A foreign matter detection visual identification system for display panel production

Abstract

The application discloses a foreign matter detection visual identification system for display panel production and relates to the technical field of display production, and comprises a monitoring area division module, a recognition detection module and the like, wherein the monitoring area division module divides a display panel into a plurality of monitoring sub-areas based on a display panel structure diagram and a process flow diagram, each monitoring sub-area is given a unique area number, and each monitoring sub-area contains an area coordinate range, a structure label and a process stage label; the recognition detection module performs image detection on each monitoring sub-area, acquires a local image feature set and performs enhancement processing on the local image feature set, and outputs a group of standardized structure enhancement image recognition data. Through the introduction of a multi-dimensional process parameter sequence modeling mechanism in the mapping module and the integration of a deep time sequence prediction model such as a long short-term memory network, the application can predict the possibility of future foreign matter defects based on the temperature, pressure, exposure power and other parameter change trends of each monitoring sub-area at different process stages.

Description

Technical Field

[0001] This invention relates to the field of display manufacturing technology, specifically to a foreign object detection visual recognition system for display panel production. Background Technology

[0002] With the large-scale application of high-resolution displays and flexible OLEDs, the requirements for cleanliness and process stability in panel manufacturing are constantly increasing. Because display panels involve complex processes such as material lamination, temperature control, and stress release in multiple processing steps including bonding, exposure, and curing, defects such as bubbles, particles, filaments, and film edge lifting are easily introduced. These foreign objects not only affect product yield but may also cause quality problems such as bright spots, shadows, and color spots during subsequent use. Therefore, accurately identifying these foreign objects and tracing their causes has become a critical issue that urgently needs to be addressed in the panel manufacturing industry.

[0003] A search revealed a Chinese patent (publication number: CN1779473A) that discloses a method and apparatus for detecting a flat panel display using a visual model. The steps of this patent include: capturing an image of the panel to be tested; adjusting the captured image and generating a test image and a reference background image; performing a recognition process on the test image and the reference image, including preprocessing, estimating image information, and integrating image information steps; finally, combining various test information to generate a detection image and a detection value; and using the detection value and the detection image as an evaluation of the image quality of the flat panel display to determine the quality of the panel.

[0004] Current technologies primarily employ visual image recognition methods for defect detection in display panels. These methods include, but are not limited to, overhead industrial cameras, multi-angle light source systems, image enhancement algorithms, and artificial intelligence recognition models. While these methods can identify significant foreign objects to a certain extent, their effectiveness remains limited when dealing with optically weak foreign objects (such as translucent particles or edge-curved films) or defects embedded within multi-layered structures. Furthermore, although some systems can record image anomalies, it is difficult to effectively correlate them with specific process steps and equipment numbers, hindering subsequent fault analysis and precise rectification. Therefore, this invention proposes a visual recognition system for foreign object detection in display panel production. Summary of the Invention

[0005] The purpose of this invention is to provide a visual recognition system for foreign object detection in display panel production, so as to solve the problems mentioned in the background art.

[0006] This invention can be achieved through the following technical solution: a visual recognition system for foreign object detection in display panel production, comprising a monitoring area division module, a recognition and detection module, a mapping module, and an analysis and processing module;

[0007] The monitoring area division module is based on the display panel structure diagram and process flow diagram, and divides the display panel into multiple monitoring sub-areas with physical boundaries and functional distinctions according to the hierarchical structure of the display panel.

[0008] Each monitoring sub-region is assigned a unique region number;

[0009] The region numbering is synchronized with the production line rolling coding, and the consistency of image frame and physical panel position is achieved through feature point alignment algorithm;

[0010] When problems such as speed fluctuations and camera acquisition delays occur during production line operation, causing the corresponding position of the monitored sub-region in the image to shift, the system calculates the drift value based on the structural points identified in the image frame and dynamically corrects the region number to achieve self-correction of the number.

[0011] The identification and detection module takes the microstructural differences of the display panel as input, and dynamically configures the image acquisition method and imaging parameters by sensing the physical structural attributes, process state characteristics and potential foreign object types of the area to be tested, so that the potential foreign objects in the monitored sub-area can obtain the optimal response effect in the image.

[0012] The identification and detection module detects the monitored sub-region by multi-view image fusion, multi-angle illumination scanning and disparity map construction, obtains local image feature set, performs enhancement processing, and outputs a set of standardized structure-enhanced image recognition data. The local image feature set is the identifiable visual response features such as brightness, contrast, edge, and texture of the corresponding area in the image formed by foreign objects under specific image acquisition conditions on the surface or structural layer of the display panel.

[0013] The structure-enhanced image recognition data includes: foreign object enhanced images of the corresponding monitoring sub-region, the spatial location of the foreign object within the region, the foreign object morphology classification results, image enhancement method labels, recognition confidence scores, and region number information, to ensure that foreign objects with different attachment methods, different structural layer positions, and different orientation characteristics can be identified with high confidence under a unified enhancement recognition process.

[0014] The structure-enhanced image recognition data improves the visibility of foreign object images while retaining the positioning information of foreign objects in the spatial structure of the display panel, which is used to support subsequent defect classification, spatial coordinate calculation and process traceability analysis.

[0015] The mapping module is used to identify the association between the structure-enhanced image recognition data output by the detection module and the corresponding processing technology, production equipment and historical parameters, and to construct the source tracing path of the foreign object recognition result;

[0016] The mapping module receives the identification results containing the area number information, and based on the mapping relationship between the area number and the production process code, queries and extracts the processing steps, processing equipment numbers, equipment operating time periods and their corresponding key process parameters corresponding to the monitored sub-area.

[0017] Key process parameters include, but are not limited to: coating thickness (μm), exposure energy (mJ / cm²), bonding pressure (kgf), UV irradiation intensity (W / cm²), bonding temperature (°C), and ambient temperature and humidity during processing (%RH, °C). These process parameters are provided by a sensor network or MES system during production and are matched and bound to the monitoring sub-area number via timestamps and panel numbers.

[0018] Furthermore, the mapping module establishes a multi-field mapping table containing "monitoring sub-area number – processing equipment number – process parameter set – timestamp" to accurately map each defect data point in the foreign object identification result to its potential processing source, thus realizing the input basis for equipment responsibility confirmation, process parameter traceability and subsequent analysis and processing;

[0019] The analysis and processing module is used to fuse the structure-enhanced image recognition data output by the identification and detection module with the processing traceability data generated by the mapping module, and to perform defect risk assessment, quality anomaly trend analysis and cross-regional defect correlation judgment.

[0020] A further technical improvement of the present invention is that the method for dynamically correcting the region number in the system includes the following steps:

[0021] S1. Determine if there is a risk of misaligned numbering:

[0022] The system calculates the time difference between the image frame acquisition time and the region number binding time;

[0023] If the theoretical production line movement distance corresponding to the time difference exceeds the preset area length threshold, it is initially determined that the current image may have a misaligned number, and the correction process is initiated.

[0024] S2. Extract structural feature points from the image:

[0025] Edge detection and feature point extraction are performed on the current image frame to identify key structural points in the image, such as display border lines, positioning corners, and fitting frame boundaries, which have strong structural stability and recognizability, as a reference for determining the actual position of the image.

[0026] S3. Match with standard structural templates:

[0027] The system extracts standard points from the standard structure template corresponding to the corresponding region number, compares them with the key structure points extracted from the image, and uses boundary contour matching, feature point pairing or image registration methods (such as affine transformation fitting) to calculate the offset between the key structure points and the standard points, thus obtaining the image drift distance and direction.

[0028] S4. Determine if the numbering needs to be corrected:

[0029] The system sets an offset tolerance threshold. If the actual position of an image structural point deviates from the template position by more than this tolerance threshold, the current image number is considered inconsistent with the actual region. Based on the direction and distance of the offset, the system calculates how many numbers should be adjusted forward or backward.

[0030] S5. Execution number correction:

[0031] Based on the offset judgment result, the system replaces the original number of the current image frame with the corrected region number and records the correction information in the number correction table, including the original number, the new number, the drift distance, the correction time, etc.

[0032] S6. Closed-loop feedback and subsequent processing:

[0033] The correction information is fed back to the monitoring area division module. If the number offset occurs in multiple consecutive image frames, the system will output an alarm message to indicate that there is a synchronization anomaly.

[0034] Meanwhile, the corrected number is used for the attribution and binding of subsequent identification data to ensure that the identification result is correctly assigned to its actual structural region.

[0035] A further technical improvement of the present invention is that the recognition and detection module includes a lateral view detection subunit, an adjustable optical axis deflection rotor unit, and a parallax enhancement recognition subunit;

[0036] The aforementioned sub-units are deployed simultaneously in the system structure, and the system can selectively activate one or more sub-units to perform image acquisition and enhancement processing based on the predicted results of the monitored sub-area structure, process status, or foreign object type.

[0037] A further technical improvement of the present invention is that the side viewing angle detection subunit is used to identify non-front-facing visible defects in the edge area or structural parts of the display panel caused by poor bonding, loose structure or attachment of micro foreign objects, including bonding residue, edge film lifting, structural component bulging and other types.

[0038] Specifically, the side-view detection subunit sets up multiple non-vertically mounted industrial cameras in a specific monitoring sub-area of ​​the display panel. The optical axis of each camera forms a fixed tilt angle with the normal direction of the panel, with the tilt angle ranging from 30° to 60°.

[0039] Deployment locations include, but are not limited to:

[0040] The four-sided border area of ​​the panel (used to inspect the edge sealing structure).

[0041] Polarizing film is attached to the edge (for detecting stacked foreign objects);

[0042] FPC bonding area (used to detect foreign objects at the side outlet);

[0043] Meanwhile, a main top-down camera is positioned directly above the panel, forming a joint acquisition array with the aforementioned side cameras;

[0044] The workflow of the lateral view detection subunit includes:

[0045] Image acquisition: When the target area enters the inspection station, the system simultaneously triggers the top-view camera and all side cameras to acquire images, obtaining top-view images and side images respectively, ensuring that the image timestamps are consistent;

[0046] Spatial registration: Affine transformation and angle correction are performed on the top view image and each side view image. Specifically, this includes: extracting the image edge point set and boundary features in the structural template for preliminary registration; solving the affine matrix using the edge corner points and performing coordinate remapping; and transforming the top view image and each side view image into a region image in a standard reference coordinate system.

[0047] Image fusion and contrast enhancement: Pixel-weighted fusion is performed on the remapped images from each viewpoint to obtain a fused image, which enhances the edge structure signal;

[0048] Local grayscale variability detection is performed on the fused image to generate a preliminary heat map of suspected foreign objects;

[0049] Edge defect feature extraction: In the fused image, a preset corner region ROI is selected, and the following steps are performed: Canny edge extraction; local texture directionality variation detection based on Gabor filter; connection analysis and morphological filtering to extract abnormal contour regions;

[0050] Defect identification: Based on contour integrity, texture interference index, and structural symmetry index, a classification model is constructed to classify and identify the following defect types that may exist in the corner area.

[0051] A further technical improvement of the present invention is that the adjustable optical axis deflector unit is used to improve the visibility of foreign objects with slight tilt, weak adhesion, or low reflectivity on the surface of the display panel in the image. It is particularly suitable for detecting defects that are difficult to image under conventional vertical incident light, such as lightly attached particulate impurities, tilted adhesive residue, and transparent foreign objects. It includes, but is not limited to, the following hardware structures:

[0052] Controllable deflection component: preferably a MEMS micromirror array, a dual-axis servo gimbal or an electric reflector mechanism, used to precisely control the incident direction of light, with a deflection angle range of ±30°;

[0053] Directional light source: emits a focused and controllable directional illumination beam, coupled with a deflection component to ensure local illumination at any angle;

[0054] Optical axis control driver: It has high-precision angle control capability, and the preferred angle adjustment step accuracy is 5°;

[0055] Incident Angle Database: Stores recommended incident angle ranges and preferred angle values ​​associated with the numbers of each monitoring sub-region, obtained based on historical training data, structural morphology, and material reflection characteristics;

[0056] Its workflow includes:

[0057] Incident Angle Configuration Acquisition: When region number i enters the current inspection station, the system reads the set of incident angle settings corresponding to that region from the incident angle database. ;

[0058] Angle scanning acquisition: The controllable deflection component sets the incident angles one by one to the set. various angles ;

[0059] Each angle is set That is, synchronously triggering industrial cameras to acquire images. This continues until all angle images have been acquired, resulting in an image sequence. ={ };

[0060] Image enhancement processing: processing of image sequences Each frame of the image undergoes local contrast enhancement processing (such as CLAHE, gradient boosting), and the response intensity of the foreign object target region is evaluated. ;

[0061] If the strongest response image exists If so, that frame image is selected as the final enhanced image;

[0062] If the response intensity difference of n frames is less than a preset difference threshold, then image stacking and fusion processing is performed to generate the final enhanced image. ;

[0063] Foreign object recognition execution: Image or Input a preset defect identification network, classify and locate weak reflective objects, low-contrast particles, and semi-transparent areas, and output defect type, location, area, and reflection intensity labels.

[0064] A further technical improvement of the present invention is that the parallax enhancement recognition subunit is used to identify slender foreign objects (such as filaments, hair, fine fibers, etc.) with strong directionality and an axis parallel to the main line of sight on the surface of the display panel, solving the problem of detection blind spots caused by weak reflection, blurred shape, and unclear boundaries of such foreign objects in vertical top-down imaging. Specifically:

[0065] The parallax enhancement recognition subunit includes two industrial cameras positioned on either side of the target monitoring sub-area of ​​the display panel. The two cameras are installed at an angle of ±15° to the left and right relative to the normal direction of the panel, so that their lines of sight are at a horizontal angle. The two cameras are equipped with a synchronous triggering device to ensure that the image acquisition time is consistent and to eliminate motion interference. The geometric relationship between the cameras is obtained through the calibration process after installation, including their respective intrinsic parameter matrices and distortion coefficients (for distortion correction), as well as their rotation and translation relationships (for geometric calibration).

[0066] Its workflow includes synchronous image acquisition, disparity map construction, edge difference enhancement, and orientation contour fitting and feature extraction;

[0067] Synchronous image acquisition:

[0068] When the detection station reaches the target area, the two cameras belonging to the parallax enhancement recognition subunit respectively acquire images of the currently monitored sub-area, denoted as the left view. and right view Furthermore, the parallax enhancement recognition subunit is for the left view. and right view Image distortion correction and geometric calibration registration were performed to obtain the left view. and right view Paired images;

[0069] Geometric calibration and registration refers to the system registering the left view based on the mounting angles and relative positions (i.e., extrinsic parameters) between cameras. and right view Geometric calibration and registration transformations are performed separately to align the two images in spatial structure, ensuring that the pixels of the target in the image correspond completely in the vertical direction when calculating disparity, while only retaining the disparity difference in the horizontal direction, thus forming a standard image pairing relationship;

[0070] Disparity map construction:

[0071] For the left view and right view Perform a stereo matching algorithm based on block matching or semi-global matching;

[0072] Output a parallax image D, where the gray value of each pixel in the parallax image D represents the horizontal offset of that pixel between the two images; the parallax image D can highlight structural features in the image that have slight variations in height or texture protrusions, especially the outline of axially elongated objects;

[0073] Enhanced edge differences:

[0074] Perform edge enhancement processing (such as Sobel operator + nonmaximum suppression) on the disparity image D, and compare it with the original left view. Image difference maps are overlaid. And perform threshold enhancement processing to obtain an enhanced image that further highlights the edges of the filamentous foreign object;

[0075] Orientation contour fitting and feature extraction:

[0076] Extract continuous, elongated boundary regions from the enhanced image and calculate their contour edge point set;

[0077] The main orientation vector, length, and center coordinates of the foreign object are calculated using a sub-pixel-level orientation fitting algorithm.

[0078] Obtain the spatial orientation properties of the foreign object, including its position coordinates (x, y), orientation angle θ, and length L.

[0079] A further technical improvement of the present invention is that the method for constructing the source tracing path by the mapping module includes the following steps:

[0080] Z1, Regional Process Information Analysis:

[0081] The mapping module first extracts the "region number" information from the structure-enhanced image recognition data, and then queries the multi-field mapping table for the corresponding production process information, including:

[0082] The type of process involved (e.g., surface cleaning, bonding, exposure, dispensing, curing);

[0083] The corresponding process segment sequence number;

[0084] Unique identifiers (such as equipment IDs) for critical equipment used in the current process;

[0085] The process parameter template ID corresponding to this process;

[0086] Z2. Production equipment parameter binding:

[0087] Based on the equipment ID and process time interval obtained in step Z1, the module mapping module calls the log data interface in the field manufacturing execution system (MES) or equipment control platform (such as PLC, SCADA) to extract the following equipment operating parameter information:

[0088] Sensor parameters such as temperature, humidity, pressure, adhesion force, exposure power, and operating current during the sampling period;

[0089] Equipment operating status codes, fault codes, and alarm event records;

[0090] Configuration parameters such as the current device software version and executable program number;

[0091] Z3. Historical Defect Record Retrieval:

[0092] The mapping module further retrieves historical records related to the current detection area from the defect history database, including:

[0093] Have similar foreign object types been identified in previous batches from the same equipment, process section, and area number?

[0094] Are there periods of high defect incidence corresponding to parameter fluctuations in this process segment?

[0095] Are there any localized abnormal clusters caused by process switching, equipment maintenance, or changes in raw material batches?

[0096] Z4. Source tracing path generation and output structure:

[0097] Based on the results of steps Z1, Z2, and Z3, the mapping module ultimately generates a source tracing path record. This record is output in a structured data format and contains at least the following fields:

[0098] Detection area number;

[0099] Foreign object type, location, and identification confidence level;

[0100] Corresponding processing technology name, process number, and processing time window;

[0101] Device ID, device status, and parameter snapshots used;

[0102] Historical anomaly correlation score and source tracing label.

[0103] This source tracing path is used by subsequent analysis and processing modules for defect classification, root cause identification, process optimization suggestions, and equipment maintenance early warnings.

[0104] A further technical improvement of the present invention lies in: in each monitoring sub-region During the corresponding production stage, the system collects key process parameters related to that stage in real time, including temperature (degrees Celsius, sampling frequency 1 Hz); pressure (Newtons); exposure power (milliwatts per square centimeter); adhesive layer thickness (micrometers); and action duration (milliseconds). Specifically, the system samples each monitoring sub-region at fixed time intervals. The key process parameters are determined using a sliding window mechanism, which takes the data from the T most recent time steps to construct a multidimensional process parameter sequence. ;

[0105] Furthermore, the system organizes key process parameters into a multidimensional process parameter sequence in the form of a T*n matrix according to time sequence. ;

[0106] The system integrates a time series prediction model based on a long short-term memory neural network in the mapping module;

[0107] When the system is working, it receives monitoring data from sub-regions. Latest multidimensional process parameter sequence The predicted probability of defect occurrence is obtained by using the time series prediction model output. ;

[0108] And the system will predict the probability. Compared with the preset risk threshold, if the predicted probability If the risk exceeds the risk threshold, the corresponding monitoring sub-area will be... The prediction results identified it as a potentially high-risk area;

[0109] When a certain monitoring sub-area Once identified as a potentially high-risk area, the system sends a scheduling instruction to the identification and detection module to improve the image acquisition quality of the area. This includes simultaneously activating the lateral view detection subunit, the adjustable optical axis deflector unit, and the parallax enhancement identification subunit to achieve collaborative imaging under multi-angle, multi-parallax, and multi-light conditions, thereby enhancing the visibility and structural resolution of foreign objects in the area.

[0110] Furthermore, each of the above sub-units is a physically independent module in the system structure. Parallel triggering and image data synchronization are achieved through a unified scheduling controller, and the output results are fused according to the timestamp alignment mechanism.

[0111] A further technical improvement of the present invention is that the mapping module monitors sub-regions. For nodes, construct a directed weighted graph structure G = (V, E), where:

[0112] , indicating all monitored sub-regions;

[0113] , representing the process dependency edges between regions, where, This indicates the monitoring sub-region. To the monitoring sub-area The strength of dependence;

[0114] In this embodiment, It is obtained through statistical analysis of the co-occurrence frequency of historical defects, configuration based on the experience of process engineering experts, or modeling of the concentration trend of foreign objects based on the output of the image recognition module;

[0115] For each downstream region Calculate its cumulative risk score The expression is as follows:

[0116] ;

[0117] In the formula, The predicted probability of the defect occurring; For region-dependent strength; To monitor sub-regions From the time it was identified as a high-risk area until the current assessment Time delay; This is the time decay coefficient, and its specific value can be adjusted according to the time sensitivity of the actual process and the characteristics of historical fault data.

[0118] If the cumulative risk score If the value exceeds the preset threshold, the downstream monitoring sub-region will be... It is marked as a secondary risk area and pushed to the identification and detection module, triggering the preset policy level upgrade.

[0119] Compared with the prior art, the present invention has the following beneficial effects:

[0120] This invention introduces a multi-dimensional process parameter sequence modeling mechanism into the mapping module and integrates deep temporal prediction models such as Long Short-Term Memory (LSTM) networks. Based on the changing trends of parameters such as temperature, pressure, and exposure power in different process stages of each monitored sub-region, it can predict the likelihood of future foreign object defects. This mechanism not only supports real-time scoring and risk threshold warnings but also adjusts the acquisition level of the detection unit according to the prediction results, enabling proactive upgrades in image acquisition and improving the defect detection rate.

[0121] Furthermore, this invention constructs a process dependency graph, treating each monitored sub-region of the display panel as a graph node, and defines the dependency edges and weights between regions by combining historical defect co-occurrence patterns and process sequence information. When a high-risk indicator appears in a certain upstream region, the system treats that node as the risk source, propagates the risk score along the dependency edges to the relevant regions, and dynamically adjusts the image acquisition strategy and analysis accuracy based on the score results, forming a risk response system for chain-like defect prediction and prevention.

[0122] On the other hand, combining the structure-enhanced image recognition data output by the identification and detection module, the mapping module in this invention further constructs a multi-field mapping table of "region number - process parameters - equipment information - timestamp" to ensure that every identified defect can be traced back to a specific equipment number and process stage. By adding defect risk prediction fields and transmission path information to the mapping table, it is possible to record, analyze, and assign responsibility for abnormal behavior, effectively supporting subsequent production optimization, quality control, and equipment maintenance. Attached Figure Description

[0123] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0124] Figure 1 This is the system logic diagram of the present invention. Detailed Implementation

[0125] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0126] Example 1

[0127] Please see Figure 1 As shown, the present invention provides a visual recognition system for foreign object detection in display panel production, including a monitoring area division module, a recognition and detection module, a mapping module and an analysis and processing module;

[0128] The monitoring area division module is based on the display panel structure diagram and process flow diagram. According to the hierarchical structure of the display panel (such as TFT substrate area, color filter area, bonding area, and encapsulation adhesive edge area), it is divided into multiple monitoring sub-areas with physical boundaries and functional distinctions.

[0129] Each monitoring sub-region is assigned a unique region number and includes the region's coordinate range, structural labels (such as glass / OCA / frame), and process stage labels (such as exposure / lamination / encapsulation).

[0130] The region numbering is synchronized with the production line rolling coding, and the consistency between the image frame and the physical panel position is achieved through feature point alignment algorithms (such as image edge detection and round dot target recognition).

[0131] When problems such as speed fluctuations and camera acquisition delays occur during production line operation, causing the corresponding position of the monitored sub-region in the image to shift, the system calculates the drift value based on the structural points identified in the image frame and dynamically corrects the region number to achieve self-correction of the number.

[0132] The system's method for dynamically correcting area codes includes the following steps:

[0133] S1. Determine if there is a risk of misaligned numbering:

[0134] The system calculates the time difference between the image frame acquisition time and the region number binding time;

[0135] If the theoretical production line movement distance corresponding to the time difference exceeds the preset region length threshold (e.g., exceeding the length of a sub-region), it is initially determined that the current image may have a numbering misalignment, and the correction process is initiated.

[0136] S2. Extract structural feature points from the image:

[0137] Edge detection and feature point extraction are performed on the current image frame to identify key structural points in the image, such as display border lines, positioning corners, and fitting frame boundaries, which have strong structural stability and recognizability, as a reference for determining the actual position of the image.

[0138] S3. Match with standard structural templates:

[0139] In this embodiment, the standard structural template extracts the structural pattern of each monitoring sub-region from the panel design file (such as CAD drawings, Gerber files, or process flow diagrams) to form a standard geometric outline, and the standard structural template includes at least the elements in Table 1 below:

[0140] Table 1

[0141]

[0142] The system extracts standard points from the standard structure template corresponding to the corresponding region number, compares them with the key structure points extracted from the image, and uses boundary contour matching, feature point pairing or image registration methods (such as affine transformation fitting) to calculate the offset between the key structure points and the standard points, thus obtaining the image drift distance and direction.

[0143] S4. Determine if the numbering needs to be corrected:

[0144] The system sets an offset tolerance threshold. If the actual position of an image structural point deviates from the template position by more than this tolerance threshold, the current image number is considered inconsistent with the actual region. Based on the direction and distance of the offset, the system calculates how many numbers should be adjusted forward or backward.

[0145] S5. Execution number correction:

[0146] Based on the offset judgment result, the system replaces the original number of the current image frame with the corrected region number and records the correction information in the number correction table, including the original number, the new number, the drift distance, the correction time, etc.

[0147] S6. Closed-loop feedback and subsequent processing:

[0148] The correction information is fed back to the monitoring area division module. If the number offset occurs in multiple consecutive image frames, the system will output an alarm message to indicate that there is a synchronization anomaly.

[0149] Meanwhile, the corrected number is used for the attribution binding of subsequent identification data to ensure that the identification result is correctly assigned to its actual structural region;

[0150] For example, taking the fifth sub-region in the upper left corner of the panel as an example, its structure template contains the following data:

[0151] Area code: N5;

[0152] Structural features: rectangular window + beveled positioning lines;

[0153] Standard feature point coordinates: {(120,45),(260,45),(260,160),(120,160)};

[0154] Image orientation angle: 0°;

[0155] Resolution: 3.5µm / pixel;

[0156] When the edge point of the border in the detected image frame deviates from the coordinate range by more than ±8 pixels, the system automatically triggers a number correction.

[0157] The identification and detection module takes the microstructural differences of the display panel as input, and dynamically configures the image acquisition method and imaging parameters by sensing the physical structural attributes, process state characteristics and potential foreign object types of the area under test, so that the potential foreign objects in the monitored sub-area can obtain the best response effect in the image.

[0158] The recognition and detection module acquires a set of local image features through multi-view image fusion, multi-angle illumination scanning, and disparity map construction, performs enhancement processing on these features, and outputs a set of standardized structure-enhanced image recognition data.

[0159] Local image feature set refers to the identifiable visual response features such as brightness, contrast, edge, and texture of the corresponding area in an image formed by foreign objects (such as particles, filaments, bubbles, and residual adhesive) on the surface or structural layers of a display panel under specific image acquisition conditions.

[0160] The structure-enhanced image recognition data includes: enhanced images of foreign objects corresponding to the monitored sub-regions, the spatial location of the foreign object within the region, the foreign object morphology classification results, image enhancement method labels, recognition confidence scores, and region numbering information. This ensures that foreign objects with different attachment methods, different structural levels, and different orientation characteristics can be identified with high confidence under a unified enhancement recognition process. In other words, when identifying foreign objects, the system can provide high-precision judgment results based on image data, meaning the system's recognition process has high accuracy and reliability. Specifically:

[0161] Accuracy: The recognition results are highly consistent with the actual situation, and the probability of misidentification and omission is low.

[0162] Reliability: The system has a high degree of trust in the results and a high confidence score, which usually means that the prediction results of the recognition model have strong stability and are not easily affected by noise or errors.

[0163] Structural enhancement image recognition data improves the visibility of foreign object images while retaining the positioning information of foreign objects in the spatial structure of the display panel, which is used to support subsequent defect classification, spatial coordinate calculation and process traceability analysis.

[0164] The recognition and detection module includes a lateral view detection subunit, an adjustable optical axis deflection rotor unit, and a parallax enhancement recognition subunit;

[0165] The above-mentioned sub-units are deployed simultaneously in the system structure, and the system can selectively activate one or more sub-units to perform image acquisition and enhancement processing based on the prediction results of the structure, process status or foreign object type of the monitored sub-area;

[0166] The side-view detection subunit is used to identify non-front-facing defects in the edge area or structural parts of the display panel caused by poor bonding, loose structure or attachment of micro-foreign objects, including types such as adhesive residue, edge film lifting, and structural component bulging.

[0167] Specifically, the side-view detection subunit sets up multiple non-vertically mounted industrial cameras in a specific monitoring sub-area of ​​the display panel. The optical axis of each camera forms a fixed tilt angle with the normal direction of the panel, with the tilt angle ranging from 30° to 60°, preferably 45°.

[0168] Deployment locations include, but are not limited to:

[0169] The four-sided border area of ​​the panel (used to inspect the edge sealing structure).

[0170] Polarizing film is attached to the edge (for detecting stacked foreign objects);

[0171] FPC bonding area (used to detect foreign objects at the side outlet);

[0172] Meanwhile, a main overhead camera is positioned directly above the panel, forming a joint acquisition array with the aforementioned side cameras;

[0173] The workflow of the lateral view detection subunit includes:

[0174] Image acquisition: When the target area enters the inspection station, the system simultaneously triggers the top-view camera and all side cameras to acquire images, obtaining top-view images and side images respectively, ensuring that the image timestamps are consistent;

[0175] Spatial registration: Affine transformation and angle correction are performed on the top view image and each side view image. Specifically, this includes: extracting the image edge point set and boundary features in the structural template for preliminary registration; solving the affine matrix using the edge corner points and performing coordinate remapping; and transforming the top view image and each side view image into a region image in a standard reference coordinate system.

[0176] Image fusion and contrast enhancement: Pixel-weighted fusion is performed on the remapped images from each viewpoint to obtain a fused image, which enhances the edge structure signal;

[0177] Local grayscale variability detection is performed on the fused image to generate a preliminary heat map of suspected foreign objects;

[0178] Edge defect feature extraction: In the fused image, a preset corner region ROI is selected, and the following steps are performed: Canny edge extraction; local texture directionality variation detection based on Gabor filter; connection analysis and morphological filtering to extract abnormal contour regions;

[0179] Defect identification: Based on contour integrity, texture interference index, and structural symmetry index, a classification model is constructed to classify and identify the following defect types that may exist in the corner area. In this embodiment, the identification results are shown in Table 2:

[0180] Table 2

[0181]

[0182] Adjustable optical axis deflector units are used to improve the visibility of foreign objects with slight tilt, weak adhesion, or low reflectivity on the surface of display panels. They are particularly suitable for detecting defects that are difficult to image under conventional vertical incident light, such as lightly attached particulate impurities, tilted adhesive residue, and transparent foreign objects. Examples include, but are not limited to, the following hardware structures:

[0183] Controllable deflection component: preferably a MEMS micromirror array, a dual-axis servo gimbal or an electric reflector mechanism, used to precisely control the incident direction of light, with a deflection angle range of ±30°;

[0184] Directional light source: emits a focused and controllable directional illumination beam, coupled with a deflection component to ensure local illumination at any angle;

[0185] Optical axis control driver: It has high-precision angle control capability, and the preferred angle adjustment step accuracy is 5°;

[0186] Incident Angle Database: Stores recommended incident angle ranges and preferred angle values ​​associated with the numbers of each monitoring sub-region, obtained based on historical training data, structural morphology, and material reflection characteristics;

[0187] Its workflow includes:

[0188] Incident Angle Configuration Acquisition: When region number i enters the current inspection station, the system reads the set of incident angle settings corresponding to that region from the incident angle database. ;

[0189] Angle scanning acquisition: The controllable deflection component sets the incident angles one by one to the set. various angles ;

[0190] Each angle is set That is, synchronously triggering industrial cameras to acquire images. This continues until all angle images have been acquired, resulting in an image sequence. ={ };

[0191] Image enhancement processing: processing of image sequences Each frame of the image undergoes local contrast enhancement processing (such as CLAHE, gradient boosting), and the response intensity of the foreign object target region is evaluated. ;

[0192] If the strongest response image exists If so, that frame image is selected as the final enhanced image;

[0193] If the response intensity difference of n frames is less than a preset difference threshold, then image stacking and fusion processing is performed to generate the final enhanced image. ;

[0194] Foreign object recognition execution: Image or Input a preset defect identification network, classify and locate weak reflective objects, low-contrast particles, and semi-transparent areas, and output defect type, location, area, and reflection intensity labels;

[0195] The defect identification network uses a convolutional neural network (CNN) model trained through deep learning. It is trained under supervised supervision using a dataset of display panel defect images collected by the system itself. The dataset covers the following types of foreign objects:

[0196] Adherent particulate foreign matter (glass shards, dust, impurities);

[0197] Semi-transparent adhesive residue spots;

[0198] Micro-reflective objects visible under oblique incidence;

[0199] Signs of membrane material peeling off;

[0200] Each image in the dataset contains manually annotated foreign object location boxes, category labels, defect outlines, and visual scores.

[0201] The parallax enhancement recognition subunit is used to identify slender foreign objects (such as filaments, hair, and fine fibers) on the surface of the display panel that are highly directional and have an axis parallel to the main line of sight. This addresses the detection blind zone problem caused by weak reflection, blurred shape, and indistinct boundaries of such foreign objects in vertical top-view imaging. Specifically:

[0202] The parallax enhancement recognition subunit includes two industrial cameras positioned on either side of the target monitoring sub-area of ​​the display panel. The two cameras are installed at an angle of ±15° to the left and right relative to the normal direction of the panel, so that their lines of sight are at a horizontal angle. The two cameras are equipped with a synchronous triggering device to ensure that the image acquisition time is consistent and to eliminate motion interference. The geometric relationship between the cameras is obtained through the calibration process after installation, including their respective intrinsic parameter matrices and distortion coefficients (for distortion correction), as well as their rotation and translation relationships (for geometric calibration).

[0203] Its workflow includes synchronous image acquisition, disparity map construction, edge difference enhancement, and orientation contour fitting and feature extraction;

[0204] Synchronous image acquisition:

[0205] When the detection station reaches the target area, the two cameras belonging to the parallax enhancement recognition subunit respectively acquire images of the currently monitored sub-area, denoted as the left view. and right view Furthermore, the parallax enhancement recognition subunit is for the left view. and right view Image distortion correction and geometric calibration registration were performed to obtain the left view. and right view Paired images;

[0206] In this embodiment, image distortion correction refers to acquiring images using a standard calibration board after camera installation to obtain distortion coefficients and intrinsic parameters, which are then used to correct imaging distortion caused by the lens. Specifically, the parallax enhancement recognition subunit, based on the calibration results of each camera (including the intrinsic parameter matrix and distortion coefficients), adjusts the left view... and right view The distorted areas in the image are remapped to eliminate radial and tangential distortions caused by the lens, generating a distortion-corrected image.

[0207] Geometric calibration and registration refers to the system registering the left view based on the mounting angles and relative positions (i.e., extrinsic parameters) between cameras. and right view Geometric calibration and registration transformations are performed separately to align the two images in spatial structure, ensuring that the pixels of the target in the image correspond completely in the vertical direction when calculating disparity, while only retaining the disparity difference in the horizontal direction, thus forming a standard image pairing relationship;

[0208] Specifically, the image distortion correction and geometric calibration registration adopt any mature method in the existing technology, such as Zhang Zhengyou camera calibration method, OpenCV stereo image correction and other publicly available and mature engineering implementation methods;

[0209] Disparity map construction:

[0210] For the left view and right view Perform a stereo matching algorithm based on block matching or semi-global matching;

[0211] Output a parallax image D, where the gray value of each pixel in the parallax image D represents the horizontal offset of that pixel between the two images; the parallax image D can highlight structural features in the image that have slight variations in height or texture protrusions, especially the outline of axially elongated objects;

[0212] Enhanced edge differences:

[0213] Perform edge enhancement processing (such as Sobel operator + nonmaximum suppression) on the disparity image D, and compare it with the original left view. Image difference maps are overlaid. And perform threshold enhancement processing to obtain an enhanced image that further highlights the edges of the filamentous foreign object;

[0214] Orientation contour fitting and feature extraction:

[0215] Extract continuous, elongated boundary regions from the enhanced image and calculate their contour edge point set;

[0216] The main orientation vector, length, and center coordinates of the foreign object are calculated using subpixel-level orientation fitting algorithms (such as least squares line fitting or ellipse major axis extraction).

[0217] Obtain the spatial orientation properties of the foreign object, including its position coordinates (x, y), orientation angle θ, and length L;

[0218] The mapping module is used to identify the association between the structure-enhanced image recognition data output by the detection module and the corresponding processing technology, production equipment and historical parameters, and to construct the source tracing path of the foreign object recognition results;

[0219] The mapping module receives the identification results containing the area number information, and based on the mapping relationship between the area number and the production process code, queries and extracts the processing steps, processing equipment numbers, equipment operating time periods and their corresponding key process parameters corresponding to the monitored sub-area.

[0220] Key process parameters include, but are not limited to: coating thickness (μm), exposure energy (mJ / cm²), bonding pressure (kgf), UV irradiation intensity (W / cm²), bonding temperature (°C), and ambient temperature and humidity during processing (%RH, °C). These process parameters are provided by a sensor network or MES system during production and are matched and bound to the monitoring sub-area number via timestamps and panel numbers.

[0221] Furthermore, the mapping module establishes a multi-field mapping table containing "monitoring sub-area number – processing equipment number – process parameter set – timestamp" to accurately map each defect data point in the foreign object identification result to its potential processing source, thus realizing the input basis for equipment responsibility confirmation, process parameter traceability and subsequent analysis and processing;

[0222] The method for constructing the source tracing path in the mapping module includes the following steps:

[0223] Z1, Regional Process Information Analysis:

[0224] The mapping module first extracts the "region number" information from the structure-enhanced image recognition data, and then queries the multi-field mapping table for the corresponding production process information, including:

[0225] The type of process involved (e.g., surface cleaning, bonding, exposure, dispensing, curing);

[0226] The corresponding process segment sequence number;

[0227] Unique identifiers (such as equipment IDs) for critical equipment used in the current process;

[0228] The process parameter template ID corresponding to this process;

[0229] Z2. Production equipment parameter binding:

[0230] Based on the equipment ID and process time interval obtained in step Z1, the module mapping module calls the log data interface in the field manufacturing execution system (MES) or equipment control platform (such as PLC, SCADA) to extract the following equipment operating parameter information:

[0231] Sensor parameters such as temperature, humidity, pressure, adhesion force, exposure power, and operating current during the sampling period;

[0232] Equipment operating status codes, fault codes, and alarm event records;

[0233] Configuration parameters such as the current device software version and executable program number;

[0234] Z3. Historical Defect Record Retrieval:

[0235] The mapping module further retrieves historical records related to the current detection area from the defect history database, including:

[0236] Have similar foreign object types been identified in previous batches from the same equipment, process section, and area number?

[0237] Are there periods of high defect incidence corresponding to parameter fluctuations in this process segment?

[0238] Are there any localized abnormal clusters caused by process switching, equipment maintenance, or changes in raw material batches?

[0239] Z4. Source tracing path generation and output structure:

[0240] Based on the results of steps Z1, Z2, and Z3, the mapping module ultimately generates a source tracing path record. This record is output in a structured data format and contains at least the following fields:

[0241] Detection area number;

[0242] Foreign object type, location, and recognition confidence (from structure-enhanced image recognition data);

[0243] Corresponding processing technology name, process number, and processing time window;

[0244] Device ID, device status, and parameter snapshots used;

[0245] Historical anomaly correlation score and source tracing label.

[0246] This source tracing path is used by subsequent analysis and processing modules for defect classification, root cause identification, process optimization suggestions, and equipment maintenance warnings.

[0247] The analysis and processing module is used to fuse the structure-enhanced image recognition data output by the identification and detection module with the processing traceability data generated by the mapping module, and to perform defect risk assessment, quality anomaly trend analysis, and cross-regional defect correlation judgment. It includes the following processing functions:

[0248] Early warning mechanism for potential defects under process fluctuations:

[0249] The analysis and processing module continuously receives process parameter data streams from the mapping module and sets historical baseline values ​​and deviation tolerance thresholds for each type of parameter. If no visible foreign matter is currently identified in a certain monitoring sub-area, but the corresponding processing parameters show an abnormal fluctuation trend (such as continuous temperature deviation, short-term surge in exposure energy, etc.), the system marks the area as a "potential abnormal area" and outputs a warning sign to prompt for subsequent stricter testing or shutdown verification.

[0250] Defect common cause relationship identification mechanism:

[0251] For foreign objects identified in multiple different areas or continuous panels, the analysis and processing module uses methods such as cluster analysis, structural similarity matching, and equipment path overlap analysis to determine whether they may be contaminated from the same source or abnormal from the same batch of equipment. If multiple conditions such as similar location, similar morphological characteristics, and consistent processing equipment are met, the system will merge and mark the above defect data as "common cause defects" and locate the suspected responsible equipment or material batch.

[0252] Numbering mismatch correction and data closed-loop feedback mechanism:

[0253] Under high-speed operation of the production line, if the image number and the area number are misaligned due to image acquisition delay or transmission error, the analysis and processing module will detect the numbering abnormality in real time based on the image feature point re-identification, number sliding window comparison and timeline consistency verification mechanism, and automatically correct the numbering of the incorrectly bound image data to ensure the unique matching relationship between image-area-equipment-parameter. At the same time, the numbering correction information will be fed back to the monitoring area division module to form a data closed loop inherent in the system.

[0254] The final output of the analysis and processing module includes: risk level label for each defect, judgment result of whether it belongs to a common cause defect cluster, identification signal of whether an early warning has been triggered, and numbered data of corrected defects, which can be used for subsequent production optimization, equipment inspection and quality accountability.

[0255] Example 2

[0256] A visual recognition system for foreign object detection in display panel production includes a monitoring area division module, a recognition and detection module, a mapping module, and an analysis and processing module;

[0257] The monitoring area division module is based on the display panel structure diagram and process flow diagram. According to the hierarchical structure of the display panel (such as TFT substrate area, color filter area, bonding area, and encapsulation adhesive edge area), it is divided into multiple monitoring sub-areas with physical boundaries and functional distinctions.

[0258] Each monitoring sub-region is assigned a unique region number and includes the region's coordinate range, structural labels (such as glass / OCA / frame), and process stage labels (such as exposure / lamination / encapsulation).

[0259] The region numbering is synchronized with the production line rolling coding, and the consistency between the image frame and the physical panel position is achieved through feature point alignment algorithms (such as image edge detection and round dot target recognition).

[0260] When problems such as speed fluctuations and camera acquisition delays occur during production line operation, causing the corresponding position of the monitored sub-region in the image to shift, the system calculates the drift value based on the structural points identified in the image frame and dynamically corrects the region number to achieve self-correction of the number.

[0261] The identification and detection module takes the microstructural differences of the display panel as input, and dynamically configures the image acquisition method and imaging parameters by sensing the physical structural attributes, process state characteristics and potential foreign object types of the area under test, so that the potential foreign objects in the monitored sub-area can obtain the best response effect in the image.

[0262] The recognition and detection module enhances the foreign object image signal through multi-view image fusion, multi-angle illumination scanning, disparity map construction, and edge feature fitting, and outputs a set of standardized structure-enhanced image recognition data. The structure-enhanced image recognition data includes: the foreign object enhanced image of the corresponding monitoring sub-region, the spatial location of the foreign object in the region, the foreign object morphology classification result, the image enhancement method label, the recognition confidence score, and the region number information, so as to ensure that foreign objects with different attachment methods, different structural levels and different orientation characteristics can be recognized with high confidence under a unified enhancement recognition process.

[0263] Structural enhancement image recognition data improves the visibility of foreign object images while retaining the positioning information of foreign objects in the spatial structure of the display panel, which is used to support subsequent defect classification, spatial coordinate calculation and process traceability analysis.

[0264] The mapping module is used to identify the association between the structure-enhanced image recognition data output by the detection module and the corresponding processing technology, production equipment and historical parameters, and to construct the source tracing path of the foreign object recognition results;

[0265] The mapping module receives the identification results containing the area number information, and based on the mapping relationship between the area number and the production process code, queries and extracts the processing steps, processing equipment numbers, equipment operating time periods and their corresponding key process parameters corresponding to the monitored sub-area.

[0266] Furthermore, the mapping module establishes a multi-field mapping table containing "monitoring sub-area number – processing equipment number – process parameter set – timestamp" to accurately map each defect data point in the foreign object identification result to its potential processing source, thus realizing the input basis for equipment responsibility confirmation, process parameter traceability and subsequent analysis and processing;

[0267] The analysis and processing module is used to fuse the structure-enhanced image recognition data output by the identification and detection module with the processing traceability data generated by the mapping module, and to perform defect risk assessment, quality anomaly trend analysis and cross-regional defect correlation judgment.

[0268] The final output of the analysis and processing module includes: risk level label for each defect, judgment result of whether it belongs to a common cause defect cluster, identification signal of whether an early warning has been triggered, and numbered data of corrected defects, which can be used for subsequent production optimization, equipment inspection and quality accountability.

[0269] Compared to Example 1, Example 2 has different monitoring sub-regions. During the corresponding production stages (such as bonding, exposure, and encapsulation), the system collects key process parameters related to the corresponding stage in real time, including temperature (unit: degrees Celsius, sampling frequency 1 Hz); pressure (unit: Newtons); exposure power (unit: milliwatts per square centimeter); and adhesive layer thickness (unit: micrometers).

[0270] Action duration (unit: milliseconds); specifically, the system samples each monitoring sub-region at fixed time intervals. The key process parameters are determined using a sliding window mechanism, which takes the data from the T most recent time steps to construct a multidimensional process parameter sequence. ;

[0271] Furthermore, the system organizes key process parameters into a multidimensional process parameter sequence in the form of a T*n matrix according to time sequence. , means as follows:

[0272] ;

[0273] In the formula, This represents the value of the i-th critical process parameter at time j, where T is the length of the sliding time window; n is the number of critical process parameters; each column represents the time series of a critical process parameter; and each row represents the sampled values ​​of all parameters at a given time.

[0274] The system integrates a time series prediction model based on a long short-term memory neural network in the mapping module;

[0275] The time series prediction model uses historical production data. The input is a sequence of process parameters within a certain time window, and the output is a label indicating whether a defect exists, as identified by the detection module, after the window ends. During training, the input is a multi-dimensional sequence of historical process parameters. It outputs defect labels confirmed by the identification and detection module after the target time period; the defect labels are back-labeled by the foreign object identification results obtained from the structure-enhanced image recognition data.

[0276] Furthermore, the training set of the time series prediction model contains at least 10,000 samples, each corresponding to a specific monitoring sub-region. ;

[0277] When the system is working, it receives monitoring data from sub-regions. Latest multidimensional process parameter sequence The predicted probability of defect occurrence is obtained by using the time series prediction model output. ;

[0278] And the system will predict the probability. Compared with the preset risk threshold, if the predicted probability If the risk exceeds the risk threshold, the corresponding monitoring sub-area will be... The prediction results identified it as a potentially high-risk area;

[0279] When a certain monitoring sub-area Once identified as a potentially high-risk area, the system sends a scheduling instruction to the identification and detection module to improve the image acquisition quality of the area. This includes simultaneously activating the lateral view detection subunit, the adjustable optical axis deflector unit, and the parallax enhancement identification subunit to achieve collaborative imaging under multi-angle, multi-parallax, and multi-light conditions, thereby enhancing the visibility and structural resolution of foreign objects in the area.

[0280] Furthermore, each of the above sub-units is a physically independent module in the system structure. Parallel triggering and image data synchronization are achieved through a unified scheduling controller, and the output results are fused according to the timestamp alignment mechanism.

[0281] Example 3

[0282] Compared to Example 1, the mapping module in Example 3 monitors sub-regions. For nodes, construct a directed weighted graph structure G = (V, E), where:

[0283] , indicating all monitored sub-regions;

[0284] , representing the process dependency edges between regions, where, This indicates the monitoring sub-region. To the monitoring sub-area The strength of dependence;

[0285] In this embodiment, It is obtained through statistical analysis of the co-occurrence frequency of historical defects, configuration based on the experience of process engineering experts, or modeling of the concentration trend of foreign objects based on the output of the image recognition module;

[0286] For example, if the upstream monitoring sub-region Monitoring sub-regions to ensure proper bonding process For the packaging process, experience shows that defects in the former can easily affect those in the latter, therefore, [the following is set]: ;

[0287] When a certain monitoring sub-area If a mapped module determines a region to be potentially high-risk, the system traverses all process-dependent edges in the directed weighted graph, that is, it traverses the directed weighted graph structure G=(V,E) and identifies the current potentially high-risk monitoring sub-region. All outgoing edges from the starting point ( → ), used to identify potentially affected downstream areas. .

[0288] Example 4

[0289] Compared to Example 2, in Example 4 the mapping module monitors sub-regions. For nodes, construct a directed weighted graph structure G = (V, E), where:

[0290] , indicating all monitored sub-regions;

[0291] , representing the process dependency edges between regions, where, This indicates the monitoring sub-region. To the monitoring sub-area The strength of dependence;

[0292] In this embodiment, It is obtained through statistical analysis of the co-occurrence frequency of historical defects, configuration based on the experience of process engineering experts, or modeling of the concentration trend of foreign objects based on the output of the image recognition module;

[0293] When a certain monitoring sub-area If a mapped module determines a region to be potentially high-risk, the system traverses all process-dependent edges in the directed weighted graph, that is, it traverses the directed weighted graph structure G=(V,E) and identifies the current potentially high-risk monitoring sub-region. All outgoing edges from the starting point ( → ), used to identify potentially affected downstream areas. ;

[0294] For each downstream region Calculate its cumulative risk score The expression is as follows:

[0295] ;

[0296] In the formula, The predicted probability of the defect occurring; For region-dependent strength; To monitor sub-regions From the time it was identified as a high-risk area until the current assessment Time delay; This is the time decay coefficient, and its specific value can be adjusted according to the time sensitivity of the actual process and the characteristics of historical fault data.

[0297] If the cumulative risk score If the value exceeds the preset threshold, the downstream monitoring sub-region will be... It is marked as a secondary risk area and pushed to the identification and detection module, triggering the preset policy level upgrade.

[0298] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0299] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A foreign matter detection visual recognition system for display panel production, characterized by, The system comprises: a monitoring area division module, which divides a display panel into a plurality of monitoring sub-areas based on a display panel structure diagram and a process flow diagram, and assigns a unique area number to each monitoring sub-area; an identification detection module, which performs image detection on each monitoring sub-area, obtains a local image feature set and performs enhancement processing thereon, and outputs a set of standardized structure-enhanced image identification data; a mapping module, which associates the structure-enhanced image identification data with corresponding processing technology, production equipment and historical parameters, constructs a source tracing path of foreign matter identification results, and establishes a multi-field mapping table of "monitoring sub-area number- processing equipment number- process parameter set- timestamp"; an analysis processing module, which fuses the structure-enhanced image identification data and the source tracing path, and performs defect risk assessment, quality abnormality trend analysis and cross-area defect correlation judgment; In each monitoring sub-region In the corresponding production stage, the system collects the key process parameters related to the corresponding stage in real time. And the system will key process parameters in chronological order into the form of a multi-dimensional process parameter sequence T*n matrix , as follows: ； In the formula, represents the value of the ith key process parameter at the jth time point, T is the length of the sliding time window; n is the number of key process parameters; the system integrates a time series prediction model based on a long short-term memory neural network in the mapping module; While the system is working, receiving monitoring sub-regions Multidimensional process parameter sequence of the latest time period And the prediction probability of defect occurrence is obtained through the time series prediction model output ; And the system will predict the probability Compare with the preset risk threshold, if the predicted probability is greater than the risk threshold, the predicted result of the corresponding monitoring sub-region is determined as a potential high-risk area; The mapping module monitors sub-regions For each node, a directed weighted graph structure G = (V, E) is constructed, wherein: , indicating all monitoring sub-areas; , indicates a process dependent edge between regions, where, , indicates a dependent strength from monitoring sub-region to monitoring sub-region . When a certain monitoring sub-region is determined as a potential high-risk region by the mapping module, the system traverses all process dependency edges in the directed weighted graph; For each downstream region , its cumulative risk score is calculated , expressed as follows: ； wherein is the predicted probability of defect occurrence; is the area-dependent intensity; is the monitoring sub-area identified as a high-risk area, to the time delay at the current evaluation ; and is the time-dependent decay coefficient, whose specific value can be adjusted according to the time sensitivity of the actual process and the characteristics of the historical failure data. If the cumulative risk score is greater than a preset judgment threshold, the downstream monitoring sub-region is marked as a secondary risk area. 2.The foreign matter detection and visual identification system for display panel production of claim 1, wherein, when performing image detection in the identification detection module, if the corresponding position of the monitoring sub-area in the image has shifted, the system performs dynamic correction on the area number, including the following steps: S1. The system calculates the time difference between the image frame acquisition time and the area number binding time; if the theoretical production line movement distance corresponding to the time difference exceeds the preset area length threshold, it is determined that the current image has number misplacement; S2. Edge detection and feature point extraction are performed on the current image frame to identify the key structure points in the image frame; S3. The standard points are extracted from the standard structure template corresponding to the corresponding area number, and compared with the key structure points extracted from the image frame to calculate the offset between the key structure points and the standard points, and obtain the image drift distance and direction; S4. It is judged whether the offset exceeds the preset offset tolerance threshold. If it exceeds, it is considered that the area number of the current monitoring sub-area has misplacement; S5. According to the offset judgment result, the original number of the current image frame is replaced by the corrected area number. 3.The foreign matter detection and visual identification system for display panel production of claim 1, wherein, The identification detection module comprises a lateral view angle detection subunit, an adjustable optical axis deflection subunit and a parallax enhancement identification subunit; The lateral view angle detection subunit, the adjustable optical axis deflection subunit and the parallax enhancement identification subunit are simultaneously deployed in the system structure, and the system selectively activates one or more subunits to perform image acquisition and enhancement processing.

4. The foreign matter detection and visual recognition system for display panel production according to claim 3, wherein The lateral view angle detection subunit synchronously acquires the top view image and the side view image of the display panel by arranging a plurality of non-vertical industrial cameras, and performs image registration and enhancement processing on the acquired top view image and side view image; Through the processed top view image and side view image, the lateral view angle detection subunit identifies and extracts defects in the edge area or structure part of the display panel.

5. The foreign matter detection and visual recognition system for display panel production according to claim 4, wherein The adjustable optical axis deflection subunit is used to adjust the optical axis direction of the image acquisition device during the display panel detection process, so that the light rays irradiate to the monitoring sub-area to be detected at different incident angles, enhancing the imaging contrast of foreign matter edges, transparent contours and multi-layer interfaces.

6. The foreign matter detection and visual recognition system for display panel production according to claim 5, wherein The parallax enhanced recognition subunit is configured to, in a monitoring sub-region of a display panel, construct a parallax signal atlas between different view angles by synchronously collecting image data of multiple fixed view angles, identify image features that change in shape due to changes in view angle, and extract displacement responses of foreign matter between different structural levels or transparent materials. 7.The foreign matter detection and visual recognition system for display panel production of claim 1, wherein, The method for constructing the source tracing path by the mapping module includes the following steps: Z1, regional process information analysis: The mapping module first extracts the "region number" information in the structural enhancement image recognition data, and queries the production process link information corresponding to the number in the multi-field mapping table, including: The type of the corresponding process; The corresponding process segment order number; The unique identifier of the key equipment used in the current process; The process parameter template ID corresponding to the process; Z2, production equipment parameter binding: The mapping module calls the log data interface in the field manufacturing execution system or equipment control platform based on the equipment ID and process segment time interval obtained in step Z1, and extracts the equipment operation parameter information; Z3, historical defect record retrieval: The mapping module further retrieves historical records related to the current detection region in the defect history database, including: Whether there are similar foreign matter type identification results in the past several batches of the same equipment, the same process segment, and the same region number; Whether there is a defect high incidence period corresponding to the process segment parameter fluctuation; Whether there is a local abnormal aggregation caused by process switching, equipment maintenance, or raw material batch change; Z4, source tracing path generation and output structure: The mapping module finally generates a source tracing path record according to the results of the above three steps Z1, Z2 and Z3. The record is output in a structured data format and at least includes the following fields: Detection region number; Foreign matter type, location, and recognition confidence; Corresponding process name, process number, and processing time window; Equipment ID, equipment state, and parameter snapshot; Historical anomaly correlation score and source tracing label.

Images