is a flowchart of the method for diagnosing foreign object defects in the backlight of a mobile phone screen according to the present invention.

[0057] Figure 2 This is a schematic diagram of the module of the mobile phone screen backlight foreign object defect diagnosis system of the present invention. Detailed Implementation

[0058] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0059] This invention provides, for example Figure 1 The method for diagnosing foreign object defects in the backlight of a mobile phone screen, as shown, includes the following steps:

[0060] S001, Construct a diffraction interference model based on electromagnetic time-domain solution, and generate a fringe prior library based on angle parameters and temperature parameters to quantify the nonlinear influence of fringes on brightness distribution;

[0061] This step relies on electromagnetic time-domain simulation to establish a diffraction interference model, generate a fringe prior library with coverage angle and temperature variables, and quantify the nonlinear interference of fringes on brightness distribution, providing a physical basis for subsequent optical image acquisition and discrimination. This process includes the following steps:

[0062] A diffraction and interference simulation model based on electromagnetic time-domain principles was constructed to simulate the propagation and interaction of incident light within and at the boundaries of a foreign object structure. In the actual modeling process, a typical mobile phone backlight module was used as a reference structure, selecting a complete optical stack including a light source layer, light guide plate, diffuser, brightness enhancement film, and exit layer. Irregular particulate foreign objects with sizes ranging from 100 nm to 700 nm were introduced onto the surface of the light guide plate. The light source was set as a monochromatic collimated beam with a center wavelength of 550 nm, and the polarization direction was fixed perpendicular to the exit surface. Spatial discretization employed fine mesh generation with a mesh step size set to one-fifteenth of the light source wavelength, and a perfectly matched absorption layer was placed at the simulation boundary to eliminate reflection interference. The entire simulation space was calculated in real-time by solving Maxwell's equations in the time domain to dynamically calculate the distribution of electric and magnetic fields near the foreign object. Through this modeling process, phenomena such as local field strength enhancement, phase perturbation, and beam deflection caused by diffraction and interference can be obtained. These phenomena constitute the basic physical process for the subsequent formation of fringe interference. Compared with the traditional method of estimating fringe distribution through empirical formulas, this step restores the real electromagnetic propagation behavior through numerical calculation, revealing every perturbation detail in the physical path, and providing traceable, high-precision source data for the subsequent establishment of fringe-parameter correspondence.

[0063] The simulation model's input mainly includes three types of detailed parameter information. The first type is structural input, including the 3D topography, thickness, refractive index distribution, and surface roughness information of all key layers in the backlight module, involving the complete optical path of the emitting layer, light guide plate, diffuser, brightness enhancement film, and emission layer. The second type is foreign object parameters, explicitly describing the foreign object's size range, material properties, embedding depth, and surface topography characteristics to ensure accurate modeling of the interaction between the foreign object and the light propagation path. The third type is excitation conditions, including the wavelength, polarization state, illumination angle, light intensity distribution, and time step interval of the incident light, used to drive the entire simulation process. The model's output is the full-field response of the light wave propagating through the foreign object and material layers, specifically manifested as the intensity distribution, direction change, and energy flow of the electric and magnetic fields at each time step in three-dimensional space, as well as the resulting local enhancement, fringe interference, edge diffraction, and scattering regions. It also outputs the time-evolving light energy distribution map, phase perturbation structure, and interference fringe pattern for subsequent image prediction and brightness nonlinear modeling.

[0064] During the modeling process, the geometry of the backlight module is first reconstructed in 3D space to ensure that the physical boundaries, thickness, refractive index, and interface relationships between each material layer and adjacent layers are consistent with the real structure. Based on this, foreign objects are embedded in designated areas, and their boundary contours and material properties are meticulously characterized to accurately induce light scattering and diffraction. Subsequently, an incident light source is introduced into the simulation space, with its illumination position, wavelength, polarization direction, and propagation angle set, allowing it to be emitted from a designated layer and begin propagating within the structure. During light wave propagation, the simulation engine advances the light field evolution layer by layer in a time-step manner, tracking the spatial response of the electric and magnetic fields in real time, focusing on monitoring beam deflection, wavefront distortion, interference enhancement regions, and the resulting bright and dark fringe structures around the foreign object. To prevent boundary reflections from interfering with the results, a highly absorptive boundary layer surrounds the simulation space, causing all propagating waves to automatically attenuate after leaving the computational region, simulating an open space environment. The entire modeling process encompasses the physical interaction between the foreign object and the background structure, enabling the model to realistically reproduce all the physical phenomena when light propagates between complex media and disturbance sources. Ultimately, a complete simulation closed loop is established, covering structural modeling, light source excitation, field response tracking, and output reconstruction.

[0065] Based on the simulation model, controllable variables were introduced to simulate the effects of different detection angles and operating temperatures on the stripe structure in the actual detection environment. For angle control, the angle between the incident light direction and the image acquisition direction was set between 0 and 20 degrees, with each 2-degree increment representing a step unit, covering common process ranges for normal illumination and oblique illumination. At each angle setting, the incident light direction was reconfigured, and electromagnetic propagation simulation was performed again. For temperature control, the temperature range was set from 20℃ to 60℃, with each 5℃ increment representing a step, for a total of nine temperature points. At each temperature, the refractive index, dielectric constant, and material dispersion parameters of the light guide plate and the foreign object were recalculated. For example, when the temperature rises from 20℃ to 60℃, the refractive index of the polycarbonate light guide plate changes from 1.586 to 1.578. The corresponding slight changes in light propagation speed and reflection conditions are sufficient to cause variations in stripe period and intensity fluctuations. Each simulation meticulously records the incident angle, acquisition angle, temperature setting, medium parameters, simulation time, fringe orientation, frequency distribution, and interference intensity pattern, laying the foundation for establishing a complete mapping relationship across parameter dimensions. This step differs from existing methods that construct fringe parameter tables through image capture and inference, achieving high-resolution modeling of the fringe response driven by both angle and temperature, thus avoiding the lack of physical parameter support in actual detection.

[0066] The electromagnetic field intensity distribution obtained from time-domain simulation was converted into a two-dimensional brightness image, and the actual impact of fringe perturbation on the brightness distribution was extracted. The electric and magnetic field data from the simulation results were first subjected to Fourier transform to obtain spectral characteristics, and then orthogonally projected onto the exit surface to obtain a light intensity image. Each pixel value represents the energy flux density per unit area of ​​that region. To match the actual acquired images, the spatial resolution of the brightness image was set to 400 pixels per millimeter, with 4096 pixels horizontally and vertically, covering the entire backlit area. During image generation, the light guide plate, foreign objects, light source position, medium refractive index, and polarization direction were input together to ensure that the final brightness image completely retains all interference-induced detail information. The fringe structure in the image appears as striped alternating bright and dark textures, with spatial frequencies ranging from 3 to 20 fringe per millimeter, and the fringe direction forming a certain angle with the incident direction. For each set of images, the mean brightness was calculated, local contrast was evaluated, frequency components were extracted, directionality was statistically analyzed, and brightness shift was estimated to extract the fringe feature quantities that form nonlinear interference. The stripe behavior exhibited by the images differed significantly under different angles and temperature combinations. For example, the stripe tilt angle increased with high-angle oblique illumination, and the stripe spacing increased with rising temperature. All images were archived with timestamps and recorded in correspondence with simulation input parameters, achieving a one-to-one mapping between parameters and image content. This processing method differs from the traditional approach of only observing brightness deviations; it is the first to convert complete physical field information into a quantifiable image representation, ensuring that the stripe effect is not only visible and traceable but also reproducible.

[0067] A fringe prior library is constructed based on image data, establishing a dual mapping relationship with angle and temperature parameters. In the actual construction process, the fringe prior library is based on an angle-temperature two-dimensional index, binding and storing fringe features extracted from the brightness image with their corresponding parameters. Each record contains 12 indicators: incident angle, receiving angle, temperature value, light guide material name and parameters, foreign object material and size, image number, fringe frequency, fringe direction, fringe contrast, brightness fluctuation amplitude, and fringe phase position. These are all quantified parameters, without subjective labels. All data items are standardized and organized using a bidirectional index table structure to support efficient data retrieval and access. Examples of access include: given an angle and temperature, querying the predicted interference intensity of the fringe in the imaging image; or, during detection, inferring possible causes based on the observed fringe frequency in the image. This fringe prior library not only serves as a data lookup table but also as input conditions in subsequent angle scanning constraints, discriminant network training, and dynamic compensation strategy design, providing data support for achieving high-precision diagnosis and interference suppression.

[0068] S002 performs angle scanning and polarization multiplexing under the constraint of the fringe prior library, and combines narrowband spectral sequence acquisition to output a three-dimensional spatiotemporal polarization data volume and complete parameter consistency calibration, so that the acquisition results are aligned with the fringe prior library.

[0069] To ensure strict consistency between the actual acquisition results and the fringe prior library, this step involves angle scanning, polarization multiplexing, and narrowband spectral acquisition under the constraints of the fringe prior library, generating a spatiotemporal polarization three-dimensional data volume. Finally, multi-dimensional calibration is used to achieve consistent parameter matching. The specific steps are as follows:

[0070] Based on the predicted angle and temperature-sensitive ranges in the fringe prior library, a detailed acquisition scheme was developed. The angle scanning process used the normal direction as the zero point, gradually increasing from 0 degrees to 20 degrees in a step of 2 degrees. Within the significant fringe frequency change range indicated by the fringe prior library, it was further subdivided into 1-degree increments. Angle adjustment was achieved using a dual-axis high-precision electrically controlled rotary table equipped with an optical encoder with a resolution of 0.01 degrees Celsius for real-time feedback of angle changes. The temperature was stabilized within the range of 25 to 35 degrees Celsius, maintained through a constant temperature control system and a sealed optical path to avoid additional fringe variations caused by air disturbances and thermal drift. The light source was a wavelength-stabilized LED array; the emitted light was shaped by an integrating homogenizing cavity before entering the backlight module, ensuring uniform intensity distribution of the incident light. The combination of acquired angle and temperature was strictly executed according to the index table of the prior library, ensuring a one-to-one correspondence between the acquisition sequence and the prior library, thus guaranteeing the feasibility of subsequent calibration from the outset. Compared with existing single-angle, single-temperature shooting methods, this step uses a priori library as a constraint framework to ensure that the acquisition path fully covers the potential interference range.

[0071] Polarization multiplexing was implemented under specific angular and temperature conditions for data acquisition. To obtain a complete polarization response, a fixed linear polarizer, a quarter-wave plate, and an electrically controlled rotating analytical polarizer were sequentially arranged in the optical path. Polarization multiplexing was performed according to six polarization states: horizontal linear polarization, vertical linear polarization, positive 45-degree linear polarization, negative 45-degree linear polarization, left-handed circular polarization, and right-handed circular polarization. Each polarization state was subjected to an exposure corresponding to an angular step point, and these polarization states were alternated in a fixed order to avoid sequence errors caused by mechanical hysteresis. The angular scale of the polarization device was calibrated using a standard polarizer before acquisition to ensure that the rotational scale was consistent with the actual polarization direction. The addition of polarization multiplexing enabled the acquired image to contain not only intensity distribution but also polarization characteristic information closely related to fringe formation, thus allowing the differences in fringe response under different polarizations to be reflected in the subsequent data volume. Unlike traditional single-polarization detection methods, this method can reveal the sensitivity characteristics of fringes to polarization direction, improving the data integrity under prior constraints.

[0072] Based on angle scanning and polarization multiplexing combined with narrowband spectral acquisition, a sequence of raw images with multi-dimensional conditional labels is generated. To avoid the fringe averaging effect caused by broadband light sources, three interference filters are used during the acquisition process, corresponding to center wavelengths of 450 nm, 550 nm, and 650 nm, respectively, each with a bandwidth of 10 nm. The filters are mounted on an electronically controlled filter wheel and embedded in the acquisition process according to angle and polarization order, so that each frame corresponds to only a single wavelength and a single polarization state. A telecentric lens is used at the imaging end to ensure that the field of view geometry is not distorted when the angle changes. The rear of the lens is connected to a 16-bit dynamic range complementary metal-oxide-semiconductor sensor with a pixel size of 3.4 micrometers and a resolution of 4096 x 4096, which can completely cover the backlight module's emission surface. Exposure is completed at each angle point for six polarization states and three wavelengths, for a total of eighteen frames, and the exposure time, light source current, filter number, polarization angle, and angle stage reading are recorded in real time. In this way, a complete sequence of image data containing angle, polarization, and spectral dimensions is formed, so that the acquisition results are clearly located in the three-dimensional conditional space, ensuring that they can be directly aligned with the parameter dimensions in the fringe prior library.

[0073] After acquisition, all image sequences were assembled into a spatiotemporal polarization 3D data volume, and parameter consistency calibration was performed to match the fringe prior library. The two spatial dimensions of the 3D data volume are the horizontal and vertical pixels of the image, and the third dimension is the ordered arrangement of angle, polarization state, and wavelength. Before assembly, flat field correction and dark field subtraction were performed to eliminate light source inhomogeneity and sensor noise. Subsequently, angle consistency calibration was performed by calibrating the zero point using the reflection spot generated by a flat metal mirror at the normal angle, and recording the deviation in an angle correction table. Polarization consistency calibration was then performed by measuring the transmission curve using a standard linear polarizer and a standard quarter-wave plate to correct the deviation between the rotating scale and the actual polarization direction. Spectral consistency calibration was performed by measuring the center wavelength drift of the interference filter at different incident angles using a miniature spectrometer, generating a wavelength correction table, and applying it to the data volume label. Radiometric consistency calibration was performed by measuring the relationship between the light source output power and exposure time using a standard reflector and an integrating sphere, generating a brightness correction coefficient, and applying it to the image intensity value. Temperature consistency calibration was performed by using the scattering brightness of a reference glass substrate at 25°C, 30°C, and 35°C as a baseline, establishing a temperature drift coefficient, and correcting it to the corresponding sequence frames. After calibration, each frame of the spatiotemporal polarization 3D data volume carries corrected angle, polarization, wavelength, and temperature labels. Its brightness distribution characteristics are consistent with the corresponding entries in the fringe prior library within the allowable error range, achieving close alignment between the prior and the acquired data.

[0074] S003, a physical-guided generation and discrimination network is trained based on the calibrated spatiotemporal polarization three-dimensional data volume, which separates the fringe driving term from the real defect boundary, outputs the boundary salience map and realizes the robust characterization of key defects;

[0075] To reliably separate the fringe-driven term from the actual defect boundary, this step uses the spatiotemporal polarization 3D data volume that has undergone consistency calibration as the sole training input, and uses the angle, temperature, polarization, and center wavelength corresponding to the fringe pattern given by the fringe prior library as physical guidance quantities to construct a physical-guided generation and discrimination training process for backlight foreign object defects. The specific steps are as follows:

[0076] Constructing training samples and baseline boundaries. Using spatiotemporal polarization 3D data as the original source, spatial slices of 512 x 512 pixels are cut from the effective backlight emission area according to a fixed grid. Each slice corresponds to an 18-frame conditional sequence, which is generated by combining 11 angles, 6 polarizations, and 3 center wavelengths in a predetermined order. To obtain a stable baseline for real defect boundaries, the group with the highest contrast among the 3 center wavelengths is first selected. Combined with industrial microscopic observation and meticulous manual drawing, an initial boundary line is generated. Then, the geometric overlap of this boundary line is compared frame by frame on other frames at the same spatial position. Only line segments with a positional deviation of less than one pixel in at least 12 frames are retained, while pseudo-line segments that swing with angle or flip with polarization are removed, forming a stable boundary set. Subsequently, for each sample, four types of quantitative descriptions corresponding to the 18 frames are extracted according to the fringe prior library index: fringe frequency, fringe direction, fringe contrast, and fringe phase. These four descriptions are projected into a fringe reference map with the same size as the frame, ensuring that each frame has clear physical guidance quantities. To cover both high-brightness and low-brightness areas, a half-step resampling is performed within each spatial slice for regions with brightness peaks above 90%, along with angle, polarization, and center wavelength labels to enrich the local details of the samples. Sample partitioning employs a spatially exclusive approach, ensuring that the same physical location enters only one of the training or validation sets, avoiding information overlap. At this point, the training samples, boundary baselines, and fringe reference maps are strictly paired in both spatial and conditional dimensions. This preparation process directly builds upon the consistency calibration results of the previous step, ensuring a frame-by-frame correspondence between the input data and the fringe prior library.

[0077] A physically constrained generation and discrimination structure is established, clearly defining the input and output. The input consists of eighteen original images, eighteen stripe reference images, eighteen angle value images, eighteen polarization angle images, eighteen center wavelength index images, and one temperature distribution image, stacked in a fixed channel order into the training process. The internal structure employs two parallel branches that intersect multiple times in the later stages. One branch, dominated by the stripe reference image, specifically reconstructs the spatial morphology, intensity distribution, and phase shift of stripes at a given angle, polarization, and center wavelength, outputting a stripe reconstruction image. The other branch focuses on the geometric properties of the real defect boundary, refining high-curvature line segments, acute-angle turns, and long-distance connectivity layer by layer, outputting a boundary saliency map. Information exchange between the two branches is achieved through several fixed-point fusions. The degree of fusion is controlled by the contrast and dominant frequency density in the stripe reference image: when the contrast is high and the dominant frequency is concentrated, the weight of the stripe branch information in the fusion is increased, so that strong stripe areas are interpreted by the stripe branch as much as possible; when the contrast is low and the boundary morphology is stable in cross-frame comparison, the weight of the boundary branch information is increased, making the output closer to the real geometry. The output simultaneously provides a fringe reconstruction map, a boundary salience map, and a fringe residue map. The fringe residue map reflects the periodic brightness fluctuations that still exist in the boundary salience map, which can be used for subsequent constraints. Compared with the traditional approach of segmenting based solely on intensity changes, this structure introduces a reference map aligned frame-by-frame with the physical quantities in the input dimension, and integrates physical knowledge into key nodes of the information flow through a fusion weighting method, thus avoiding mistaking fringes for boundaries at the structural level.

[0078] A training process centered on physical consistency was developed and progressive convergence was implemented. Training was conducted in batches, with each batch containing an 18-frame conditional sequence for that slice. This ensured that a single parameter update could simultaneously address the responses at the same location under different angles, polarizations, and center wavelengths. Each forward pass consisted of two phases: the initial phase froze the boundary branches, driving them solely based on the fringe reference image until the differences between the reconstructed fringe image and the reference image in terms of dominant frequency position, orientation distribution, and contrast amplitude decreased to within a preset threshold; subsequently, the boundary branches were unfrozen, focusing on extracting geometric details orthogonal to the fringes within the context of the interpretation provided by the fringe branches. To ensure the output conformed to physical laws, four types of constraints were introduced and gradually activated as training progressed. The first type was the prior consistency constraint, requiring the reconstructed fringe image to match the reference image in terms of dominant frequency position, fringe orientation, and fringe contrast. Dominant frequency deviation was evaluated using spatial frequency as the scale, orientation deviation using angle difference, and contrast deviation using standardized amplitude difference. The second type is the separation constraint, which requires that the boundary saliency map and the fringe reconstruction map be spatially uncorrelated. By comparing the same positions in eighteen frames, the residual fringes in the boundary saliency map that swing synchronously with the angle or flip with the polarization are suppressed. The third type is the cross-condition consistency constraint, which requires that the boundary saliency maintains strength and position stability when corresponding to the same real boundary position across eighteen frames. The allowed offset is limited to the sub-pixel range, and stability is comprehensively evaluated by the overlapping area and centerline offset. The fourth type is the energy balance constraint, which requires that the intensity fluctuations explained by the fringe branches and the effective contrast retained by the boundary branches be consistent with the original brightness distribution after superposition. It does not allow any arbitrary increase or loss of energy. The training schedule adopts a phased learning rate control. First, the fringe branches are optimized individually to obtain reliable interpretations. Then, the two branches are jointly optimized to ensure coupling stability. Finally, the boundary stability index on the validation set is used as the stopping condition. To alleviate the imbalance of sample types, a defect type ratio list is established so that the proportions of the four types of samples, including dust particles, fiber filaments, bubble spots, and micro-scratches, are similar during the training process. The combination of angle, polarization, and center wavelength is mixed in a fixed ratio to ensure that the structure maintains consistent performance within the complete condition space. Compared to processes that rely solely on manual annotation and intensity thresholds, this training process directly embeds physical quantification descriptions into sample organization, branch collaboration, and convergence monitoring, enabling the output results to be interpretable and traceable across conditions.

[0079] The process involves inference, verification, and robust characterization generation, with the results integrated into subsequent workflows. Eighteen original images of any spatial slice, along with angle, polarization, center wavelength, and temperature labels, are input into the trained structure to obtain the corresponding boundary saliency map and fringe reconstruction map. Consistent fusion is performed on the eighteen boundary saliency maps of the same spatial slice. The fusion weights are determined by the stability of each frame after cross-frame alignment, with stability derived from the overlap ratio of boundary positions across frames and the integrity of endpoint connections. For regions with low stability, entries for adjacent angles and polarizations in the fringe prior library are reviewed to check the physical fringe contrast and dominant frequency drift records at that location. When the records show a significant enhancement of the fringes under this conditional combination, the weight of that frame in the fusion is reduced, and local re-inference is triggered if necessary. Local re-inference fixes the fringe branch parameters, fine-tuning only the boundary branches within a small range to concentrate the response of that region onto the geometric boundary. The fused boundary saliency map is then thinned and pruned to obtain a continuous contour with a single pixel width. The contour is reprojected and verified against the original brightness map to confirm the correspondence between closure, coherence, and brightness transitions. A quality report is then generated, listing the boundary consistency score, fringe interpretation ratio, and energy balance deviation under three center wavelengths, six polarizations, and eleven angles. When all three indicators fall within the preset range, the spatial slice is marked as meeting the robust characterization requirements, and the boundary salience map and fringe reconstruction map are stored in the intermediate dataset used in subsequent steps as input for phase unwrapping and structural sparse decomposition. Compared with traditional methods that rely solely on single-frame intensity or single polarization, the above steps integrate the consistency calibration results, the fringe prior library description, and the physical constraints within the training structure, ensuring that the inference results remain stable and verifiable under multiple conditions. Ultimately, this achieves a reliable characterization of key defect boundaries, providing a highly reliable input basis for subsequent generation of foreign object influence fields and anti-spoofing strategies.

[0080] S004, based on the boundary saliency map, performs stripe phase unwrapping and structural sparse decomposition to generate foreign object influence field and extract the perturbation spectrum most unfavorable to edge features, so that potential interference can be quantitatively described.

[0081] This step, based on the obtained boundary saliency map and the spatiotemporal polarization 3D data volume with completed consistency calibration, sequentially performs fringe phase unwrapping, structural sparse decomposition, foreign object influence field construction, and most unfavorable perturbation spectrum extraction, forming quantitative results that can be directly used for subsequent simulations and closed-loop execution. The specific steps are as follows:

[0082] Guided by the boundary saliency map, fringe phase unwrapping is completed, establishing consistent phase coordinates across angles, polarizations, and wavelengths. For each frame, pixels with response intensities greater than a set threshold in the boundary saliency map are used as seeds, and a guard band is formed by expanding outwards by five to eight pixels along the boundary normal direction to shield the interference of the real geometric boundary on fringe phase estimation. Outside the guard band, a 30x30 pixel sliding window is used to statistically analyze the intensity of local grayscale changes in each direction. The direction with the weakest change is taken as the fringe direction, and the average interval between bright and dark alternations is taken as the dominant frequency. The direction and dominant frequency of entries with the same angle, polarization, and center wavelength in the fringe prior library are used as verification standards. If the deviation exceeds a set threshold, the window is reduced to 20x20 pixels and the statistics are recalculated until they are consistent. Subsequently, a wrapped phase map is generated outside the guard band. Phase anchor points are set in the annular low-response area near the boundary, with a spacing of no more than 20 pixels, as the starting reference for phase propagation. The phase is unfolded along the fringe direction starting from the phase anchor points, and the phase difference between adjacent high-contrast areas is used as a connection in low-contrast areas to ensure phase continuity throughout the entire image. After completing the single-frame phase calculation, the phase difference at adjacent angles and polarizations at the same spatial location is selected as the alignment criterion. The absolute phase of each frame is unified to a common coordinate system with the normal angle, horizontal linear polarization, and center wavelength of 550 nm as zero points, allowing the phase of the same pixel to be directly compared under all conditions. Through dual verification of guard bands, orientation, and dominant frequency, as well as the propagation method supported by phase anchor points, phase unwrapping avoids the risk of misjudging strong boundary gradients as fringe transitions and transforms the boundary saliency map obtained from the previous training step into a reliable phase estimation constraint.

[0083] Based on continuous phase, structural sparse decomposition is performed on the brightness map of each frame, separating the periodic fringe component, geometric boundary component, and gradually varying background component one by one, and performing cross-condition consistency verification. Specifically, based on the fringe direction and dominant frequency obtained in the previous section, a directional bandpass processing method is constructed that can only pass through the dominant frequency and only along the direction to obtain the initial fringe component; using the boundary saliency map as a shape mask, the abrupt brightness components are extracted in the boundary normal direction, and the continuity is forced to be maintained in the boundary tangential direction to obtain the geometric boundary component; the remaining brightness is smoothed with a spatial scale of more than 200 pixels to obtain the gradually varying background component. Subsequently, energy verification is performed on the three components, requiring that the deviation of the three components from the original brightness at each pixel is less than 2% of the full grayscale, and the dominant frequency of the fringe component is checked at the corresponding pixels of adjacent angles and adjacent polarizations to see if it is within the allowable drift range (the dominant frequency change does not exceed one strip per millimeter). If it exceeds the range, the process returns to the phase unwrapping stage, the window is reduced at that position, the direction and dominant frequency are re-estimated, and the components are separated again. To prevent fringe components from swallowing up real boundary details, a protection threshold is set for the region five pixels to either side of the boundary normal direction. The maximum amplitude of the fringe component in this region must not exceed 30% of the boundary gradient at that location. If this value is exceeded, the bandpass intensity is reduced and the weight of the boundary component in this region is increased. Through this chain-like processing of "directional constraint - shape constraint - energy check - cross-condition verification", the fringes and the boundary are physically separated, and the separation quality is fixed with quantifiable indicators to ensure that subsequent calculations based on fringe components are not contaminated by geometric boundaries.

[0084] An object influence field is constructed using fringe and boundary components, and the intensity of the fringe's effect on edge features is quantified in both spatial and conditional dimensions. For each frame, the brightness modulation depth per unit pixel is calculated along the phase gradient direction on the fringe component, and the angle between this depth and the normal direction of the boundary component is determined to obtain the fringe effect efficiency coefficient. The fringe modulation depth is multiplied by the effect efficiency to obtain the instantaneous reduction value of boundary contrast, and the sub-pixel offset of the boundary centerline is calculated using the equivalent displacement corresponding to the fringe phase change. To achieve comparability across wavelengths and polarizations, the reduction values ​​for the three center wavelengths are normalized using the energy scale under the reference reflector and weighted according to the angle between the polarization direction and the fringe direction, giving higher weight to polarizations aligned with the fringe direction. The quantization results are stacked into three-dimensional volume data for eleven angle values, so that each pixel position has an influence curve that varies with the angle. Subsequently, two metrics are summarized point-by-point on the boundary curve with a step size of one pixel: one is the reduction in boundary contrast, which is the difference in grayscale gradient when there are no superimposed stripe components in the boundary normal direction; the other is the boundary position offset, which is the sub-pixel displacement of the boundary centerline in two cases. These two metrics, along with the angle, polarization, and center wavelength labels, are written into the foreign object influence field to form a searchable, superimposed, and replayable quantitative description. To avoid local extreme values ​​misleading subsequent calculations, outlier detection is performed at each spatial location under all conditions. If the weakening value or displacement of a certain condition combination is higher than three times the standard deviation of the mean of the condition set at that location, the sparse decomposition results are checked to see if there is incomplete separation between the stripes and the boundary. If necessary, the boundary protection threshold is increased at that location and the separation is repeated before being written into the influence field.

[0085] Based on the foreign object influence field, the most unfavorable perturbation spectrum for edge features is extracted, and a priority list for simulation and execution is output. At each sampling point of the boundary curve, all angles, polarizations, and center wavelength combinations in the influence field at that point are considered as candidates, and the corresponding fringe components are superimposed onto the boundary components one by one to calculate the comprehensive degradation value. The comprehensive degradation value consists of two linearly weighted parts: the magnitude of boundary contrast reduction and the boundary position offset. The weights are given before the task begins. For example, when positioning accuracy is the primary objective, the weight of the position offset is 70%, and the weight of the magnitude of contrast reduction is 30%; when visually detectable targets are the primary objective, the weight of the magnitude of contrast reduction is 70%, and the weight of the position offset is 30%. After completing the traversal, the combination with the largest overall degradation value is selected, and its dominant fringe frequency, fringe orientation, phase gradient amplitude, polarization direction, angle value, and center wavelength are recorded as the most unfavorable perturbation spectrum for that sampling point. The most unfavorable perturbation spectra of all sampling points along the boundary are connected in spatial order to form a perturbation band. The width of the perturbation band depends on the spatial diffusion range of the fringe effect efficiency. By default, five pixels are taken on each side of the boundary normal. If the phase gradient changes rapidly locally, it is extended to eight pixels on each side. To verify the effectiveness of the perturbation band and perturbation spectrum, playback simulation is performed on all frames with three center wavelengths and six polarizations. The playback results should reproduce the contrast decrease trend and position jitter direction in the actual measurement. If the deviation exceeds the allowable range, the previous segment is returned to recalculate the effect efficiency coefficient and energy normalization factor at the corresponding position until the playback matches the actual measurement. The final output includes three-dimensional volume data of the foreign object influence field, perturbation bands along the boundary, and a list of the most unfavorable perturbation spectra for each sampling point. These results will be directly used as inputs for subsequent virtual relighting simulations and on-site execution strategy generation, and will closely connect this processing with the results of the previous boundary saliency map, forming a complete closed loop from discrimination output to physical quantification.

[0086] S005 injects the foreign object influence field back into the optical model, runs a virtual re-illumination simulation, and generates a false exposure strategy, acquisition trajectory, and temperature control trajectory to provide a validated optimal action set.

[0087] This step embeds the foreign object influence field into optical propagation, performs virtual relighting, and outputs a validated action set to guide online execution and effect verification. The specific steps are as follows:

[0088] An optical propagation link capable of accepting the influence field of foreign objects is established, and the quantization results obtained in the previous step are fully written into each layer of the medium. Using the backlight structure as the framework, the light-emitting diode (LED) surface source, light guide plate, diffuser, brightness enhancement film, and exit surface are defined layer by layer. The thickness, refractive index, and geometric dimensions are within the range confirmed in the consistency calibration stage: light guide plate thickness 1 mm to 1 mm², refractive index 1.58 to 1.579; diffuser thickness 50 μm to 100 μm; brightness enhancement film thickness 40 μm to 80 μm; exit surface roughness between 20 nm and 50 nm. The phase gradient amplitude in the foreign object influence field is written into the local refractive index perturbation map of the light guide plate and diffuser, and a set of refractive index increments varying with angle and polarization is stored at each pixel position. The contrast reduction amplitude is written into the scattering intensity map of the exit surface, and weights are assigned according to the angle between the pixel in the normal direction and the fringe direction, with the weights derived from the efficiency coefficient. To express polarization dependence, an orientation field is established in the brightness enhancement film layer based on the fringe direction recorded in the influence field, so that the incident polarization and the fringe direction form a definite coupling. The light source employs three center wavelength spectral lines: 450 nm, 550 nm, and 650 nm, with a bandwidth of 10 nm. The angular emission distribution uses the actual curve measured during the flat-field correction phase. The imaging side uses a telecentric imaging model with a pixel size of 3.4 micrometers, a quantization depth of 16 bits, and a fixed entrance pupil diameter to ensure consistency between simulated spatial sampling and actual acquisition. Through this writing method, the foreign object influence field is transformed from an "image domain index" into a "propagation domain perturbation," ensuring that every adjustment to the angle, polarization, and center wavelength corresponds to a calculable image plane response—a fundamental difference from methods that merely add or subtract weights in image post-processing.

[0089] Virtual relighting is performed within the propagation path, traversing combinations of angle, polarization, spectrum, and exposure to evaluate fringe suppression and edge enhancement performance frame by frame. Angle coverage ranges from 0 to 20 degrees, with a standard step size of 2 degrees, and a step size of 1 degree for highly sensitive regions indicated by prior knowledge. Six defined polarization states are used: horizontal linear polarization, vertical linear polarization, positive 45-degree linear polarization, negative 45-degree linear polarization, left-handed circular polarization, and right-handed circular polarization. The spectrum uses three center wavelengths with a bandwidth of 10 nanometers. Exposure settings are in three levels: 2 milliseconds, 10 milliseconds, and 20 milliseconds. Each combination is propagated to the image plane within the propagation path to obtain frames consistent with actual acquisition. Three types of metrics are calculated for each frame: 1) contrast enhancement at key boundary locations, calculated as the difference in grayscale gradients under conditions of removing and retaining the influence field; 2) fringe retention ratio, calculated as the percentage of fringe dominant frequency energy in the total energy; and 3) geometric stability, calculated as the maximum sub-pixel offset of the boundary centerline under conditions of cross-angle and cross-polarization. A comprehensive weighting is set based on the detection objective: if the primary objective is to determine the presence or absence of defects, the weight of contrast enhancement is increased; if the primary objective is to accurately locate defects, the weight of geometric stability is increased. All combinations form a scoring matrix, with the matrix index consistent with the prior library index, ensuring that each spatial location has a sortable scoring trajectory within the parameter space. This process closes the causal chain of "influence field—propagation—image plane index," avoiding the loss of the most favorable conditions for the boundary by relying on experience to lock onto a single angle or exposure.

[0090] Based on the scoring matrix, a spoofing exposure strategy, acquisition trajectory, and temperature control trajectory are generated, and time and energy constraints are introduced for feasibility verification. The spoofing exposure strategy consists of three parts: an exposure time table, a gain table, and a sampling priority table. The exposure time table automatically allocates duration according to the combined scores, with higher-scoring combinations receiving longer durations. Double exposures of 2ms and 10ms are introduced in high-brightness areas to prevent saturation. The gain table sets analog gains for the three center wavelengths to ensure consistent effective dynamic range of the image plane. The sampling priority table includes pixels along the previous perturbation band in the resampling list, with a resampling frequency twice that of non-critical areas. The acquisition trajectory follows the principle of angular continuity and polarization alternation: the angle starts at 0 degrees, advances clockwise to 20 degrees, and returns to 0 degrees, with a standard step size of 2 degrees, switching to a 1-degree step in highly sensitive regions. Polarization alternates between linear and circular polarization to reduce the superposition of fringes in the same direction within a short period. The spectrum cycles in the order of 550 nm, 450 nm, and 650 nm to make the mid-wavelength a stable reference. The temperature control trajectory employs a gradually increasing and decreasing stepped curve with a rate of increase / decrease of 0.5 degrees Celsius per minute. Steady-state points are set at 25, 30, and 35 degrees Celsius, each held for 60 seconds to eliminate hysteresis in refractive index and light source output. The disturbance zone coverage area is preferentially sampled at the 30-degree Celsius steady-state point to reduce phase uncertainty introduced by thermal drift. All three trajectories adhere to two hard constraints and one soft constraint: the first hard constraint is that the total duration of a single batch should not exceed 120 seconds to meet production line cycle time; the second hard constraint is that the light source output power should not exceed 80% of its rated capacity to ensure lifespan and thermal stability. The soft constraint aims to cover the top 20% of combinations in the scoring matrix; if this conflictes with the hard constraints, combinations contributing more to geometric stability are prioritized. After verification, a list of candidate action sequences to be validated is generated. Each sequence includes an angle sequence, polarization order, center wavelength order, exposure time per frame, gain per frame, temperature and time table, and dwell time.

[0091] Candidate action sequences are virtually executed and robustness checked. The optimal action set is selected, and an execution list that can be directly deployed is output. Virtual execution replays the candidate sequences one by one according to the actual execution order, generating corresponding images in the propagation path. Three key metrics are summarized for each sequence: minimum boundary contrast, maximum fringe residue ratio, and maximum subpixel offset. The criteria are: minimum boundary contrast is improved by at least 20% compared to the baseline, maximum fringe residue is no more than 5% of the total energy, and maximum subpixel offset is no more than 0.5 pixels. If any metric fails to meet the criteria, the influence field entry for that sequence in the failed frame is read. Based on the most unfavorable perturbation spectrum at that location, a subsequence is regenerated at an adjacent angle or adjacent polarization, and replayed until the three thresholds are met without exceeding the total duration and power constraints. After passing the judgment, a robustness test is performed: a small random perturbation within 0.05 degrees is added to the angle reading, a slight oscillation of 0.2 degrees Celsius is added to the steady-state temperature, and an intensity fluctuation of 1% is added to the light source current. This is repeated three to five times. If all three indicators remain within the threshold, the sequence is confirmed as a member of the optimal action set. The final output list lists the angle advancement direction and step size, polarization switching order and switching time, center wavelength switching order, exposure and gain per frame, temperature target and dwell time, redundant sampling allocation ratio, and the corresponding spatial target area identifier. Unlike the process of collecting data first and then manually fine-tuning, this method completes the dual checks of propagation domain playback and perturbation test before action generation, so that repeatable stripe suppression and boundary enhancement effects can be obtained on the first launch. The quantitative results of the foreign object influence field in the previous step are integrated into the entire chain of strategy generation, feasibility verification, and robustness verification.

[0092] S006, based on the action set generated by the virtual re-illumination simulation, implements closed-loop adaptive optical field polarization temperature coordinated control, uses Bayesian design mechanism to update phase, polarization, spectrum and temperature in real time, and triggers online resampling in key areas to maximize the contrast of real defect boundaries and complete the diagnostic closed loop;

[0093] This step implements closed-loop adaptive control of the optical field, polarization, and temperature based on the action set generated by the virtual reillumination simulation. It also uses a Bayesian design mechanism to update optical parameters in real time, maximizing the contrast of defect boundaries through an adaptive strategy, ultimately achieving efficient diagnostic closure. The specific steps are as follows:

[0094] A dynamic feedback control mechanism is constructed based on the action set generated by virtual relighting simulation. The action set generated by virtual simulation provides the optimal combination of exposure time, angle, polarization, spectrum, and temperature, which is adjusted according to real-time feedback during the detection process. Specifically, after each frame of image acquisition, the contrast of defect boundaries in key areas is extracted in real time and compared with the preset target contrast. If the detected defect boundary contrast is lower than the set value, the optical parameters are adjusted. The adjustment strategy includes optimizing the polarization angle of the light source, precisely controlling the temperature (to avoid thermal drift affecting fringes), and fine-tuning the exposure time. The control mechanism determines the parameters required for the next acquisition by measuring the deviation between the current image and the simulation result, ensuring that fringes and interference in the real-time environment are effectively suppressed, maximizing the accuracy of defect identification. In this way, the light source and acquisition conditions will be adjusted according to the actual image feedback of each frame, rather than relying solely on preset static configurations, thereby improving the flexibility and adaptability of the system.

[0095] The parameter update process is optimized using a Bayesian design mechanism. This mechanism allows for dynamic adjustment of optical parameters such as phase, polarization, spectrum, and temperature based on each detection feedback, calculating the optimal adjustment scheme in real time. Specifically, after each image frame is acquired, the system compares key boundary information in the image with the expected standard to calculate the deviation. The Bayesian inference mechanism adjusts control parameters based on the defect characteristics (such as contrast and positional offset) of the current image feedback, enabling the next round of acquisition to improve diagnostic accuracy in the shortest possible time. By evaluating defect performance and environmental changes in real time, Bayesian design optimizes parameter updates and automatically adjusts the control strategy, ensuring that each round of acquisition adapts to the current conditions. In implementation, the Bayesian model considers the real-time feedback of each image frame and the statistical results of historical data to infer the most suitable combination of optical parameters, further improving image quality and reducing errors caused by mismatched acquisition conditions.

[0096] Online resampling is triggered in key areas to improve the contrast of true defect boundaries. The resampling operation is automatically initiated based on feedback information from each frame, especially when the contrast of defect boundaries is low. The resampling operation in key areas includes two main steps: first, local brightness enhancement is applied to low-contrast areas to make defect boundaries more prominent; second, the stripe contrast in these areas is optimized by adjusting the polarization direction and exposure time to reduce artifact interference caused by temperature changes or other factors. This process ensures that key defect areas are accurately captured during image acquisition and receive enhanced contrast feedback, avoiding the omission of minor defects or misidentification of artifacts. During resampling, exposure is dynamically adjusted not only based on image contrast changes but also by incorporating factors such as boundary changes and temperature fluctuations, ensuring that each resampling session specifically optimizes image quality. Unlike traditional methods that rely solely on a fixed exposure time setting, this implementation adjusts the resampling strategy based on the feedback results of each frame, ensuring that the boundaries of each image are at their sharpest.

[0097] The system executes closed-loop adaptive control, updating and verifying optimization results in real time to complete the entire diagnostic closed loop. After each adjustment and resampling, the image quality is immediately assessed to confirm whether key indicators such as boundary contrast, fringe residue, and geometric stability meet the set requirements. This process is continuously adjusted through a feedback mechanism, and the control parameters are further optimized based on the image quality feedback from each round. If the boundary contrast of a certain region fails to meet the expected value, the system will adjust the phase, spectrum, temperature, or polarization angle based on the inference results of the aforementioned Bayesian design mechanism, triggering acquisition again until satisfactory image quality is finally obtained. Through repeated adaptive adjustments, the system can continuously improve the accuracy and contrast of defect localization, ultimately achieving accurate defect identification in the entire image.

[0098] This invention constructs a diffraction interference model based on electromagnetic time-domain solutions to generate a fringe prior library, accurately quantifying the influence of fringes on brightness distribution. It also incorporates multi-dimensional data such as angle, polarization, and temperature for acquisition and calibration, ensuring consistency between the acquired results and the fringe prior library. Based on this, a physically guided discriminant network is trained to effectively separate fringes from defect boundaries, outputting a boundary saliency map to provide precise support for subsequent phase unwrapping and sparse structural decomposition. Finally, the optimal action set generated by virtual re-illumination simulation is used for closed-loop adaptive control, updating imaging parameters in real time and triggering resampling in key areas, thereby enhancing defect boundary contrast and ensuring accurate defect identification and location. This method solves the problems of artifact effects and inaccurate defect identification in traditional detection methods, improving the quality control level of backlight modules.

[0099] This invention provides, for example Figure 2The mobile phone screen backlight foreign object defect diagnosis system shown includes a stripe modeling module, a data acquisition module, a feature separation module, an interference analysis module, a simulation optimization module, and a closed-loop control module.

[0100] The fringe modeling module constructs a diffraction interference model based on electromagnetic time-domain solutions and generates a fringe prior library based on angle and temperature parameters to quantify the nonlinear influence of fringes on brightness distribution.

[0101] The data acquisition module performs angle scanning and polarization multiplexing under the constraints of the fringe prior library, and combines narrowband spectral acquisition to obtain a three-dimensional spatiotemporal polarization data volume and complete parameter calibration, so that the acquisition results are consistent with the fringe prior library.

[0102] The feature separation module trains a physically guided generation and discrimination network based on the calibrated spatiotemporal polarization 3D data volume to separate the stripe driving term from the defect boundary and output a boundary salience map.

[0103] The interference analysis module performs stripe phase unwrapping and structural sparse decomposition based on the boundary saliency map, generates the foreign object influence field, and extracts the perturbation spectrum of edge features.

[0104] The simulation optimization module injects the foreign object influence field back into the optical model, runs a virtual re-illumination simulation, generates an exposure strategy, acquisition trajectory, and temperature control trajectory, and obtains the optimal action set.

[0105] The closed-loop control module implements closed-loop adaptive optical field polarization temperature control based on the optimal action set, updates imaging parameters in real time using a Bayesian design mechanism, and triggers resampling in key areas to enhance the contrast of defect boundaries and complete the diagnostic closed loop.

[0106] The method for diagnosing foreign object defects in mobile phone screen backlight provided in this embodiment of the invention is implemented through the aforementioned mobile phone screen backlight foreign object defect diagnosis system. For details of the specific methods and processes of the mobile phone screen backlight foreign object defect diagnosis system, please refer to the embodiments of the above-mentioned method for diagnosing foreign object defects in mobile phone screen backlight, which will not be repeated here.

[0107] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for diagnosing foreign object defects in the backlight of a mobile phone screen, characterized in that, Includes the following steps: S001, Construct a diffraction interference model based on electromagnetic time-domain solution, and generate a fringe prior library based on angle parameters and temperature parameters; S002, under the constraint of the fringe prior library, performs angle scanning and polarization multiplexing, and combines narrowband spectral acquisition to obtain a three-dimensional spatiotemporal polarization data volume and complete parameter calibration; S003, a physical-guided generation and discrimination network is trained based on the calibrated spatiotemporal polarization three-dimensional data volume to separate the stripe driving term from the defect boundary and output the boundary salience map; S004, based on the boundary saliency map, stripe phase unwrapping and structural sparse decomposition are performed to generate the foreign object influence field and extract the perturbation spectrum of edge features; S005, reinject the foreign object influence field into the optical model, run virtual re-illumination simulation, generate exposure strategy, acquisition trajectory and temperature control trajectory, and obtain the optimal action set; S006 implements closed-loop adaptive optical field polarization temperature control based on the optimal action set, updates imaging parameters in real time using a Bayesian design mechanism, and triggers resampling in key areas to enhance defect boundary contrast and complete diagnostic closure.

2. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 1, characterized in that, Step S001 includes: A diffraction interference model based on electromagnetic time-domain solution is constructed to simulate the propagation and interaction of incident light inside and at the boundary of a foreign object structure, and a fringe prior library is generated to quantify the nonlinear influence of fringes on brightness distribution. Based on the simulation model, angle and temperature control variables are introduced to simulate the influence of different detection angles and working temperatures on the stripe structure and obtain stripe feature data under different conditions. The electromagnetic field intensity data obtained from the simulation is converted into a two-dimensional brightness image, and the actual impact of stripe perturbation on the brightness distribution is extracted. A stripe prior library is constructed based on image data, and the stripe feature quantities in the brightness image are dually mapped with the corresponding parameters to generate a complete stripe feature data record.

3. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 1, characterized in that, Step S002 includes: Based on the prediction results of the angle and temperature sensitive range in the stripe prior library, a detailed acquisition plan was formulated to perform angle scanning, polarization multiplexing and narrowband spectral acquisition to generate a spatiotemporal polarization three-dimensional data volume. Polarization multiplexing acquisition is performed under angle scanning and temperature conditions, and narrowband spectroscopy is combined with acquisition to obtain a raw image data sequence with complete angle, polarization and wavelength information; Flat field correction and dark field subtraction are performed on the acquired image data, and calibrates are made for angular consistency, polarization consistency, spectral consistency, radiometric consistency and temperature consistency. The calibrated image data is used to generate a spatiotemporal polarization three-dimensional data volume, which is then matched with the fringe prior library to ensure accurate alignment of the data acquisition results.

4. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 1, characterized in that, Step S003 includes: Using the spatiotemporal polarization three-dimensional data volume that has completed consistency calibration as training input, training samples are constructed, and four types of quantitative descriptions, namely fringe frequency, fringe direction, fringe contrast and fringe phase, are extracted to generate a fringe reference image. Guided by the fringe reference map, a physical-guided generation and discrimination network is established, and the fringe driving term is separated from the defect boundary through the network to output the boundary salience map. By training the generated network, stripe phase unwrapping and structural sparse decomposition are performed on the boundary saliency map to generate the foreign object influence field and extract the perturbation spectrum of edge features. Based on the generated foreign object influence field reinjection optical model, a virtual re-illumination simulation was performed to optimize the exposure strategy, acquisition trajectory, and temperature control trajectory, and to trigger resampling in key areas.

5. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 4, characterized in that, During training, the local details of the samples are further optimized by introducing resampling of high-brightness and low-brightness regions in each training sample to ensure comprehensive capture of the stripes. At the same time, the spatial and conditional dimensions of the samples are strictly aligned with the stripe prior library by coordinating angle, polarization and center wavelength labels.

6. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 1, characterized in that, Step S004 includes: Based on the spatiotemporal polarization three-dimensional data volume that has been uniformly calibrated, the fringe phase is unwrapped based on the boundary saliency map, and uniform phase coordinates across angles, polarizations, and wavelengths are established, while avoiding interference from the real geometric boundary on the fringe phase estimation. Based on the unwrapped phase, structural sparse decomposition is performed on the brightness map of each frame to separate the periodic stripe component, geometric boundary component and slowly varying background component, ensuring that the stripe component does not interfere with the extraction of the true boundary. The foreign object influence field is constructed using the fringe component and the boundary component. The influence of the fringes on the boundary contrast and position is quantified, and the influence information is recorded with angle, polarization and wavelength labels. Based on the foreign object influence field, the perturbation spectrum most unfavorable to edge features is extracted and perturbation bands are generated to ensure that all results are consistent with the measured data in the playback simulation, providing input for subsequent simulation and execution strategy generation.

7. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 1, characterized in that, Step S005 includes: The foreign object influence field is embedded into the optical propagation link, defining the optical structure of the LED surface source, light guide plate, diffuser, brightness enhancement film and emission surface, and writing phase gradient and contrast reduction amplitude into the propagation model; Perform virtual relighting simulations, traversing different angles, polarizations, spectral and exposure combinations, and calculate the boundary contrast enhancement, stripe retention ratio and geometric stability of each frame of image to form a scoring matrix; Based on the scoring matrix, a false exposure suppression strategy, acquisition trajectory, and temperature control trajectory are generated, and time and energy constraints are introduced for executability verification. The candidate action sequences are virtually executed and robustness tested. The optimal action set is selected and an execution list that can be directly issued is output, ensuring that the preset contrast, stripe residue and geometric stability indicators are achieved under all conditions.

8. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 7, characterized in that, For each frame of the candidate action sequence, boundary saliency map consistency fusion is performed. The weights are determined by the stability of each frame after cross-frame alignment. The entries of adjacent angles and polarizations in the stripe prior library are backtracked to adjust the local weights to optimize boundary contrast.

9. The method for diagnosing foreign object defects in mobile phone screen backlight according to claim 1, characterized in that, Step S006 includes: Based on the action set generated by the virtual relighting simulation, a dynamic feedback control mechanism is constructed to adjust the combination of exposure time, angle, polarization, spectrum and temperature through real-time feedback. By employing a Bayesian design mechanism, optical parameters are dynamically adjusted based on each image feedback, and the optimal adjustment scheme is calculated in real time. In key areas, online resampling is triggered to optimize stripe contrast through local brightness enhancement and adjustments to polarization direction and exposure time. Closed-loop adaptive control is executed, and the image quality feedback after each adjustment is used to verify whether the boundary contrast, stripe residue, and geometric stability indicators meet the preset requirements.

10. A mobile phone screen backlight foreign object defect diagnosis system, used to implement the mobile phone screen backlight foreign object defect diagnosis method according to any one of claims 1-9, characterized in that, It includes a stripe modeling module, a data acquisition module, a feature separation module, an interference analysis module, a simulation optimization module, and a closed-loop control module. The fringe modeling module constructs a diffraction interference model based on electromagnetic time-domain solutions and generates a fringe prior library based on angle and temperature parameters. The data acquisition module performs angle scanning and polarization multiplexing under the constraints of the fringe prior library, and combines narrowband spectral acquisition to obtain a three-dimensional spatiotemporal polarization data volume and complete parameter calibration. The feature separation module trains a physically guided generation and discrimination network based on the calibrated spatiotemporal polarization 3D data volume to separate the stripe driving term from the defect boundary and output a boundary salience map. The interference analysis module performs stripe phase unwrapping and structural sparse decomposition based on the boundary saliency map, generates the foreign object influence field, and extracts the perturbation spectrum of edge features. The simulation optimization module injects the foreign object influence field back into the optical model, runs a virtual re-illumination simulation, generates an exposure strategy, acquisition trajectory, and temperature control trajectory, and obtains the optimal action set. The closed-loop control module implements closed-loop adaptive optical field polarization temperature control based on the optimal action set, updates imaging parameters in real time using a Bayesian design mechanism, and triggers resampling in key areas to enhance the contrast of defect boundaries and complete the diagnostic closed loop.

Images