Film layer overflow detection method and device based on thin film interference imaging

Abstract

The embodiment of the application provides a film layer overflow detection method and device based on thin film interference imaging, information is effectively extracted through innovatively designing a feature extraction system, through geometric analysis and texture analysis, a similarity evaluation mechanism is constructed, a reliable detection strategy is established in combination with feature fusion and region positioning, and abnormal analysis is introduced, parameter quantization and level evaluation are used to ensure the accuracy of detection. The method effectively solves the shortcomings of traditional technologies in feature extraction, region positioning and abnormal analysis, and provides technical support for film layer detection.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for detecting membrane overflow based on thin-film interferometry imaging. Background Technology

[0002] Existing methods for detecting membrane overflow have significant shortcomings. Traditional systems perform poorly in image processing and feature extraction, failing to effectively utilize interference fringe information and thus affecting detection accuracy.

[0003] Furthermore, existing technologies face bottlenecks in regional localization and similarity analysis. Most systems lack robust feature fusion mechanisms and matching strategies, resulting in inaccurate localization.

[0004] Existing systems have technical shortcomings in anomaly analysis. The lack of in-depth analysis of interference fringes makes efficient anomaly identification through parameter evaluation difficult, thus affecting detection results. Solving these problems is crucial for improving film detection capabilities. Summary of the Invention

[0005] To address the problems in the existing technology, this application provides a method and apparatus for detecting membrane overflow based on thin-film interferometry imaging, which can effectively solve the shortcomings of traditional technologies in feature extraction, region localization and anomaly analysis, and provide technical support for membrane detection.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] In a first aspect, this application provides a method for detecting film overflow based on thin-film interferometry imaging, comprising:

[0008] The system receives an image of the panel to be inspected, determines the outer boundary of the display area based on edge detection, analyzes process structure parameters to generate a film overflow inspection area and an uncoated inspection area, acquires a bright and dark texture image of the interference fringes in the reference area, extracts the geometric contour features and texture distribution features of the inspection area, constructs a detection formula data structure containing the coordinate information of the inspection area, interference fringe feature parameters, and process limit parameters, and writes the detection formula data structure into the inspection system.

[0009] The proportional relationship between the current panel and the reference panel is calculated based on the detection formula data structure to generate the initial search range of the test area. The straight and circular geometric features and gray-level co-occurrence matrix texture features of the interference fringes are extracted. The geometric similarity of Hu's moments and the similarity of texture distribution are calculated. The geometric feature similarity and texture feature similarity are fused according to the preset weight. The region with the highest comprehensive matching score is located as the precise test area. Gaussian filtering and histogram equalization enhancement are performed on the precise test area image. A dual threshold segmentation rule is constructed based on the mean and standard deviation of the region gray level. The bright stripe contours are extracted and projected onto the detection baseline for stripe clustering. The stripe direction angle and the proportion of broken pixels are calculated.

[0010] The stripe direction angle and the proportion of broken pixels are quantitatively analyzed, and the ratio of effective interference stripe pixels to the total number of pixels in the precise inspection area is statistically analyzed. The ratio is compared with the process limit parameters in the detection formula to determine the threshold. Based on the determination result, the type and severity level of the film coating abnormality are determined, and a detection report containing the abnormal location coordinates, abnormal area area, and abnormality level is generated. The detection report is then output to the production control system.

[0011] Furthermore, it also includes: loading the image data of the panel to be inspected, performing edge enhancement according to a preset brightness threshold, constructing an image gradient operator to extract the outline of the display area, generating a sequence of boundary points of the display area based on the outline coordinates, and fitting the boundary point sequence to the outer boundary curve of the display area;

[0012] Based on the outer boundary curve of the display area, a process structure parameter mapping function is constructed. The membrane overflow detection reference point and the uncoated detection reference point are calculated according to the mapping function. The outline of the area to be inspected is generated with the detection reference point as the center and the expansion radius is set as the membrane overflow inspection area and the uncoated inspection area.

[0013] Furthermore, it also includes: acquiring interference fringe imaging data of the reference area, performing gray-level normalization processing on the imaging data, extracting the intensity distribution map of bright and dark fringes, calculating the edge contour vector and gray-level co-occurrence matrix features of the area to be inspected, and generating a feature description set containing geometric feature parameters and texture feature parameters.

[0014] A detection area positioning template is established based on the feature description set. The detection area coordinate information, interference fringe feature parameters, and process limit parameters are written into the detection formula data structure. A formula parameter mapping table is generated based on the detection formula data structure and loaded into the detection system database.

[0015] Furthermore, it also includes: obtaining the current panel size parameters and the reference panel parameters, establishing a size mapping function, transforming the coordinates of the area to be detected in the detection formula to the current panel coordinate system according to the mapping function, generating the initial search range boundary, extracting the features of straight line segments and arc segments within the search range, and constructing a gray-level co-occurrence matrix to calculate texture statistics;

[0016] Based on the features of the straight line segment and the arc segment, the Hu's moment invariant is calculated, the texture statistics are normalized to generate a feature vector, and the geometric feature similarity and texture feature similarity are weighted and combined using a preset weight coefficient. The candidate regions are sorted according to the comprehensive similarity score, and the region with the highest score is selected as the precise location of the inspection area.

[0017] Furthermore, it also includes: applying a Gaussian kernel function to filter the precise image of the area to be inspected, calculating the image gray-level histogram distribution, determining the gray-level mapping curve based on the distribution characteristics, performing histogram equalization operation, calculating double threshold segmentation parameters based on the mean and standard deviation of the region gray-level, and performing differential operation on the double threshold segmentation results to extract the bright stripe contours.

[0018] The bright stripe contour is projected onto the detection baseline, and stripe clustering rules based on the projection distance are constructed. Straight line fitting is performed on the clustered stripes to calculate the direction angle, the continuous distribution of stripe pixels is statistically analyzed, the proportion of broken pixels is calculated, and a feature parameter set containing stripe direction angle and broken pixel ratio is generated.

[0019] Furthermore, it also includes: constructing a fringe validity judgment criterion based on the fringe direction angle and the proportion of broken pixels, marking the fringe that meets the requirements of direction angle range and continuity as valid interference fringes, calculating the total number of pixels of the valid interference fringes, and generating a fringe coverage index for the area to be inspected;

[0020] The stripe coverage index is compared with the preset upper and lower process thresholds in the test formula. Based on the comparison results, the anomaly type of the film layer is determined. The stripe coverage value is used to construct an anomaly degree quantification index, and a test result containing anomaly type markers and anomaly degree parameters is generated.

[0021] Furthermore, it also includes: determining the specific type of film coating anomaly based on the threshold judgment result, calculating the centroid coordinates and boundary contour of the abnormal area, calculating the area of ​​the abnormal area based on the boundary contour, constructing an anomaly degree quantification function based on the stripe coverage rate, and generating an anomaly feature record containing anomaly type identifier, location coordinates, area value, and severity level.

[0022] The abnormal feature records are organized according to the preset report template format, and detection timestamps and equipment identification information are added to construct a standardized detection report document. The detection report document is then transmitted to the production control system database through a data interface to trigger process parameter optimization instructions.

[0023] Secondly, this application provides a membrane overflow detection device based on thin-film interferometry imaging, comprising:

[0024] The feature extraction module is used to receive the image of the panel to be inspected, determine the outer boundary of the display area based on edge detection, analyze the process structure parameters to generate the film overflow inspection area and the uncoated inspection area, acquire the light and dark texture image of the interference fringes in the reference area, extract the geometric contour features and texture distribution features of the inspection area, construct the detection formula data structure containing the coordinate information of the detection area, the interference fringe feature parameters, and the process limit parameters, and write the detection formula data structure into the detection system.

[0025] The overflow detection module is used to calculate the ratio between the current panel and the reference panel according to the detection formula data structure, generate the initial search range of the inspection area, extract the straight and arc geometric features and gray-level co-occurrence matrix texture features of the interference fringes, calculate the geometric similarity of Hu's moments and the similarity of texture distribution, fuse the geometric feature similarity and texture feature similarity according to preset weights, locate the area with the highest comprehensive matching score as the precise inspection area, perform Gaussian filtering denoising and histogram equalization enhancement on the image of the precise inspection area, construct a dual threshold segmentation rule based on the mean and standard deviation of the region gray level, extract the bright stripe contour, project it onto the detection baseline for stripe clustering, and calculate the stripe direction angle and the proportion of broken pixels.

[0026] The overflow handling module is used to quantitatively analyze the stripe direction angle and the proportion of broken pixels, statistically analyze the ratio of effective interference stripe pixels to the total number of pixels in the precise inspection area, compare the ratio with the process limit parameters in the detection formula to determine the threshold, determine the type and severity level of the film coating abnormality based on the determination result, generate a detection report including the abnormal location coordinates, abnormal area area, and abnormality level classification, and output the detection report to the production control system.

[0027] As described above, this application provides a method and apparatus for detecting membrane overflow based on thin-film interferometry imaging. Through an innovative feature extraction system, it achieves effective information extraction via geometric and texture analysis. A similarity evaluation mechanism is constructed, combining feature fusion and region localization to establish a reliable detection strategy. Anomaly analysis is introduced, and through parameter quantification and level evaluation, the accuracy of detection is ensured. This method effectively addresses the shortcomings of traditional techniques in feature extraction, region localization, and anomaly analysis, providing technical support for membrane detection. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic flowchart of the membrane overflow detection method based on thin-film interferometry imaging in the embodiments of this application;

[0030] Figure 2 This is a structural diagram of the membrane overflow detection device based on thin-film interferometry imaging in the embodiments of this application;

[0031] [Figure Labels]

[0032] 10: Feature Extraction Module

[0033] 20: Overflow detection module

[0034] 30: Overflow Handling Module Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0037] To address the problems existing in current technologies, this application provides a method and apparatus for detecting membrane overflow based on thin-film interferometry imaging. Through an innovative feature extraction system, effective information extraction is achieved via geometric and texture analysis. A similarity evaluation mechanism is constructed, combining feature fusion and region localization to establish a reliable detection strategy. Anomaly analysis is introduced, and the accuracy of detection is ensured through parameter quantification and level evaluation. This method effectively solves the shortcomings of traditional technologies in feature extraction, region localization, and anomaly analysis, providing technical support for membrane detection.

[0038] To effectively address the shortcomings of traditional techniques in feature extraction, region localization, and anomaly analysis, and to provide technical support for membrane layer detection, this application provides an embodiment of a membrane layer overflow detection method based on thin-film interferometry imaging. See [link to relevant documentation]. Figure 1 The membrane overflow detection method based on thin-film interferometry specifically includes the following:

[0039] Step S101: Receive the image of the panel to be inspected, determine the outer boundary of the display area based on edge detection, analyze the process structure parameters to generate the film overflow inspection area and the uncoated inspection area, collect the light and dark texture image of the interference fringes in the reference area, extract the geometric contour features and texture distribution features of the inspection area, construct the detection formula data structure containing the detection area coordinate information, interference fringe feature parameters, and process limit parameters, and write the detection formula data structure into the detection system;

[0040] This embodiment receives the image of the panel to be inspected as input. The image is submitted by the upstream acquisition unit in single-frame format, along with exposure parameters and panel size identification for subsequent scaling. The image enters the grayscale and edge enhancement process, where a gradient operator is used to construct an intensity change field. In the intensity change field, non-maximum values ​​are used to suppress the formation of thinner edges, and then a set of contour lines is generated based on connectivity.

[0041] For the set of contour lines, the centroid constraint and the principle of long-side consistency are used to eliminate inner holes and noise contours, retaining the contour line with the largest coverage area. This contour line is fitted as the outer boundary curve of the display area, and the output is a sequence of boundary points and its parameterized representation, which serves as the benchmark for subsequent process mapping.

[0042] The process structure parameters are provided on the equipment side, including the positional relationship of the AA area, the geometric reference points, local spacing, and expansion radius of the DAM and NOINK areas. A mapping function from boundary coordinates to process coordinates is constructed based on the outer boundary curve of the display area. This mapping function is used to calculate the reference points for film overflow detection and uncoated detection.

[0043] Centered on each inspection reference point, a contour of the area to be inspected is generated according to a preset expansion radius. The contour is represented by a polygon and marked as the film overflow inspection area and the uncoated inspection area. The spatial constraint here ensures that the inspection area falls within the manufacturing tolerance range, avoiding subsequent matching deviations to non-target areas.

[0044] The light and dark texture images of the interference fringes in the reference area are acquired during the calibration stage and include reference samples under the same batch of process conditions. In this embodiment, grayscale normalization is performed on the reference image to obtain the fringe intensity distribution, and then the geometric contour and texture distribution are extracted within the reference window of each area to be inspected.

[0045] The geometry side uses straight lines and arcs as basic elements, recording endpoints, radii, and normal directions; the texture side calculates the energy, contrast, and homogeneity of the gray-level co-occurrence matrix, forming a fixed-length feature vector. This feature vector, along with the geometric parameters, is stored in a feature description set for the current panel's location and subsequent quality measurements.

[0046] On the current panel image, a proportional relationship is established based on the panel size markings and the reference panel parameters. The reference coordinates of the area to be inspected are mapped to the current coordinate domain to obtain the initial search range.

[0047] To mitigate the impact of scene variations, illumination equalization and small-scale smoothing are first performed within the initial search area to suppress non-striate brightness gradients. Then, straight line segments and arc segments are extracted, and Hugo's moment invariant is calculated as a measure of contour similarity. Simultaneously, a gray-level co-occurrence matrix is ​​constructed to obtain texture statistics. The two types of similarity are weighted and combined, and candidate windows are ranked. The window with the highest comprehensive score is selected as the precise detection region. This localization result maintains a coordinate system consistent with the boundary curve, facilitating cross-region aggregation.

[0048] To preserve reusable information in the formulation, this embodiment captures a precise statistical snapshot of the stripes within the area to be inspected, including a stripe direction histogram, texture feature vector, and geometric contour parameters. Considering the differences in reflectivity among different materials, a dual-threshold segmentation rule is employed in the coarse stripe extraction stage, with the following formula:

[0049] T = μ + k·σ,

[0050] Where μ represents the mean gray level of the precise inspection area, σ represents the standard deviation of gray level, and k represents an empirical coefficient. Two segmentation images are generated using two different k values, and the difference between the two images yields bright stripe candidates, avoiding stripe breakage and propagation under a single threshold. This candidate result is used for subsequent directional statistics and is not used as the final basis for defect determination.

[0051] The candidate fringe set is projected onto the detection baseline and clustered according to projection distance to avoid interference between multiple adjacent fringes during orientation estimation. Line fitting is performed within each cluster to obtain the fringe orientation angle. Pixel continuity is then statistically analyzed, and the proportion of broken pixels is calculated. The fringe orientation angle and the proportion of broken pixels together constitute the interference fringe feature parameters within the region. These parameters, combined with the texture feature vector, form a structured feature entry. This entry, along with the detection area coordinate information, becomes a key component of the formulation data.

[0052] The process limit parameters are derived from the process section input, including the upper limit of stripe coverage for membrane overflow and the lower limit of coverage for uncoated areas, as well as the effective range of stripe direction and continuity. In this embodiment, the limits are recorded separately for each detection area and linked to specific coordinates and feature entries in the formulation data using a mapping table. To ensure consistency across batches, the limit records include version tags and source batch numbers to avoid subsequent misuse.

[0053] The completed detection formula data structure contains three categories: detection area coordinate information, interference fringe characteristic parameters, and process limit parameters. At the end of this step, this data structure is submitted to the parameter management unit of the detection system, and a verification summary is returned to confirm its completeness. In subsequent steps, this formula provides a template and similarity metric for localization, and boundary conditions for threshold determination. Fringe direction and breakage ratio serve as validity constraints, limiting coverage statistics to only within valid fringes, ensuring the stability and reliability of the basis for anomaly type determination.

[0054] Step S102: Calculate the ratio between the current panel and the reference panel according to the detection formula data structure, generate the initial search range of the test area, extract the straight and arc geometric features and gray-level co-occurrence matrix texture features of the interference fringes, calculate the geometric similarity of Hu's moments and the similarity of texture distribution, fuse the geometric feature similarity and texture feature similarity according to the preset weight, locate the area with the highest comprehensive matching score as the precise test area, perform Gaussian filtering denoising and histogram equalization enhancement on the precise test area image, construct a double threshold segmentation rule based on the mean and standard deviation of the region gray level, extract the bright stripe contour, project it onto the detection baseline for stripe clustering, and calculate the stripe direction angle and the proportion of broken pixels.

[0055] In this embodiment, after the test formula is prepared, the coordinate information of the test area and the feature description set in the formula are read. The size identifier and resolution parameters of the current panel are used as input to establish the proportional relationship between the panel and the reference panel.

[0056] The scaling relationship employs a linear scale mapping and records the horizontal and vertical scaling factors. Combined with the circumscribed rectangle constraint of the outer boundary curve of the display area, an initial search range is generated based on the reference coordinates of the detection area. The search range is represented by a rectangular window, with the boundary clipped at the intersection of the shielded area and the display area to ensure that subsequent matching does not exceed the boundary.

[0057] To reduce interference from lighting differences and random noise, the image within the initial search range is first subjected to Gaussian filtering for noise reduction, and then histogram equalization is performed by determining the mapping curve based on the local gray-level histogram. The processed window then undergoes both geometric and texture feature extraction.

[0058] On the geometric side, based on the connected contours obtained from edge detection, straight line segments and arc segments are identified. The endpoints, lengths, center positions, and radius estimates of each segment are recorded, and Hugh's moment invariant is calculated as a criterion for shape similarity. On the texture side, a gray-level co-occurrence matrix is ​​constructed. Four statistical measures—energy, contrast, homogeneity, and correlation—are taken and combined into a fixed-length vector, which is then normalized within a window to suppress the influence of overall brightness drift on the results.

[0059] The matching process calculates geometric similarity and texture similarity separately. Geometric similarity is converted into a similarity score using the distance metric between Huygen's moments features, while texture similarity is scored using the cosine similarity of the co-occurrence matrix statistics.

[0060] The two are fused according to preset weights, which are derived from the formulation and vary depending on the type of detection area. For areas with film overflow, the focus is on geometric contour consistency, while for uncoated areas, the focus is on texture distribution consistency. A sliding window evaluation is performed within the initial search range to obtain a comprehensive matching score map. The window with the highest score is selected as the precise detection area, and the second highest score is retained as a confidence boundary in case of abnormal scenarios.

[0061] After the precise area to be inspected is determined, steady-state input is prepared for stripe analysis. Sub-pixel interpolation matching the filter kernel is applied again within the window to smooth fine-grained noise, and then a dual-threshold segmentation rule is constructed based on local brightness statistics.

[0062] Two binary images are generated using two different sets of k, and the difference between the two is used to obtain a candidate set of bright stripes. This approach can reduce stripe adhesion or breakage when high-reflectivity materials coexist with low-contrast backgrounds, while preserving the continuous bright stripe ridges.

[0063] After obtaining bright stripe candidates, the candidate pixels are projected onto the detection baseline along the normal direction to form a one-dimensional projection distribution. Stripe clustering is then performed based on an interval threshold set according to the projection distance. Robust line fitting is performed within each cluster, outputting the stripe direction angle. Simultaneously, pixel connectivity along the fitting direction is statistically analyzed, and the proportion of broken pixels is calculated. To avoid local defects shifting the angle estimation, a penalty weight is applied to low-confidence clusters. If the proportion of broken pixels exceeds the allowable range of the formulation, the cluster is not included in the valid stripe set.

[0064] Intermediate quantities for geometry and texture matching are used in a closed loop within this step and are also saved as the basis for subsequent threshold determination. Specifically, these include: the window coordinates of the precise area to be inspected, the comprehensive matching score, the distribution of stripe direction angles, the proportion of broken pixels, and the bright stripe mask. These quantities are directly called in subsequent coverage statistics and anomaly type determination. Coverage is only calculated within the effective stripe mask, and the direction angle and the proportion of broken pixels serve as validity constraints to avoid misjudgments caused by background textures.

[0065] To improve temporal consistency, this embodiment maintains consistent versions of the scaling and matching weights for consecutive frames, with the scaling factor only updated when the outer boundary curve of the display area shifts. If the overall matching score falls below the historical quantile threshold, a fallback strategy is triggered: the boundary scaling of the initial search range is increased, while the geometric similarity weight is raised until the confidence level of the precise detection area recovers to an acceptable range. In this way, localization and stripe extraction are completed under the same coordinate system, providing stable input for subsequent coverage calculations and process limit determination.

[0066] Step S103: Quantitatively analyze the stripe direction angle and the ratio of broken pixels, statistically analyze the ratio of effective interference stripe pixels to the total number of pixels in the precise inspection area, compare the ratio with the process limit parameters in the detection formula to determine the threshold, determine the type and severity level of the film coating abnormality based on the determination result, generate a detection report including the abnormal location coordinates, abnormal area area, and abnormality level classification, and output the detection report to the production control system.

[0067] In this embodiment, after obtaining the precise area to be inspected and its bright stripe mask, the stripe direction angle distribution and the proportion of broken pixels generated in the previous step are read, and the process limit parameters of the corresponding detection area are obtained simultaneously.

[0068] To form a stable metric that can be used for judgment, the fringe direction angle is first screened for validity: an allowable deviation range from the detection baseline is set, and fringes that fall within the range and whose corresponding cluster's proportion of broken pixels does not exceed the limit are marked as valid interference fringes; other fringes are not included in the coverage statistics. This screening is carried out at the pixel level, and the connected patches within the fringe cluster are further subdivided to prevent abnormal breakpoints from being mixed into a single cluster and causing statistical bias.

[0069] The output of the validity screening is a binary valid fringe map and a corresponding fringe label table. The number of pixels for each valid fringe is counted according to the fringe label table, and the total number of valid interference fringe pixels is obtained by summing these numbers. The total number of pixels in the precise inspection area is directly given by the window boundary; the ratio of the two constitutes the fringe coverage index. To suppress counting errors caused by edge interpolation, the pixel-width sideband around the window is removed from the coverage statistics, and integration is performed only within the main area.

[0070] This coverage is independent of material reflectivity differences because it originates from structural stripe geometry and texture consistency constraints, and is applicable to both overflow and uncoated categories.

[0071] Regarding the relationship between coverage and anomaly type, this embodiment uses upper and lower thresholds specified in the formulation for differentiation. When the coverage exceeds the upper threshold, it is determined to be membrane overflow; when the coverage is below the lower threshold, it is determined to be uncoated; samples between the two thresholds are recorded as normal or pending verification.

[0072] To avoid jitter near the threshold, a hysteresis interval is added. The width of the hysteresis interval is derived from historical statistical quantiles and remains consistent across batches. The stripe direction angle and the proportion of broken pixels are used as validity constraints. If either of these two indicators fails to meet the standards, the corresponding area is directly marked as low confidence and is not included in this threshold judgment, but is transferred to the subsequent re-inspection queue.

[0073] The severity level of anomalies is constructed based on the degree of deviation from coverage, using a piecewise monotonic quantization function to map the distance between coverage and the threshold into discrete levels.

[0074] To ensure comparability of the same type of anomaly across different inspected areas, the quantization function is normalized along with the limits in the formula during loading, outputting a grade code and corresponding confidence score. The confidence score is calculated using a compensation factor that considers the concentration of stripe directions and the proportion of breaks; the more concentrated the direction and the fewer the breaks, the higher the confidence score. This grade and confidence score are written into the header field of the anomaly feature record.

[0075] The coordinates of the anomaly location and the area of ​​the region are calculated on the valid fringe map. Based on the intersection of the valid fringe map and the candidate bright fringe map, connected component analysis is performed to obtain the boundary contour of the suspected anomaly region.

[0076] For each connected component, the centroid coordinates are calculated as the region location output, and the region area is obtained using the polygon area formula of the boundary contour. To avoid region fragmentation caused by multiple neighboring stripes, a consistent merging rule for distance and angle is set to merge adjacent connected components with similar directions before calculating the area and location. The parameters of the merging rule are bound to the detection region type and are recorded in the recipe.

[0077] The intermediate results of the quantitative analysis form a complete chain of anomaly features, including precise window coordinates of the inspection area, coverage value, anomaly type, severity level, centroid coordinates, area, and confidence level. This chain, along with the timestamp and device identifier, is encapsulated into a standardized record for generating the inspection report.

[0078] The report generation process fills in the above fields according to a preset template and attaches formula version information and proportion mapping snapshots for traceability to ensure comparability within the same batch. For low-confidence samples, a retesting mark is set in the report, preventing them from entering the production and disposal process.

[0079] Before outputting the report, a consistency check is performed, including whether the coverage is calculated from the effective stripe pixels, whether the anomaly type and severity level match the threshold range, whether the location coordinates fall within the window of the area to be inspected, and whether the area meets the minimum reportable area constraint.

[0080] If any verification fails, the record is downgraded to a temporary result and pushed to the re-inspection queue, without triggering the process adjustment interface of the production control system. A report that passes verification is written to the production control system via the data bus. The system then adjusts or performs spot checks on the parameters of the corresponding process based on this report. Simultaneously, the core fields of the report are written back as historical samples for periodic updates of thresholds and hysteresis intervals.

[0081] The above processing is performed in a closed loop at the frame-level granularity. The precise inspection area, bright stripe candidates, and directional breakage parameters provided in the previous steps are fully utilized in this step, and the output anomaly feature records and detection reports are fed back into parameter maintenance and version management, ensuring that the coverage threshold and severity level classification remain stable under actual production line data.

[0082] As described above, the membrane overflow detection method based on thin-film interferometry provided in this application can effectively extract information through an innovatively designed feature extraction system and geometric and texture analysis. A similarity evaluation mechanism is constructed, combining feature fusion and region localization to establish a reliable detection strategy. Anomaly analysis is introduced, and the accuracy of detection is ensured through parameter quantification and level evaluation. This method effectively solves the shortcomings of traditional techniques in feature extraction, region localization, and anomaly analysis, providing technical support for membrane detection.

[0083] In one embodiment of the film overflow detection method based on thin-film interferometry imaging in this application, it may further include the following:

[0084] Step S201: Load the image data of the panel to be inspected, perform edge enhancement according to the preset brightness threshold, construct an image gradient operator to extract the outline of the display area, generate a sequence of boundary points of the display area based on the outline coordinates, and fit the boundary point sequence to the outer boundary curve of the display area.

[0085] Step S202: Construct a process structure parameter mapping function based on the outer boundary curve of the display area, calculate the membrane overflow detection reference point and the uncoated detection reference point according to the mapping function, generate the outline of the area to be inspected with the detection reference point as the center and according to the preset expansion radius, and mark the outline of the area to be inspected as the membrane overflow inspection area and the uncoated inspection area.

[0086] In this embodiment, after the image acquisition unit completes exposure and storage, it loads the image data of the panel to be inspected. To ensure that the boundary of the display area remains separable in low-contrast scenes, edge enhancement is first performed based on the brightness threshold given by the device, and the grayscale gradient is improved by combining local contrast stretching and high-pass filtering.

[0087] Subsequently, an image gradient operator is constructed to calculate the intensity change fields in the horizontal and vertical directions respectively, and the gradient ridges are thinned by non-maximum suppression to obtain candidate edges. Connected component tracking is performed on the candidate edges to remove fragments that are too short or excessively curved, retaining only the contour lines with continuous coverage. The contour line coordinates are resampled into equally spaced point sequences according to the scanning order to form a sequence of boundary points of the display area. The local normals and curvatures of the point sequences are recorded as fitting constraints.

[0088] The boundary point sequence is input into the curve fitting stage. Considering that the outer boundary of the display area is mostly a structure with long straight sides and short rounded corners in manufacturing, a piecewise model fitting is adopted: robust linear regression is performed on the long side segments, and arc regression is performed on the corner segments. The segments are constrained by tangential and normal continuity to synthesize the outer boundary curve of the display area. The fitting process outputs the segment type, parameter set, and residual statistics. Points with residuals exceeding the threshold are marked as suspicious and temporarily masked from subsequent mappings to avoid irregular gaps affecting the baseline geometry.

[0089] Based on this outer boundary curve, a process structure parameter mapping function is constructed. The mapping function uses the arc length parameter of the boundary curve as the independent variable to establish a two-way relationship from panel pixel coordinates to process coordinates; its internal components include fixed process quantities such as the AA area reference box, the DAM area sideband width, and the relative position of the NOINK area.

[0090] To ensure consistency, monotonicity and continuity constraints are applied to the mapping function, and independent scaling factors are set for line segments and circular arc segments to handle cases of non-uniform scaling within the surface. The mapping function is frozen as the version effective in this batch after being verified by several reference feature points.

[0091] The overflow detection reference points and uncoated detection reference points are calculated according to the mapping function. Specifically, an overflow detection reference point series is obtained by sampling at an offset distance within the sideband on the DAM zone side, and an uncoated detection reference point set is generated in the NOINK zone according to the functional area centerline and equidistant distribution rules given by the process.

[0092] Each reference point inherits the type label of its respective area and carries local normal information for orientation definition of the area to be inspected.

[0093] Centered on each inspection reference point, the outline of the area to be inspected is generated according to a preset expansion radius. The expansion radius is taken from the allowable deviation range set by the process, and the outline adopts an oriented rectangle consistent with the local boundary normal, with the long side placed tangentially to cover the potential overflow or missing extension direction.

[0094] During the generation process, intersection clipping with the outer boundary of the display area is performed to prevent the inspected area from crossing the boundary into the invalid imaging area. All inspected areas are bound with a unique number, area type, and arc length parameter to form a structured list of areas.

[0095] To avoid cumulative errors in subsequent positioning, the area list includes two types of verification quantities: the normal distance from the boundary to the center of the area, and the overlap ratio between the area and the reference boundary. If either verification quantity exceeds the allowable range of the process, the area to be inspected is temporarily suspended from distribution, and the scaling factor and boundary fit are re-evaluated in the next frame. Through this step, the area to be inspected is spatially aligned with the process coordinates and geometrically conforms to the true boundary.

[0096] The output of the above processing is a set of outer boundary curves of the display area and contours of the area to be inspected. These are directly used as two inputs in subsequent steps: first, to generate an initial search range for multimodal feature matching; and second, to limit the effective window for coverage statistics after stripe extraction. The mapping function is reused throughout the entire process as an implicit intermediate quantity. When the panel size or boundary shape changes, it serves as the only geometric link that needs to be updated, ensuring that the coordinates of the area to be inspected, stripe positioning, and threshold determination are all within a unified coordinate system.

[0097] In one embodiment of the film overflow detection method based on thin-film interferometry imaging in this application, it may further include the following:

[0098] Step S301: Collect interference fringe imaging data of the reference area, perform gray-level normalization processing on the imaging data, extract the intensity distribution map of bright and dark fringes, calculate the edge contour vector and gray-level co-occurrence matrix features of the area to be inspected, and generate a feature description set containing geometric feature parameters and texture feature parameters.

[0099] Step S302: Establish a detection area positioning template based on the feature description set, write the detection area coordinate information, interference fringe feature parameters, and process limit parameters into the detection formula data structure, generate a formula parameter mapping table based on the detection formula data structure, and load the mapping table into the detection system database.

[0100] In this embodiment, after boundary fitting and the generation of the region to be inspected are completed, the data acquisition and feature construction of the reference region begins. The reference region is selected from the reference panel provided by the process provider, covering typical locations of the DAM and NOINK regions to ensure the representativeness of subsequent matching. The acquired interferometric fringe imaging data is stored in the original grayscale format of linear array scanning, including timestamps and region identifiers.

[0101] To eliminate the impact of exposure fluctuations, grayscale normalization is first performed within the region, and extreme brightness and darkness values ​​are suppressed using piecewise linear mapping to obtain a stable intensity distribution map of bright and dark fringes. The intensity distribution map preserves the spatial frequency and phase information of the fringes and serves as a common base map for geometric and texture calculations.

[0102] Two types of features are extracted within the reference window corresponding to each inspection area. On the geometric side, connected contours are obtained using edge operators, decomposed into straight lines and circular arcs, and the endpoint vectors, tangential directions, and local curvatures are calculated to form edge contour vectors. Simultaneously, the normal distance distribution to the detection baseline is recorded to distinguish stripe direction from the background contour. On the texture side, a gray-level co-occurrence matrix is ​​constructed, and energy, contrast, homogeneity, and correlation are statistically analyzed. Averaging is performed across multiple directions to reduce disturbances caused by slight deviations in stripe tilt angles. Both feature paths are normalized to zero-mean, unit-variance units, merged into a fixed-length feature vector, and appended with region numbers and region types to form a feature description set.

[0103] The feature description set, after verification, is used to create a detection region localization template. The template contains three parts:

[0104] First, the geometric prototype is represented by a parametric combination of the long side straight line and the short side rounded corner, giving the target interval of the Hu's moment invariant;

[0105] Second, the texture prototype is the target vector and tolerance of the co-occurrence matrix statistics;

[0106] Thirdly, a weighting scheme is used, which stipulates the fusion ratio of geometric similarity and texture similarity and corresponds one-to-one with the region type.

[0107] The template also records the allowable deviation angle of the stripe direction and the upper limit of the breakage ratio, which will be used as a priori for subsequent validity screening.

[0108] Based on this, a data structure for the test formulation is constructed. This structure binds the coordinate information of the test area with the characteristic parameters of the interference fringes according to the area number, and adds process limit parameters, including the upper limit of fringe coverage, the lower limit of coverage, the allowable range of direction, and the continuity requirements. To ensure consistent mapping across panels of different sizes, the data structure stores the version identifier of the scaling conversion caliber and the outer boundary curve of the display area, avoiding parameter crosstalk between different batches. After the structure is completed, a formulation parameter mapping table is generated, clearly defining the index relationship from the area number to the template, limit, coordinate, and scaling caliber.

[0109] Before being loaded into the detection system database, the mapping table undergoes a consistency check. The check items include consistency of feature vector dimensions, matching of template weights and region types, legality of limit ranges, and consistency of coordinate and outer boundary curve versions. Once the check is passed, the data is written to the database, and a verification summary and effective time window are returned. This mapping table is directly called in subsequent localization and segmentation stages: in the localization stage, geometric and texture similarity is calculated and fused according to the template; in the segmentation stage, valid stripes are selected based on limit values ​​and directional constraints. Thus, the bright and dark stripes and parametric features collected from the baseline region are continuously reused in the detection chain, ensuring consistent and traceable matching and judgment criteria.

[0110] In one embodiment of the film overflow detection method based on thin-film interferometry imaging in this application, it may further include the following:

[0111] Step S401: Obtain the current panel size parameters and the reference panel parameters, establish a size mapping function, transform the coordinates of the area to be detected in the detection formula to the current panel coordinate system according to the mapping function, generate the initial search range boundary, extract the features of straight line segments and arc segments within the search range, construct the gray-level co-occurrence matrix, and calculate the texture statistics.

[0112] Step S402: Calculate the Hu's moment invariant based on the features of the straight line segment and the arc segment, normalize the texture statistics to generate a feature vector, use a preset weight coefficient to weight and combine the geometric feature similarity and texture feature similarity, sort the candidate regions according to the comprehensive similarity score, and select the region with the highest score as the precise location of the area to be inspected.

[0113] This embodiment reads the physical dimensions and imaging resolution parameters of the current panel, and simultaneously obtains the version identifier of the reference panel parameters and the outer boundary curve.

[0114] To achieve consistent positioning across dimensions, a size mapping function is established, employing a linear model with independent scaling in both the horizontal and vertical directions and an overlay of origin offset, constrained to be monotonic and continuous. The mapping function takes the coordinates of the area to be inspected in the recipe as input and outputs the coordinates of the center point and the four corners of the oriented rectangle in the current panel coordinate system, thus generating the initial search range boundary. If the mapped boundary overlaps with the outer boundary of the display area, the intersection is clipped to eliminate out-of-bounds areas.

[0115] Preprocessing is performed within the initial search range, mild Gaussian filtering is applied to stabilize noise, and local histogram equalization is used to suppress slowly varying illuminance.

[0116] Next, feature construction proceeds along two paths: geometry and texture. On the geometry side, edges are extracted based on connected component tracking using gradient extrema. Least squares line fitting is used to identify line segments, and circular arc fitting is performed in regions of continuous curvature to obtain estimates of endpoints, lengths, center positions, and radii. Simultaneously, the angle distribution between each segment and the detection baseline is recorded, and structural edges that significantly deviate from the stripe orientation are removed. On the texture side, a gray-level co-occurrence matrix is ​​established for several candidate windows, and four quantities—energy, contrast, homogeneity, and correlation—are statistically analyzed. The orientation is averaged from multiple angles to reduce the directional bias caused by slight stripe tilt.

[0117] Based on the contour set of straight line segments and circular arc segments, Hu's moment invariants are calculated as shape representations. To reduce the influence of scale and rotation on moments, moment calculations are performed on a normalized scale of the edge binary map, and in-threshold truncation is introduced to suppress the contribution of fragmented edges.

[0118] The distance between the obtained Hu's moment vector and the geometric prototype in the template is measured, and the geometric feature similarity is obtained by exponential mapping. At the same time, the aforementioned texture statistics are normalized to zero mean and unit variance to form a fixed-dimensional texture feature vector, and the texture feature similarity is obtained by comparing it with the texture prototype of the template using cosine similarity.

[0119] The geometric and texture similarity are combined using weighted coefficients given by the formula. The weights are constrained by the type of the area to be inspected. For areas with film overflow, the geometric weight is increased, and for areas without coating, the texture weight is increased.

[0120] To improve robustness, confidence adjustments are made for the two types of similarity before combination: if the number of line segments in the geometric candidate is less than the minimum number of segments, the geometric similarity is weighted less; if the pixel count in the texture co-occurrence matrix is ​​insufficient, the texture similarity is weighted less. The combined similarity score map is then obtained, and the sliding candidate window is evaluated and sorted one by one. The region with the highest score is retained as the precise detection area, while the second and third highest scores are recorded as backups to address misselection caused by local occlusion or reflection.

[0121] Once the precise location of the area to be inspected is determined, its window coordinates are read back and solidified as input for subsequent stripe extraction. To ensure consistency with subsequent steps, geometric consistency is checked again within this window: the residuals of the fitted straight line and arc are checked to see if they fall within the template tolerance; if they exceed the limit, the next highest-scoring window is used instead and the check is repeated until the constraints are met or a backtracking is triggered to expand the search range. The window that passes the check caches its geometric and texture features together as intermediate localization values ​​for this frame and writes them into frame-level metadata for tracking and re-inspection.

[0122] The inputs to the above process are the coordinates of the area to be inspected and the feature template in the detection formula, as well as the size parameters of the current panel and the reference panel; the core intermediate quantities are the size mapping function, the initial search range boundary, the Hughes moment invariant, and the co-occurrence matrix statistics; the output is the precise location of the area to be inspected and the corresponding comprehensive similarity score. This output is directly connected to stripe segmentation and validity screening, ensuring that double-threshold segmentation and stripe clustering are performed within the correct spatial window, and using the geometric and texture features cached in this step as the prior basis for valid stripe judgment.

[0123] In one embodiment of the film overflow detection method based on thin-film interferometry imaging in this application, it may further include the following:

[0124] Step S501: Apply Gaussian kernel function to filter the image of the precise area to be inspected, calculate the gray-level histogram distribution of the image, determine the gray-level mapping curve according to the distribution characteristics, perform histogram equalization operation, calculate the double threshold segmentation parameters based on the mean and standard deviation of the gray level of the region, and perform difference operation on the double threshold segmentation results to extract the bright stripe contours.

[0125] Step S502: Project the bright stripe contour onto the detection baseline, construct stripe clustering rules based on the projection distance, perform straight line fitting on the clustered stripes to calculate the direction angle, statistically analyze the continuous distribution of stripe pixels, calculate the proportion of broken pixels, and generate a feature parameter set containing stripe direction angle and broken pixel ratio.

[0126] In this embodiment, the precise inspection area determined in the previous stage is used as input. First, a Gaussian kernel function is applied for filtering within this window. The size and standard deviation of the filter kernel are set according to the window scale to preserve the dominant frequency of the stripes while suppressing high-frequency noise. After filtering, the gray-level histogram distribution of the window is statistically analyzed to identify the main peak and shoulder structure, thereby determining the gray-level mapping curve. The mapping of dark area enhancement and highlight suppression is combined for histogram equalization, resulting in a more balanced stripe image. To form a stable segmentation threshold, the mean and standard deviation of the region's gray levels are calculated based on the equalization result, constructing a dual-threshold segmentation parameter.

[0127] Two different sets of k are selected to generate two binary images. The difference results highlight the ridges of the bright stripes and reduce the drag of the segmentation boundary on the background's gradually changing brightness. The difference binary images are then subjected to morphological thinning to remove isolated pixels and connect short breakpoints, outputting candidate contours for the bright stripes.

[0128] To ensure geometric consistency in stripe clustering, this embodiment projects the candidate contours of bright stripes onto the detection baseline along the normal direction. The detection baseline is defined by the tangent normal of the boundary of the region, ensuring that the projection axis aligns with the process direction. After projection, a one-dimensional distance distribution is obtained. Stripe clustering is constructed based on a distance threshold and a minimum interval rule. The threshold references the stripe spacing range recorded in the formulation, and the minimum interval prevents wide stripes from being mistakenly split into multiple clusters. During clustering, a minimum pixel threshold is set for low-amplitude isolated ridges to eliminate false stripes caused by noise.

[0129] Within each cluster, a straight-line fitting is performed, fitting the bright stripe skeleton points within the cluster. To reduce the influence of reflective spots, a robust loss is used for fitting, and large residual points are iteratively removed until the residuals converge or the iteration limit is reached. The stripe direction angle is obtained through fitting, and the angle is expressed as the angle with the detection baseline, facilitating subsequent comparison with the directional effectiveness range. Simultaneously, connectivity statistics are performed on the pixels within the cluster along the fitting direction, and the proportion of broken pixels is calculated, defined as the ratio of the number of non-connected pixels to the total number of pixels in the cluster. For stripes spanning multiple rows, connectivity is first evaluated within sub-blocks and then merged along the direction to avoid overestimating the overall breakage ratio due to local occlusion.

[0130] To improve the stability of angle and breakage indicators, a cluster-level weighted strategy is adopted, with the weight determined by both cluster length and pixel density. Longer and denser stripe clusters have greater weights, and their angles contribute more to the regional angle distribution. Sparsely dense or insufficiently long clusters serve only as auxiliary indicators and are not included in subsequent threshold determination. The angle of each cluster is statistically analyzed for the region's angle under its weight, including the mean and concentration, to determine directional stability in critical situations. The proportion of broken pixels is weighted and averaged, and the maximum breakage proportion is recorded as a risk indicator for use in subsequent validity screening.

[0131] The stripe direction angle and the proportion of broken pixels are organized together with the bright stripe mask into a feature parameter set. The parameter set includes window coordinates, mean stripe angle, angle concentration, weighted breakage ratio, cluster number, and a summary of the length distribution of each cluster, and is bound to timestamps and region numbers. This parameter set serves two purposes: firstly, as a direct input for stripe validity determination, filtering out stripes with directional deviations or insufficient continuity; and secondly, as an effective mask for coverage statistics, calculating area ratios only within the set of valid stripe pixels, avoiding bias caused by background textures or light spots entering the denominator.

[0132] Considering brightness drift under production conditions, this embodiment performs a slight review of the dual-threshold segmentation parameters at the end of the current frame. If the skeleton width of the bright stripe candidates systematically thickens or thins, the value range of k is adjusted to make the threshold of the next frame closer to the current grayscale distribution, but without exceeding the safety range set by the formula. This fine-tuning is recorded in the frame-level metadata and stored in the database along with the stripe direction and breakage statistics of the current frame, so that it can be directly reused for coverage calculation and anomaly type determination in subsequent steps, realizing a closed loop from filtering, equalization, segmentation to clustering and fitting.

[0133] In one embodiment of the film overflow detection method based on thin-film interferometry imaging in this application, it may further include the following:

[0134] Step S601: Construct a fringe validity judgment criterion based on the fringe direction angle and the proportion of broken pixels, mark the fringe that meets the requirements of direction angle range and continuity as valid interference fringes, calculate the total number of pixels of the valid interference fringes, and generate the fringe coverage index of the area to be inspected.

[0135] Step S602: Compare the stripe coverage index with the preset upper and lower process thresholds in the test formula, determine the anomaly type of the film layer based on the comparison results, construct an anomaly degree quantification index using the stripe coverage value, and generate a test result containing anomaly type markers and anomaly degree parameters.

[0136] This embodiment establishes a fringe validity criterion after obtaining the fringe direction angle, angle concentration, and weighted breakage ratio. The criterion uses two types of threshold constraints: first, the direction angle falls within the allowable deviation range of the formula record, and the angle concentration limits the inter-cluster dispersion; second, the weighted breakage ratio does not exceed the continuity upper limit, and the maximum breakage ratio of a single cluster does not trigger a risk warning. Fringe clusters that meet both conditions are marked as valid interference fringes, and their pixel set is obtained by intersecting the bright fringe mask with the morphological dilation result of the cluster skeleton, avoiding pixel loss caused by skeleton refinement. To prevent noise from introducing artificial height areas, a minimum effective width and minimum length threshold are set for each valid cluster, and pixels that do not meet the threshold are removed from the set.

[0137] The total number of effective interference fringe pixels is accumulated in clusters and recorded in the window-level statistics. The total number of pixels in the precise inspection area is fixed during window initialization. To prevent errors caused by edge interpolation, this embodiment excludes the outermost one-pixel-width edge band during calculation and uses the kernel region as the denominator. Fringe coverage is defined as the ratio of the number of effective interference fringe pixels to the total number of pixels in the kernel region. Coverage is updated with each frame. When an abnormal jump in coverage occurs in the same area within a short window, the double threshold parameters and cluster-level weights saved in the previous frame are read for verification. If it is confirmed to be an illumination disturbance, the previous coverage is maintained and marked as pending re-inspection.

[0138] The coverage index is compared with the upper and lower thresholds of the process in the test formulation to determine the anomaly type. A coverage rate greater than the upper threshold is recorded as film overflow; a coverage rate less than the lower threshold is recorded as uncoated; and a rate in between is recorded as normal. To suppress reciprocating jitter near the threshold, a hysteresis band is introduced, with bandwidth determined based on historical coverage quantile statistics and remaining constant within the same batch. Samples within the hysteresis band retain their type based on the previous valid determination, while simultaneously triggering a low-cost resegmentation to confirm the stability of the current mask.

[0139] Anomaly quantification constructs a piecewise monotonic function with the deviation of coverage from a threshold as the independent variable. For overflow, the greater the deviation from the upper limit, the higher the level; for uncoated areas, the greater the deviation from the lower limit, the higher the level. The function outputs a discrete level code and an auxiliary score, where the auxiliary score integrates two correction factors: angle concentration and breakage ratio. The higher the angle concentration and the lower the breakage ratio, the higher the auxiliary score, reflecting a more reliable stripe signal. This quantification result, along with the anomaly type, is written into the core field of the detection result for easy report generation and process traceability.

[0140] Before generating test results, a consistency check is performed. The check items include: whether the valid stripe set originates from clusters that meet both direction and continuity standards; whether the denominator of the coverage rate uses the number of pixels in the kernel region; whether the determination within the hysteresis band follows the previous valid state; and whether the anomaly level matches the corresponding range of the type. If any item fails, the current window result is downgraded to a re-inspection marker and enters the reprocessing queue, instead of being processed in the production line.

[0141] The detection results output by the above processing include stripe coverage, anomaly type markers, and anomaly severity parameters, along with window coordinates, timestamps, recipe versions, and scale-mapped snapshots, used for subsequent report generation and system-side closed-loop control. Coverage continues to be used as an intermediate quantity in batch statistical updates, participating in the adjustment of quantile estimates for hysteresis band and minimum effective width threshold; anomaly type and severity drive sampling inspection or process parameter tuning in the production control system, forming a stable link from validity determination to threshold comparison and then to quantitative output.

[0142] In one embodiment of the film overflow detection method based on thin-film interferometry imaging in this application, it may further include the following:

[0143] Step S701: Determine the specific type of film coating anomaly based on the threshold judgment result, calculate the centroid coordinates and boundary contour of the abnormal area, calculate the area of ​​the abnormal area based on the boundary contour, construct the anomaly degree quantification function based on the stripe coverage, and generate an anomaly feature record containing anomaly type identifier, location coordinates, area value, and severity level.

[0144] Step S702: Organize the abnormal feature records according to the preset report template format, add detection timestamps and equipment identification information, construct a standardized detection report document, and transmit the detection report document to the production control system database through the data interface to trigger process parameter optimization instructions.

[0145] In this embodiment, after completing the coverage comparison, the threshold judgment result is used as input to determine the specific type of film coating anomaly. For the inspection area marked as abnormal, the intersection of the effective stripe mask and the bright stripe candidate image is taken as the abnormal pixel set to avoid area deviation caused by skeletonization. Connectivity analysis is performed on the abnormal pixel set, and the first and zero moments of space are calculated for each connected component to obtain the centroid coordinates; the boundary is generated into a closed polygon using contour tracking, the vertex sequence is recorded and local normal information is retained for subsequent position alignment. To mask occasional isolated points, objects with a connected component size lower than the number of pixels corresponding to the minimum reportable area are removed.

[0146] The area of ​​anomaly regions is calculated using a polygonal formula based on the boundary contour, ensuring internal consistency between concavity and convexity and independent of directional assumptions. If multiple connected components with similar directions and a distance less than the merging threshold exist within the same region, they are first merged based on directional consistency and minimum spacing, and then a unified boundary and centroid are calculated to reduce the impact of fragmentation on area statistics. The merging threshold and directional consistency range are bound to the detection area type, derived from the formula, and remain unchanged across batches. For anomalies crossing window boundaries, the area of ​​missing sidebands is estimated through window extrapolation. The compensation amount is given by the normal extension provided by the outer boundary curve and is separately marked in the record.

[0147] Severity grading is based on a quantization function constructed from stripe coverage. This quantization function is a piecewise monotonic mapping that converts the deviation of coverage from the upper or lower limit of the process specifications into a grade code and an auxiliary score. For overflow, the higher the coverage exceeds the upper limit, the higher the grade; for uncoated areas, the lower the coverage falls below the lower limit, the higher the grade. The auxiliary score corrects for angular concentration and breakage ratio; samples with high concentration and low breakage receive higher confidence, thus exhibiting higher severity grade confidence for the same coverage deviation. The quantization output, along with the anomaly type identifier, centroid coordinates, and area value, constitutes the core fields of the anomaly feature record.

[0148] The anomaly feature record also includes location caliber and version information, specifically: precise window coordinates of the inspection area, size mapping function version, outer boundary curve version, and the formula number used. To ensure traceability, the record includes the double-threshold segmentation coefficients used in this frame, stripe clustering parameters, and a similarity weight summary, facilitating subsequent verification of critical samples. If multiple anomalies exist within the same window, only the type with the largest coverage deviation is retained as the primary anomaly, while the others are marked as secondary anomalies and listed separately to avoid duplicate statistics.

[0149] In the report compilation stage, the aforementioned anomaly records serve as the data source, and the reports are structured according to a pre-defined report template. The template includes a basic information area, an anomaly summary area, and a details area. The basic information area contains the detection timestamp and equipment identifier; the summary area provides the anomaly type and severity level; and the details area lists the location coordinates, area value, coverage, auxiliary score, and key parameter version. To facilitate rapid location by the process side, the location coordinates simultaneously provide pixel coordinates and process coordinates derived through a mapping function, with each corresponding to the same inspected area number.

[0150] After the report is generated, a consistency check is performed. This check includes verifying whether the area is calculated from a closed contour, whether the centroid falls within the window boundary, whether the anomaly type and coverage deviation direction are consistent, and whether the template fields are complete. Reports that pass the check are packaged into standardized documents and written to the production control system database via a data interface. The data interface uses a synchronous confirmation mechanism, returning the write result and record number; if the write fails, the report enters a retry queue and is retained in the local cache until successful confirmation.

[0151] Once the report is successfully entered into the database, process parameter optimization instructions associated with the anomaly type are triggered. For overflow, the instructions suggest narrowing adjacent spray paths or reducing the spray volume per unit volume; for uncoated areas, the instructions suggest increasing spray coverage density or re-inspecting the nozzle status. The triggering content is based solely on the anomaly type and severity level, without directly rewriting equipment parameters. Actual adjustments are executed at the process end, and the execution status is reported back. The execution status is linked to the record number of this report, serving as a reference sample for subsequent statistics and threshold maintenance.

[0152] The output of the above process is a standardized test report containing anomaly type identifier, location coordinates, area value, and severity level. The anomaly feature records are stored as intermediate quantities for a long time and used in batch-level statistics to evaluate the stability of the threshold and the suitability of the quantization function, forming a closed loop from judgment, quantification to reporting and process linkage.

[0153] To effectively address the shortcomings of traditional technologies in feature extraction, region localization, and anomaly analysis, and to provide technical support for membrane layer detection, this application provides an embodiment of a membrane layer overflow detection device based on thin-film interferometry (TIA) for implementing all or part of the aforementioned membrane layer overflow detection method based on thin-film interferometry. See [link to embodiment]. Figure 2 The membrane overflow detection device based on thin-film interferometry imaging specifically includes the following components:

[0154] Feature extraction module 10 is used to receive the image of the panel to be inspected, determine the outer boundary of the display area based on edge detection, analyze the process structure parameters to generate the film overflow inspection area and the uncoated inspection area, collect the light and dark texture image of the interference fringes in the reference area, extract the geometric contour features and texture distribution features of the inspection area, construct the detection formula data structure containing the coordinate information of the detection area, the interference fringe feature parameters, and the process limit parameters, and write the detection formula data structure into the detection system.

[0155] Overflow detection module 20 is used to calculate the ratio between the current panel and the reference panel according to the detection formula data structure, generate the initial search range of the inspection area, extract the straight and arc geometric features and gray-level co-occurrence matrix texture features of the interference fringes, calculate the geometric similarity of Hu's moments and the similarity of texture distribution, fuse the geometric feature similarity and texture feature similarity according to preset weights, locate the area with the highest comprehensive matching score as the precise inspection area, perform Gaussian filtering denoising and histogram equalization enhancement on the precise inspection area image, construct a dual threshold segmentation rule based on the mean and standard deviation of the region gray level, extract the bright stripe contour, project it onto the detection baseline for stripe clustering, and calculate the stripe direction angle and the proportion of broken pixels.

[0156] The overflow handling module 30 is used to quantitatively analyze the stripe direction angle and the proportion of broken pixels, statistically analyze the ratio of effective interference stripe pixels to the total number of pixels in the precise inspection area, compare the ratio with the process limit parameters in the detection formula to determine the threshold, determine the type and severity level of the film coating abnormality based on the determination result, generate a detection report including the abnormal location coordinates, abnormal area area, and abnormality level classification, and output the detection report to the production control system.

[0157] As described above, the membrane overflow detection device based on thin-film interferometry imaging provided in this application can effectively extract information through an innovatively designed feature extraction system, using geometric and texture analysis. A similarity evaluation mechanism is constructed, combining feature fusion and region localization to establish a reliable detection strategy. Anomaly analysis is introduced, and the accuracy of detection is ensured through parameter quantification and level evaluation. This method effectively solves the shortcomings of traditional technologies in feature extraction, region localization, and anomaly analysis, providing technical support for membrane detection.

[0158] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting film overflow based on thin-film interferometry imaging, characterized in that, The method includes: The system receives an image of the panel to be inspected, determines the outer boundary of the display area based on edge detection, analyzes process structure parameters to generate a film overflow inspection area and an uncoated inspection area, acquires a bright and dark texture image of the interference fringes in the reference area, extracts the geometric contour features and texture distribution features of the inspection area, constructs a detection formula data structure containing the coordinate information of the inspection area, interference fringe feature parameters, and process limit parameters, and writes the detection formula data structure into the inspection system. The proportional relationship between the current panel and the reference panel is calculated based on the detection formula data structure to generate the initial search range of the test area. The straight and arc geometric features and gray-level co-occurrence matrix texture features of the interference fringes are extracted. The geometric similarity of Hu's moments and the similarity of texture distribution are calculated. The geometric feature similarity and texture feature similarity are fused according to preset weights. The area with the highest comprehensive matching score is located as the precise test area. Gaussian filtering and histogram equalization enhancement are performed on the precise test area image. The double threshold segmentation parameters are calculated based on the mean and standard deviation of the region gray level. The double threshold segmentation results are subjected to difference operation to extract the bright stripe contours. The contours are projected onto the detection baseline for stripe clustering. The stripe direction angle and the proportion of broken pixels are calculated. Based on the fringe direction angle and the ratio of broken pixels, a fringe validity judgment criterion is constructed. Fringe that meets the requirements of direction angle range and continuity is marked as valid interference fringes. The ratio of valid interference fringe pixels to the total number of pixels in the precise inspection area is calculated. The ratio is compared with the process limit parameters in the detection formula to determine the threshold. Based on the judgment result, the type and severity level of film coating abnormality are determined. A detection report containing the abnormal location coordinates, abnormal area area, and abnormality level is generated. The detection report is output to the production control system.

2. The method for detecting film overflow based on thin-film interferometry imaging according to claim 1, characterized in that, The process of receiving the image of the panel to be inspected, determining the outer boundary of the display area based on edge detection, and analyzing process structural parameters to generate the film overflow inspection area and the uncoated inspection area includes: Load the image data of the panel to be inspected, perform edge enhancement according to the preset brightness threshold, construct an image gradient operator to extract the outline of the display area, generate a sequence of boundary points of the display area based on the outline coordinates, and fit the boundary point sequence to the outer boundary curve of the display area. Based on the outer boundary curve of the display area, a process structure parameter mapping function is constructed. The membrane overflow detection reference point and the uncoated detection reference point are calculated according to the mapping function. The outline of the area to be inspected is generated with the detection reference point as the center and the expansion radius is set as the membrane overflow inspection area and the uncoated inspection area.

3. The method for detecting film overflow based on thin-film interferometry imaging according to claim 1, characterized in that, The acquisition of the light and dark texture image of the interference fringes in the reference area is used to extract the geometric contour features and texture distribution features of the area to be inspected. A detection formula data structure is constructed, containing the coordinate information of the detection area, interference fringe feature parameters, and process limit parameters. This detection formula data structure is then written into the detection system, including: Interference fringe imaging data of the reference area is collected, gray-level normalization is performed on the imaging data, intensity distribution map of bright and dark fringes is extracted, edge contour vector and gray-level co-occurrence matrix features of the area to be inspected are calculated, and feature description set containing geometric feature parameters and texture feature parameters is generated. A detection area positioning template is established based on the feature description set. The detection area coordinate information, interference fringe feature parameters, and process limit parameters are written into the detection formula data structure. A formula parameter mapping table is generated based on the detection formula data structure and loaded into the detection system database.

4. The method for detecting film overflow based on thin-film interferometry imaging according to claim 1, characterized in that, The process involves calculating the ratio between the current panel and the reference panel based on the detection formula data structure, generating an initial search range for the area to be inspected, extracting the straight-line and circular-arc geometric features and gray-level co-occurrence matrix texture features of the interference fringes, calculating the geometric similarity of Hu's moments and the similarity of texture distribution, fusing the geometric feature similarity and texture feature similarity according to preset weights, and locating the region with the highest comprehensive matching score as the precise area to be inspected, including: Obtain the current panel size parameters and the reference panel parameters, establish a size mapping function, transform the coordinates of the area to be detected in the detection formula to the current panel coordinate system according to the mapping function, generate the initial search range boundary, extract the features of straight line segments and arc segments within the search range, construct a gray-level co-occurrence matrix to calculate texture statistics; Based on the features of the straight line segment and the arc segment, the Hu's moment invariant is calculated, the texture statistics are normalized to generate a feature vector, and the geometric feature similarity and texture feature similarity are weighted and combined using a preset weight coefficient. The candidate regions are sorted according to the comprehensive similarity score, and the region with the highest score is selected as the precise location of the inspection area.

5. The method for detecting film overflow based on thin-film interferometry imaging according to claim 1, characterized in that, The process involves performing Gaussian filtering for noise reduction and histogram equalization on the precise image of the target area, calculating dual-threshold segmentation parameters based on the mean and standard deviation of the region's grayscale, extracting bright stripe contours from the dual-threshold segmentation results through differential operations, projecting them onto the detection baseline for stripe clustering, and calculating parameters such as stripe direction angle and break pixel ratio. The Gaussian kernel function is applied to filter the image of the precise inspection area, the gray-level histogram distribution of the image is calculated, the gray-level mapping curve is determined according to the distribution characteristics, histogram equalization operation is performed, and the double threshold segmentation parameters are calculated based on the mean and standard deviation of the gray level of the region. The double threshold segmentation results are then subjected to difference operation to extract the bright stripe contour. The bright stripe contour is projected onto the detection baseline, and stripe clustering rules based on the projection distance are constructed. Straight line fitting is performed on the clustered stripes to calculate the direction angle, the continuous distribution of stripe pixels is statistically analyzed, the proportion of broken pixels is calculated, and a feature parameter set containing stripe direction angle and broken pixel ratio is generated.

6. The method for detecting film overflow based on thin-film interferometry imaging according to claim 1, characterized in that, The process involves constructing a fringe validity criterion based on the fringe direction angle and the proportion of broken pixels, marking fringes that meet the requirements of direction angle range and continuity as valid interference fringes, calculating the ratio of valid interference fringe pixels to the total pixels in the precise inspection area, and using this ratio to perform threshold determination against the process limit parameters in the detection formula. Based on the parameters of fringe direction angle and broken pixel ratio, a fringe validity judgment criterion is constructed. Fringe that meets the requirements of direction angle range and continuity is marked as valid interference fringes. The total number of pixels of the valid interference fringes is calculated to generate the fringe coverage index of the area to be inspected. The stripe coverage index is compared with the preset upper and lower process thresholds in the test formula. Based on the comparison results, the anomaly type of the film layer is determined. The stripe coverage value is used to construct an anomaly degree quantification index, and a test result containing anomaly type markers and anomaly degree parameters is generated.

7. The method for detecting film overflow based on thin-film interferometry imaging according to claim 1, characterized in that, The process of determining the type and severity level of film coating abnormalities based on the judgment results, generating a detection report including the coordinates of the abnormal location, the area of ​​the abnormal region, and the severity level of the abnormality, and outputting the detection report to the production control system includes: The specific type of film coating anomaly is determined based on the threshold judgment result. The centroid coordinates and boundary contour of the abnormal area are calculated. The area of ​​the abnormal area is calculated based on the boundary contour. An anomaly degree quantification function is constructed based on the stripe coverage rate to generate an anomaly feature record containing anomaly type identifier, location coordinates, area value, and severity level. The abnormal feature records are organized according to the preset report template format, and detection timestamps and equipment identification information are added to construct a standardized detection report document. The detection report document is then transmitted to the production control system database through a data interface to trigger process parameter optimization instructions.

8. A membrane overflow detection device based on thin-film interferometry imaging, characterized in that, The device includes: The feature extraction module is used to receive the image of the panel to be inspected, determine the outer boundary of the display area based on edge detection, analyze the process structure parameters to generate the film overflow inspection area and the uncoated inspection area, acquire the light and dark texture image of the interference fringes in the reference area, extract the geometric contour features and texture distribution features of the inspection area, construct the detection formula data structure containing the coordinate information of the detection area, the interference fringe feature parameters, and the process limit parameters, and write the detection formula data structure into the detection system. The overflow detection module is used to calculate the ratio between the current panel and the reference panel according to the detection formula data structure, generate the initial search range of the inspection area, extract the straight and arc geometric features and gray-level co-occurrence matrix texture features of the interference fringes, calculate the geometric similarity of Hu's moments and the similarity of texture distribution, fuse the geometric feature similarity and texture feature similarity according to preset weights, locate the area with the highest comprehensive matching score as the precise inspection area, perform Gaussian filtering denoising and histogram equalization enhancement on the precise inspection area image, calculate the double threshold segmentation parameters based on the mean and standard deviation of the region gray level, perform differential operation on the double threshold segmentation results to extract the bright stripe contours, project them onto the detection baseline for stripe clustering, and calculate the stripe direction angle and break pixel ratio parameters. The overflow handling module is used to construct a stripe validity judgment criterion based on the stripe direction angle and the proportion of broken pixels. Stripes that meet the requirements of direction angle range and continuity are marked as valid interference stripes. The ratio of valid interference stripe pixels to the total number of pixels in the precise inspection area is calculated. The ratio is compared with the process limit parameters in the detection formula to determine the threshold. Based on the judgment result, the type and severity level of the film coating abnormality are determined. A detection report containing the abnormal location coordinates, abnormal area area, and abnormality degree classification is generated. The detection report is output to the production control system.

Images