Multi-source stability collaborative compensation method and device for panel scanning system

Abstract

This application provides a multi-source stability collaborative compensation method and apparatus for a panel scanning system. Through innovative data processing system design, it achieves effective information integration via parameter analysis and feature extraction. A compensation model mechanism is constructed, combining mapping analysis and rule formulation to establish a reliable correction strategy. Dynamic optimization is introduced, ensuring imaging stability through multi-dimensional correction and collaborative compensation. This method effectively addresses the shortcomings of traditional technologies in data processing, model construction, and dynamic compensation, providing technical assurance for panel scanning.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a multi-source stability collaborative compensation method and apparatus for a panel scanning system. Background Technology

[0002] Existing stability compensation methods for panel scanning systems have significant shortcomings. Traditional systems perform poorly in data acquisition and parameter analysis, failing to effectively integrate multi-source information and thus affecting the compensation effect.

[0003] Furthermore, existing technologies suffer from bottlenecks in mapping relationships and model construction. Most systems lack robust parameter correlation mechanisms and feature evaluation strategies, resulting in inaccurate corrections.

[0004] Existing systems have technical shortcomings in dynamic compensation. A lack of in-depth analysis of imaging problems makes it difficult to achieve efficient stability improvements through collaborative optimization, thus affecting image quality. Solving these problems is crucial for improving panel scanning capabilities. Summary of the Invention

[0005] To address the problems in the prior art, this application provides a multi-source stability collaborative compensation method and apparatus for panel scanning systems, which can effectively solve the shortcomings of traditional technologies in data processing, model building and dynamic compensation, and provide technical support for panel scanning.

[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 multi-source stability collaborative compensation method for a panel scanning system, comprising:

[0008] Vibration acceleration data of the air-floating platform, flatness data of the laser displacement sensor array, and motion straightness data of the guide rail encoder are collected. The vibration acceleration data is analyzed according to three-dimensional components. Based on the flatness data, the local warping height and overall tilt angle parameters of the platform are extracted. The lateral offset between the actual trajectory and the theoretical trajectory of the platform is calculated according to the motion straightness data, and a multi-source stability parameter dataset is constructed.

[0009] A parameter correlation model is established based on the multi-source stability parameter dataset. The mapping relationship between vibration amplitude and image pixel offset, local warped area and out-of-focus area, overall tilt angle and perspective distortion, and lateral offset and periodic fluctuation of regular pattern is calculated. A collaborative compensation model including compensation threshold, feature evaluation function and correction rule is generated and written into the system controller.

[0010] Dynamic correction is performed based on the collaborative compensation model. Adjacent frame registration compensation is performed for image misalignment caused by vibration. Local sharpness optimization is performed by driving the focusing mechanism for out-of-focus areas. The perspective transformation matrix is ​​calculated and geometric correction is performed based on the overall tilt angle. The periodic fluctuation function is constructed using the lateral offset to reverse the coordinates of the regular pattern and output a stable imaging result.

[0011] Furthermore, it also includes: collecting vibration acceleration data streams of the air-floating platform, planar height data streams of the laser displacement sensor array, and position coordinate data streams of the guide rail encoder based on multiple sampling channels, performing time alignment and preprocessing on the three types of data streams, and generating a data sampling sequence under a unified time base;

[0012] Based on the data sampling sequence, the vibration acceleration data is decomposed into a three-dimensional spatial coordinate system, the relative change of plane height data is calculated, the lateral offset component of position coordinate data is extracted, and a multi-source stability dataset is constructed.

[0013] Furthermore, it also includes: receiving flatness data from a laser displacement sensor and motion trajectory data from a guide rail encoder, calculating the plane fitting residual based on the flatness data, extracting the relative height difference in local areas and the overall surface tilt angle, and generating a platform state feature vector;

[0014] Based on the comparison between the platform state feature vector and the theoretical uniform linear motion trajectory, the lateral deviation value of the actual motion trajectory is calculated, and a multi-source stability parameter dataset including local warping height, overall tilt angle, and lateral offset is established.

[0015] Furthermore, it also includes: classifying and organizing the multi-source stability parameter dataset according to vibration parameters, flatness parameters, and straightness parameters, constructing a three-dimensional feature space, performing statistical analysis on the distribution characteristics of various parameters, and generating a parameter feature mapping matrix;

[0016] Based on the parameter feature mapping matrix, the correlation coefficients between vibration amplitude and pixel offset, local warping and out-of-focus areas, overall tilt and perspective distortion, and lateral offset and periodic fluctuation are calculated, and a correlation model containing multi-source parameter mapping relationships is established.

[0017] Furthermore, it also includes: dynamically acquiring calibration plate images, extracting the position coordinate sequence of regular patterns, calculating the quantitative correspondence between vibration amplitude and pixel offset, warped area and out-of-focus distribution, tilt angle and perspective distortion, lateral offset and wave characteristics, and generating a set of feature mapping functions;

[0018] Based on the aforementioned feature mapping function set, a compensation threshold determination criterion is constructed, feature evaluation functions for various stability parameters are designed, corresponding correction rules are formulated, and the compensation threshold, evaluation functions, and correction rules are integrated into a collaborative compensation model, which is then written into the system controller.

[0019] Furthermore, it also includes: reading the correction rules of the collaborative compensation model, extracting feature points from continuously acquired image frames, calculating the displacement deviation vector between adjacent image frames, and performing spatial registration transformation on the image based on the displacement deviation vector;

[0020] The sharpness evaluation index of the local region is calculated based on the image after spatial registration transformation. The sharpness evaluation index is compared with a preset threshold, and the focusing mechanism is driven to adjust the focal plane of the region below the threshold.

[0021] Furthermore, it also includes: establishing a perspective transformation model based on the overall tilt angle, calculating the mapping matrix from image space to physical plane, using the mapping matrix to correct perspective distortion in the image, and generating corrected image data;

[0022] By fitting a periodic fluctuation function to the lateral offset sequence, the fluctuation amplitude and period parameters are extracted. The fluctuation function is then used to perform reverse compensation calculation on the coordinates of the regular pattern to generate a stable image coordinate sequence, and the corrected imaging result is output.

[0023] Secondly, this application provides a multi-source stability collaborative compensation device for a panel scanning system, comprising:

[0024] The dataset construction module is used to collect vibration acceleration data of the air-floating platform, flatness data of the laser displacement sensor array, and motion straightness data of the guide rail encoder. The vibration acceleration data is analyzed according to three-dimensional components. Based on the flatness data, the local warping height and overall tilt angle parameters of the platform are extracted. The lateral offset between the actual trajectory and the theoretical trajectory of the platform is calculated according to the motion straightness data to construct a multi-source stability parameter dataset.

[0025] The compensation model construction module is used to establish a parameter correlation model based on the multi-source stability parameter dataset, calculate the mapping relationship between vibration amplitude and image pixel offset, local warped area and out-of-focus area, overall tilt angle and perspective distortion, lateral offset and periodic fluctuation of regular pattern, generate a collaborative compensation model including compensation threshold, feature evaluation function and correction rule, and write the collaborative compensation model into the system controller.

[0026] The collaborative compensation module is used to perform dynamic correction based on the collaborative compensation model, perform adjacent frame registration compensation for image misalignment caused by vibration, drive the focusing mechanism to perform local sharpness optimization for out-of-focus areas, calculate the perspective transformation matrix based on the overall tilt angle and perform geometric correction, construct a periodic fluctuation function using the lateral offset to reverse correct the coordinates of the regular pattern, and output a stable imaging result.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the multi-source stability collaborative compensation method of the panel scanning system.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multi-source stability collaborative compensation method for the panel scanning system described above.

[0029] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the multi-source stability collaborative compensation method for the panel scanning system.

[0030] As can be seen from the above technical solution, this application provides a multi-source stability collaborative compensation method and apparatus for a panel scanning system. Through innovative design of the data processing system, it achieves effective information integration via parameter analysis and feature extraction. A compensation model mechanism is constructed, and a reliable correction strategy is established by combining mapping analysis and rule formulation. Dynamic optimization is introduced, and imaging stability is ensured through multi-dimensional correction and collaborative compensation. This method effectively solves the shortcomings of traditional technologies in data processing, model construction, and dynamic compensation, providing technical support for panel scanning. Attached Figure Description

[0031] 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.

[0032] Figure 1 This is a flowchart illustrating the multi-source stability collaborative compensation method for the panel scanning system in this application embodiment;

[0033] Figure 2 This is a structural diagram of the multi-source stability collaborative compensation device of the panel scanning system in the embodiments of this application. Detailed Implementation

[0034] 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.

[0035] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0036] To address the problems existing in current technologies, this application provides a multi-source stability collaborative compensation method and apparatus for panel scanning systems. Through innovative data processing system design, it achieves effective information integration via parameter analysis and feature extraction. A compensation model mechanism is constructed, combining mapping analysis and rule formulation to establish a reliable correction strategy. Dynamic optimization is introduced, ensuring imaging stability through multi-dimensional correction and collaborative compensation. This method effectively solves the shortcomings of traditional technologies in data processing, model construction, and dynamic compensation, providing technical assurance for panel scanning.

[0037] To effectively address the shortcomings of traditional technologies in data processing, model building, and dynamic compensation, and to provide technical support for panel scanning, this application provides an embodiment of a multi-source stability collaborative compensation method for a panel scanning system. See [link to relevant documentation]. Figure 1 The multi-source stability collaborative compensation method for the panel scanning system specifically includes the following:

[0038] Step S101: Collect vibration acceleration data of the air-floating platform, flatness data of the laser displacement sensor array, and motion straightness data of the guide rail encoder. Analyze the vibration acceleration data according to three-dimensional components. Extract the local warping height and overall tilt angle parameters of the platform based on the flatness data. Calculate the lateral offset between the actual trajectory and the theoretical trajectory of the platform based on the motion straightness data. Construct a multi-source stability parameter dataset.

[0039] In this embodiment, the time synchronization calibration of the multi-source sensing channels was completed before the system went online, providing a unified time base and a unified coordinate system as the input basis. This step receives continuous vibration acceleration data, flatness data, and motion straightness data. All three types of data are accompanied by timestamps and sampling numbers, and a fixed-length sliding window is maintained in the buffer to ensure sufficient time domain overlap.

[0040] To avoid phase deviations introduced by cross-channel jitter, a combination of interpolation alignment and integer sampling point backoff is used to generate aligned data sequences under a unified time base, providing stable input for subsequent component analysis and feature extraction.

[0041] Vibration acceleration data is mapped to the system's three-dimensional coordinate system based on the installation attitude, resulting in component sequences along the X, Y, and Z directions. The mapping matrix is ​​obtained from the attitude angle calibration of the equipment under installation conditions and remains unchanged during the maintenance period. To suppress high-frequency noise and low-frequency drift, a bandpass filter is used to process each component independently, followed by calculation of the short-window energy and peak value, outputting vibration amplitude and phase information. This information serves as a driving force in subsequent correlation modeling with image inter-line pixel offset and is also a necessary prerequisite for triggering registration compensation.

[0042] Flatness data is obtained from the grid scan of the plane under test by the laser displacement sensor array. To eliminate zero position difference between arrays, translation correction in the row and column directions is performed first, and then the reference plane is fitted by the least squares method to obtain the tilt angle parameters of the overall surface shape and the vertical residual distribution of the point cloud.

[0043] The local warpage height is calculated as the vertical distance of each measurement point relative to the fitted plane, and the output is a height raster and statistical summary at the same resolution. The overall tilt angle parameter is represented by the tilt angles in two orthogonal directions, along with the fitting residuals, which are used to measure the reliability of the geometric correction.

[0044] The motion straightness data consists of the actual motion trajectory recorded by the guide rail encoder and is compared with the theoretical uniform straight line at the same sampling step size.

[0045] To obtain the lateral offset, a one-dimensional reference axis is first established in the scanning direction. The projection distance of the actual trajectory onto the normal of this axis is calculated, resulting in a lateral offset sequence. For real-time considerations, sliding fit is used to remove low-order trends, while mid-to-high frequency components are retained for periodic feature recognition. The lateral offset is subsequently used to construct a periodic fluctuation function and to reverse-correct the coordinates of the regular pattern.

[0046] During the data aggregation phase, the three types of data are aligned using timestamps as keys and uniformly packaged into a multi-source stability parameter dataset. The dataset includes the amplitude and phase of the three vibration quantities, local warp height grids and overall tilt angle parameters, lateral offset sequences, and their respective trusted labels. To ensure that subsequent models are scale-insensitive, continuous quantities are subjected to interval scaling, and the scaling caliber is recorded. Original values ​​are retained by reference to avoid losing physical meaning due to scaling.

[0047] Considering the geometric constraints of this system, the calculation method for lateral offset is given as follows:

[0048] ΔL(n) = X_actual(n) X_theory(n),

[0049] Where X_actual is the projection of the actual position recorded by the encoder onto the reference axis normal, and X_theory is the theoretical position of uniform linear motion; n is the sampling number. The ΔL serves as the baseline expression for the lateral offset, used both for periodic fitting and for connecting the monitoring logic on the motion control side. This formula only defines the difference relationship and does not introduce a fixed threshold, facilitating consistent interpretation under different speeds and sampling frequencies.

[0050] At the end of this step, the aforementioned dataset is written to the buffer and a version fingerprint is added. Subsequent parameter association models read this dataset and establish mapping relationships between vibration amplitude and pixel offset, local warping and defocusing, overall tilt and perspective distortion, and lateral offset and periodic fluctuations. Dynamic collaborative compensation directly calls the lateral offset sequence and tilt angle parameters to drive the execution chain of image registration, focusing, and perspective correction. Through this organization, inputs, intermediate processing quantities, and outputs are integrated under the same time base, reducing cross-module uncertainties.

[0051] Step S102: Establish a parameter correlation model based on the multi-source stability parameter dataset, calculate the mapping relationship between vibration amplitude and image pixel offset, local warped area and out-of-focus area, overall tilt angle and perspective distortion, lateral offset and periodic fluctuation of regular pattern, generate a collaborative compensation model including compensation threshold, feature evaluation function and correction rule, and write the collaborative compensation model into the system controller.

[0052] This embodiment uses a multi-source stability parameter dataset as input to establish a parameter correlation model under a unified time base and coordinate system. The dataset includes the amplitude and phase of the three vibration components, local warp height grids and overall tilt angle parameters, lateral offset sequences, and trusted labels.

[0053] To ensure dimensional consistency, continuous quantities undergo interval scaling and inverse transformation caliber recording before being included in modeling, while discrete confidence labels are used for sample weighting. The modeling process uses a calibrated image with a regular pattern as a reference, and all image quantization metrics are derived from calibrated frames within the same sampling window to avoid cross-window drift.

[0054] The mapping between vibration and pixel offset adopts the "interline displacement regression" approach. First, feature points of regular patterns are extracted on adjacent frames and sub-pixel level displacements are calculated to obtain a pixel offset sequence.

[0055] A robust regression is performed on the sequence and vibration amplitude within the same time window, outputting the slope and residual distribution. This generates a linear mapping function from vibration amplitude to pixel offset, along with a confidence interval. This function is used to subsequently trigger registration and estimate the displacement to be compensated, avoiding accidental triggering of the compensation link under low amplitude conditions.

[0056] The correlation between local warping and defocus depends on the sharpness evaluation function. Using the warping height raster as an index, the sharpness index of the corresponding image block is extracted, and a "warping height-sharpness" lookup table is constructed. A piecewise monotonic model is used to fit the defocus threshold region to obtain the mapping between the local warping height threshold and the corresponding focusing step size.

[0057] The relationship between the overall tilt angle and perspective distortion is quantified by the curvature of the straight line features. A perspective transformation model is established based on the calibration geometry, and the mapping coefficient from the tilt angle to the perspective distortion amplitude is obtained by regression. This coefficient is used to generate prior constraints for solving the homography matrix, thereby improving the convergence and stability of the geometric correction.

[0058] The mapping between lateral offset and periodic fluctuations of a regular pattern is based on sine fitting. The lateral offset sequence and the pattern coordinate offset are time-series paired, and a periodic model is fitted using weighted least squares to extract the amplitude, period, and phase. The offset compensation function is defined as follows:

[0059] Δ (n) = A·sin(2πn / T+ ),

[0060] Where A is the amplitude and T is the period. Let Δ be the phase, and n be the sampling number. It serves as a generator for subsequent reverse correction and is bound to a trusted label of the encoder state. Its participation is reduced when the trust level decreases to prevent miscalibration.

[0061] After obtaining the above single correlation, they are integrated into a collaborative compensation model. The model contains three types of elements. The first is the compensation threshold, which provides the vibration trigger threshold, warping sharpness threshold, and tilt perspective threshold. It adopts a quantile strategy instead of a fixed constant and is versioned with each production batch. The second is the feature evaluation function, which defines the calculation caliber of inter-line displacement estimation, image block sharpness assessment, and straight line feature curvature measurement, and is used to determine the compensation intensity online. The third is the correction rule, which stipulates that no action is taken within the threshold, intervention is carried out with a gradually varying weight near the threshold, and full compensation is performed outside the threshold; priority is set when multiple channels are running concurrently to avoid mutual cancellation of compensation.

[0062] The output interfaces of the collaborative compensation model include the displacement registration target, focus step size estimation, initial value of the perspective homography matrix, and periodic reverse correction. The model is loaded into the system controller as a read-only parameter set and a subset of operations, with version and effective time window tagged. Subsequent dynamic correction directly calls the aforementioned interface quantities to perform image registration, local focusing, perspective correction, and periodic reverse correction, and writes the execution results back as samples for the next round of modeling, forming a traceable closed loop.

[0063] Step S103: Perform dynamic correction based on the collaborative compensation model, perform adjacent frame registration compensation for image misalignment caused by vibration, drive the focusing mechanism to perform local sharpness optimization for the out-of-focus area, calculate the perspective transformation matrix according to the overall tilt angle and perform geometric correction, use the lateral offset to construct a periodic fluctuation function to reverse correct the coordinates of the regular pattern, and output a stable imaging result.

[0064] In this embodiment, dynamic correction is performed based on the multi-source stability parameter dataset generated in step S101 and the loaded collaborative compensation model.

[0065] Image acquisition is triggered by line scans. Upon arrival of each frame, the controller synchronously reads the amplitude and phase of the three vibration parameters, the local warp height grid, the overall tilt angle parameters, and the lateral offset sequence. To reduce processing latency, a two-level buffer is established: a fast loop for registration and deblurring between the current and previous frames, and a slow loop for updating parameters related to profile tilt and periodic fluctuations. Both loops share the same time base, avoiding cross-loop drift.

[0066] To address image misalignment caused by vibration, adjacent frame registration compensation is performed. Based on the feature evaluation function given by the collaborative compensation model, the controller extracts high-confidence feature points on the current frame and calculates the displacement deviation vector from the previous frame.

[0067] Displacement estimation employs a pyramid-level coarse-to-fine matching approach, with matching cost jointly measured by gradient consistency and texture similarity. The resulting displacement deviation vector, after robust estimation to remove outliers, generates translation transformation parameters, which are applied to the current frame to achieve spatial alignment. For cases where the amplitude exceeds the model compensation threshold, a first-order affine degree of freedom is added to absorb slight shearing and prevent the accumulation of interline misalignment during stitching.

[0068] For out-of-focus areas, a sharpness evaluation index is calculated based on the local warp height grid of the flatness data. The index is then compared with the threshold in the model to determine the areas that need adjustment and the predicted focal length direction.

[0069] The focusing mechanism searches in small steps, employing a monotonic, hill-climbing strategy that only accepts steps when sharpness improvement is achieved, preventing back-and-forth oscillations. Sharpness evaluation is represented as a weighted sum of high-frequency energy and edge sharpness, and is limited to areas with significant warping using a region mask to avoid applying invalid actions to the entire image. After local sharpness optimization, the sharpness marker for that area is updated for the defect detection stage to filter out low-confidence segments.

[0070] To address the perspective distortion introduced by the overall tilt, the overall tilt angle parameters are read, and a perspective transformation matrix is ​​constructed. To ensure geometric consistency, the homography matrix is ​​derived from the normal vector obtained by plane fitting and the imaging geometry. The transformation center is aligned with the principal point of the image, and the boundary uses inverse mapping and sub-pixel interpolation to reduce resampling jagged edges.

[0071] To avoid mutual interference with registration of adjacent frames, geometric correction is performed after registration, and the updated extrinsic parameters are cached in the slow loop for initial values ​​in the next frame.

[0072] To address the periodic fluctuations caused by the straightness of motion, a periodic fluctuation function is constructed using lateral offset to perform a reverse correction on the coordinates of the regular pattern. The difference between the captured regular pattern coordinate sequence and the ideal coordinates is calculated, and a sinusoidal basis function is used to fit and obtain the combined parameters of amplitude, period, and phase. These parameters are then smoothly updated within the model's bandwidth constraints to avoid parameter jumps. The reverse correction is applied point-by-point to the current frame coordinates, and the corrected coordinates are written back to the stitching engine to ensure geometric consistency during large-scale stitching.

[0073] To standardize terminology, this embodiment expresses the coordinate correction of the lateral offset as a simple relationship:

[0074] Xcorrection(n) = Xobservation(n) ΔX(n),

[0075] Where Xobservation represents the observation coordinates of the regular pattern in the image, ΔX represents the periodic fluctuation function value obtained based on the lateral offset fitting, and n is the index. This expression only indicates the direction of the backcorrection; the specific values ​​of the parameters are determined jointly by the fitting result of the slow loop and the confidence label. If necessary, the weights of low-confidence samples are reduced to avoid error propagation.

[0076] The branches of dynamic correction are scheduled within the same frame period. The processing order is registration compensation, local sharpness optimization, perspective geometry correction, and periodic reverse correction. After processing, a stable imaging result is output, and the feature summary and parameter snapshot of the current frame are updated simultaneously. These intermediate quantities are directly used in the subsequent defect detection stage for threshold selection and coordinate mapping; they are also retained in log form on the control side for traceability and version consistency verification. The entire link maintains a loop-free dependency, and if any branch fails, it is rolled back using the most recent valid snapshot to ensure continuous output.

[0077] As described above, the multi-source stability collaborative compensation method for panel scanning systems provided in this application can effectively integrate information through innovative data processing system design, parameter analysis, and feature extraction. It constructs a compensation model mechanism, combining mapping analysis and rule formulation to establish a reliable correction strategy. Dynamic optimization is introduced, and through multi-dimensional correction and collaborative compensation, imaging stability is ensured. This method effectively addresses the shortcomings of traditional technologies in data processing, model construction, and dynamic compensation, providing technical support for panel scanning.

[0078] In one embodiment of the multi-source stability collaborative compensation method for the panel scanning system of this application, it may further include the following:

[0079] Step S201: Collect vibration acceleration data stream of air-floating platform, plane height data stream of laser displacement sensor array, and position coordinate data stream of guide rail encoder based on multiple sampling channels, perform time alignment and preprocessing on the three types of data streams, and generate a data sampling sequence under a unified time base;

[0080] Step S202: Decompose the vibration acceleration data according to the three-dimensional spatial coordinate system based on the data sampling sequence, calculate the relative change of the plane height data, extract the lateral offset component of the position coordinate data, and construct a multi-source stability dataset.

[0081] This embodiment is designed for online production line environments, setting up three sampling channels: vibration acceleration, planar height, and position coordinates. The sampling cycle is constrained by both scan triggering and encoder mileage interruption. Data from each channel is queued with both timestamp and sequence number, and the buffer maintains a window covering multiple frame cycles. To eliminate trigger jitter, a phase deviation estimate between the sampling clock and the main control clock is established. Spline interpolation is used to resample on the main time base to obtain the original data sequence under a unified time base. The interpolation order and confidence label of each sample point are recorded for subsequent weight allocation.

[0082] Vibration acceleration data preprocessing considers sensor bias and installation attitude. First, the bias is estimated using the mean of the static segment and then subtracted. Next, bandpass filtering is used to suppress low-frequency drift and high-frequency noise. Installation attitude is derived from the Euler angles calibrated online. A fixed-direction cosine matrix is ​​constructed to map the sensor coordinate samples to the system's three-dimensional coordinate system. After mapping, the acceleration sequences of the X, Y, and Z components, along with their corresponding energy indices and peak times, are output. This decomposition result is used for subsequent calculations corresponding to inter-line displacement in the image and also serves as input to the vibration-triggered compensation link.

[0083] Planar height data is provided by a laser displacement sensor array on an in-plane grid. There is a fixed zero-position difference and a small-scale drift between the arrays. In the preprocessing stage, the offset correction in the row and column directions is performed first, and then the consistency is checked within the same scan line. Abnormal channels are replaced with nearest neighbor interpolation.

[0084] For a height grid on a unified time base, a plane fitting method is used to obtain a reference plane and a tilt normal. The relative change is defined by the vertical distance of each measurement point relative to this plane. The output is a relative height map and an overall tilt angle parameter at the same resolution. The relative height map is directly indexed when locating out-of-focus areas, and the overall tilt angle parameter provides prior information for perspective geometry correction.

[0085] The position coordinate data comes from the displacement readings of the guide rail encoder. To improve the stability of lateral offset extraction, a one-dimensional reference axis is established in the same direction as the scanning direction, and the projection distance of the actual trajectory in the normal direction of this axis is calculated.

[0086] Considering that encoder readings include quantization errors and jitter, a sliding polynomial fitting is first used to extract low-order trends, and the residual portion is retained as a candidate for lateral offset. This candidate is then aligned with the platform velocity curve to eliminate instantaneous spurious offsets caused by velocity transitions, resulting in a smooth sequence of lateral offset components. This sequence is used for fitting periodic fluctuations and generates a reverse correction during the dynamic correction phase.

[0087] The time alignment strategy adopts the rule of "alignment priority, content fidelity preservation". The three types of data sequences are aligned with the timestamp as the primary key. If a sample point is missing in a certain channel, it is filled in according to the interpolation order of the samples on both sides not exceeding a predetermined upper limit, and the confidence label of the sample point is set to low.

[0088] After alignment, a consistency check is run to compare the phase relationship between the peak vibration time and the lateral offset. If an out-of-bounds difference is found, the alignment parameters are rolled back and the interpolation span is shortened until the check passes. Through this process, the data sampling sequence under a unified time base remains physically interpretable in the time domain.

[0089] Based on the alignment and preprocessing results described above, a multi-source stability dataset is constructed. This dataset is indexed by frame period and contains three structured parts:

[0090] One is the three-dimensional component sequence of vibration acceleration, energy index, and peak position;

[0091] Secondly, there is the relative height map and the overall tilt angle parameters, along with the plane fitting residual as a reliability measure;

[0092] Thirdly, the time series of the lateral offset component and its correspondence with the velocity curve.

[0093] To ensure scale consistency in subsequent models, the scaling of continuous quantities is recorded, while the original quantities are retained by reference to avoid loss of physical meaning.

[0094] After the dataset is generated, it is lightly bound to the metadata of the image frames to establish a mapping relationship between the frame index and the three types of parameters.

[0095] In subsequent parameter correlation modeling, vibration components and image pixel offsets are regressed within a window; local relative height and sharpness evaluation are paired for analysis; overall tilt angle and straight line feature curvature are used for perspective mapping; and lateral offset components and regular pattern coordinate offsets are fitted with a periodic model. The dynamic correction stage directly reads the corresponding entries from the dataset, reducing the risk of redundant calculations and cross-module data inconsistencies. The entire process operates under the same time base and coordinate system, ensuring consistent interpretation and traceability of parameters before and after compensation.

[0096] In one embodiment of the multi-source stability collaborative compensation method for the panel scanning system of this application, it may further include the following:

[0097] Step S301: Receive flatness data from the laser displacement sensor and motion trajectory data from the guide rail encoder, calculate the plane fitting residual based on the flatness data, extract the relative height difference of the local area and the overall surface tilt angle, and generate the platform state feature vector;

[0098] Step S302: Based on the comparison between the platform state feature vector and the theoretical uniform linear motion trajectory, calculate the lateral deviation value of the actual motion trajectory, and establish a multi-source stability parameter dataset including local warping height, overall tilt angle, and lateral offset.

[0099] In this embodiment, flatness data and motion trajectory data are read under a unified time base. Both types of data are bound to the same batch of image frames using frame indices and timestamps. Flatness data is provided by an in-plane grid from a laser displacement sensor array, and preprocessing has completed zero-difference correction and abnormal channel replacement.

[0100] This step first constructs a reference plane on the grid, then uses least-squares fitting to obtain the plane's normal and intercept, and calculates the perpendicular distance from each measurement point to the reference plane, which serves as the plane fitting residual field. The residual field retains the original resolution and is used for subsequent local statistics and region mask generation.

[0101] The relative height difference of local regions is extracted from the residual field. The grid is divided into sliding windows of fixed size, and the difference between the median and extreme heights in each window is calculated as a robust estimate of the local warping intensity. At the same time, the center position and area of ​​the window are recorded to form a region-level structured entry.

[0102] The overall surface tilt angle is converted into tilt angles in two orthogonal directions using the normal vector of the reference plane, with the overall root mean square of the fitted residuals used as a confidence metric. To prevent statistical distortion of the window at the boundaries, a symmetrical filling strategy is introduced when the window falls to the edge, avoiding suppressing the local warping estimate.

[0103] The above quantities are organized in chronological order into a platform state feature vector. The fixed fields of the vector include the frame index, the two components of the overall tilt angle, the root mean square of the overall residual, a spatial distribution summary of the local warp intensity, and a window index table for localization.

[0104] The spatial distribution summary consists of quantiles and maximum values, which facilitates threshold determination without losing information about extreme regions. This vector serves as the direct input for subsequent trajectory comparison and compensation determination, and the sensor array configuration and fitting order are recorded in the metadata to ensure consistent interpretation.

[0105] The motion trajectory data comes from the guide rail encoder and includes the principal displacement and lateral micro-displacement in the scanning direction. To calculate the lateral deviation of the actual motion trajectory, the scanning principal axis and its normal reference axis are first defined in the installation geometry. The encoder sampling trajectory is projected onto the normal reference axis to obtain the original lateral displacement sequence; then, a sliding polynomial fitting is used to remove extremely low-frequency drift and retain the mid-to-high frequency components that reflect straightness characteristics. Considering that platform speed changes may introduce phase folding, this step aligns the lateral displacement sequence with the speed curve and reduces the weight of short-window samples on both sides of the speed transition point to reduce the influence of spurious deviations.

[0106] The theoretical uniform linear motion trajectory is generated based on the scan segment length and the desired velocity, sampled at equal intervals on a unified time base, and only includes the ideal position along the principal axis. The projection of the actual trajectory onto the normal reference axis is compared with the theoretical trajectory, and the lateral deviation is calculated according to the sampling sequence number. To provide a consistent standard, the following formula is used:

[0107] ΔL(n) = X_actual(n) X_theory(n),

[0108] Where ΔL is the lateral deviation value, X_actual is the projection of the actual trajectory onto the normal direction, X_theory is the ideal projection of the theoretical trajectory onto the normal direction, and n is the sampling number. Confidence labels are introduced during the calculation process to mark incomplete samples caused by interpolation or anomaly removal, thus avoiding overfitting in subsequent period fitting.

[0109] When constructing the multi-source stability parameter dataset, the platform state feature vector and the lateral deviation value are associated one by one by frame index, and three types of core quantities are output.

[0110] One is the local warp height, defined by the relative height difference within the sliding window, and accompanied by a corresponding spatial index, for use in focusing link positioning.

[0111] Secondly, the overall tilt angle serves as a priori input for perspective correction, and carries the overall fitting residual as a confidence level.

[0112] Thirdly, there is the lateral offset, which is a time series of ΔL, used for modeling periodic fluctuations and reverse correction.

[0113] The dataset retains necessary original metric references and scaling calibers to ensure consistent interpretation across different batches and sampling rates.

[0114] To support the stable operation of subsequent collaborative compensation, this step performs a consistency check after the dataset is generated. The check examines whether the correlation between the overall tilt angle and the local warping intensity is abnormally high. If large-scale structural warping causes threshold false triggering, the intervention threshold of the focusing link is increased. Simultaneously, it checks whether the phase relationship between the lateral offset and the encoder speed is stable. If drift occurs in a specific frequency band, the weight of that segment is reduced. The check results are written into the dataset metadata in the form of tags for use in association modeling and dynamic correction strategy selection. Through the above processing, the dataset remains computable and traceable in both spatial and temporal dimensions, providing reliable input for the online mapping and compensation in steps S102 and S103.

[0115] In one embodiment of the multi-source stability collaborative compensation method for the panel scanning system of this application, it may further include the following:

[0116] Step S401: Classify and organize the multi-source stability parameter dataset according to vibration parameters, flatness parameters, and straightness parameters, construct a three-dimensional feature space, perform statistical analysis on the distribution characteristics of each type of parameter, and generate a parameter feature mapping matrix;

[0117] Step S402: Based on the parameter feature mapping matrix, calculate the correlation coefficients between vibration amplitude and pixel offset, local warping and out-of-focus area, overall tilt and perspective distortion, and lateral offset and periodic fluctuation, and establish a correlation model that includes multi-source parameter mapping relationships.

[0118] This embodiment reads a multi-source stability parameter dataset organized by frame, and classifies and organizes vibration parameters, flatness parameters, and straightness parameters to ensure that the timestamps and coordinate scales of the three types of quantities are consistent within the same frame. Vibration parameters include statistics of the amplitude and phase of the three components, flatness parameters include local warping height grids and overall tilt angle, and straightness parameters include lateral offset sequences and their trusted labels.

[0119] To avoid scale differences affecting subsequent estimations, continuous quantities are scaled in intervals without changing their physical meaning. The reversible transformation caliber is recorded, and missing segments are marked with a visibility mask instead of being filled in numerically.

[0120] Based on this, a three-dimensional feature space is constructed. Each frame is mapped to a joint feature vector, which is composed of a vibration sub-vector, a flatness sub-vector, and a straightness sub-vector in a fixed order. The vibration sub-vector takes the three components of energy and the dominant frequency position, the flatness sub-vector is represented by the overall tilt angle and the local warp quantile summary, and the straightness sub-vector is represented by the amplitude and phase of the lateral offset short window.

[0121] Distribution statistics were performed on the space, calculating the marginal distribution and pairwise joint distribution, identifying anomalous tails, and controlling their impact on the estimation using a weighted decay method. The statistical results are presented as a parametric feature mapping matrix, where the rows correspond to the source parameters, the columns correspond to the response indices of the image and motion sides, and the elements are the correlation and scaling factors under robust statistical standards, along with the sample size and confidence interval.

[0122] To establish a mapping relationship that can be invoked by the controller, the source parameters need to be paired one-to-one with the image response. Pixel offsets are obtained by registration of adjacent frames, out-of-focus areas are obtained by sharpness evaluation within a local mask, perspective distortion is calculated from the curvature of straight line features, and periodic fluctuations are given by the offset sequence of regular pattern coordinates.

[0123] On the parameter feature mapping matrix, normalized correlation coefficients are further calculated, and robust regression and piecewise monotonic constraints are used to avoid unreasonable backward gradients near the threshold. For each pair of source parameters and response quantities, the point estimate and variance of the correlation coefficient are output, and the constraint range of the visibility mask on the effective samples is recorded.

[0124] The correlation between vibration amplitude and pixel offset is established using inter-row displacement regression. The model is a linear kernel, and the slope reflects the average pixel offset caused by a unit amplitude.

[0125] The correlation between local warping and out-of-focus areas is established using a segmented threshold model. A monotonically decreasing sharpness curve is set in the middle and high segments of the warping height. The out-of-focus threshold is determined by the quantile points, and an empirical mapping of the focus step size is given.

[0126] The relationship between overall tilt and perspective distortion is constrained by geometric consistency. The regression variables are the two components of the tilt angle, and the response is the curvature measure. Joint regression yields the mapping coefficient matrix and residual shape, which facilitates the initial value setting for subsequent homography matrix solving. The relationship between lateral offset and periodic fluctuation is characterized by weighted sine fitting, with weights taken from the encoder's reliable labels. The amplitude, period, and phase of the fitted solution constitute the core parameters for period compensation.

[0127] To improve the usability of the model under different working conditions, an association model containing multi-source parameter mapping relationships is constructed, using an organization method of "threshold set plus feature function library".

[0128] The threshold set is derived from the quantile rules of the aforementioned distribution statistics and correlation coefficient estimation. The feature function library includes interline displacement estimation function, sharpness evaluation function and curvature evaluation function, and gives the caliber definitions of input domain and output domain.

[0129] The model also carries conflict resolution rules. When vibration and straightness act on the coordinates at the same time, the high-frequency offset is explained by the periodic fluctuation term first, and the remaining low-frequency part is absorbed by the registration to avoid repeated compensation.

[0130] After generation, the association model undergoes a consistency check. The check compares the stability of association coefficients across different batches of samples; if cross-batch drift occurs, it reverts to a more robust quantile threshold. Simultaneously, it checks whether the contribution signs of different source parameters to the same response are consistent within the same frame; if contradictions exist, the frame is marked as low confidence and removed from the training window. Models that pass the check are solidified in a versioned format. Their core output consists of four types of mapping relationships and corresponding threshold and feature evaluation function configurations, which are directly invoked by the controller during the dynamic correction phase.

[0131] In the control path, the correlation model is loaded using a combination of read-only parameters and updatable statistics. Dynamic correction calls correlation coefficients and thresholds to determine whether to trigger adjacent frame registration, local focus, perspective correction, and periodic reverse correction. The output of the feature evaluation function is used to quantify the compensation intensity and is written back as new statistical samples after execution, becoming the basis for the next round of matrix updates. Through this organization, the causal chain from source parameters to image response converges within the same time base, reducing cross-stage uncertainty and providing a stable model foundation for subsequent collaborative compensation.

[0132] In one embodiment of the multi-source stability collaborative compensation method for the panel scanning system of this application, it may further include the following:

[0133] Step S501: Dynamically acquire the calibration plate image, extract the position coordinate sequence of the regular pattern, calculate the quantitative correspondence between vibration amplitude and pixel offset, warped area and out-of-focus distribution, tilt angle and perspective distortion, lateral offset and wave characteristics, and generate a set of feature mapping functions.

[0134] Step S502: Based on the feature mapping function group, construct the compensation threshold judgment criterion, design feature evaluation functions for various stability parameters, formulate corresponding correction rules, integrate the compensation threshold, evaluation function, and correction rules into a collaborative compensation model, and write it into the system controller.

[0135] This embodiment uses a calibration board for dynamic data acquisition, with the acquisition process consistent with the mass production scanning cycle to ensure true coupling between the imaging and motion links. The calibration board contains a regular pattern with pre-calibrated ideal coordinates. After an image frame arrives, the corner points and edge intersections of the regular pattern are extracted under stable exposure and synchronous triggering conditions. Subpixel positioning is used to obtain a position coordinate sequence, which is then bound to the frame index and timestamp. To suppress occasional false detections, a geometric consistency check is introduced during the positioning stage to eliminate point pairs that do not match the pattern topology; the remaining coordinates are then used for subsequent correlation calculations.

[0136] The correspondence between vibration amplitude and pixel offset is obtained within the same time window. Using the vibration three-component amplitude from the multi-source stability parameters, the synthesized vibration amplitude is calculated according to the dominant frequency energy within the window, and paired with the pattern coordinate difference of adjacent frames to form sample pairs.

[0137] Robust linear regression is performed on the sample pairs to obtain the average pixel offset slope corresponding to a unit amplitude, and the residual distribution is output as uncertainty. This slope constitutes the first-order coefficient of the "vibration to pixel offset" mapping function, and the weights are automatically reduced according to the residual variance in the low amplitude range to avoid false registration.

[0138] The correspondence between warped regions and out-of-focus distributions is indexed by the relative height map, and a sharpness map with the same resolution as the height grid is constructed on the image. For each grid cell, the joint distribution of its sharpness evaluation index and relative height is statistically analyzed. A piecewise monotonic model is used to fit the segments where "sharpness decreases as height increases," obtaining the mapping from local warped height to out-of-focus probability, and an empirical function for the focus step size is also given.

[0139] To reduce interference from structured lighting, sharpness evaluation introduces bandpass weights in the frequency domain to mask low-frequency brightness trends.

[0140] The relationship between tilt angle and perspective distortion is based on the curvature of straight lines. Long edges within the image and reference lines are selected, straight lines are fitted, and the curvature index of the residual curve is calculated as the perspective distortion variable. The two components of the overall tilt angle are jointly regressed with this index, constrained to a linear model passing through the origin, resulting in a mapping matrix from tilt to distortion, which serves as prior information for subsequent homography matrix calculation.

[0141] The magnitude estimate of the quadratic term in the nonlinear part of the residual is retained as a warning signal of model mismatch, triggering incremental iteration when geometric correction fails to converge.

[0142] The correspondence between lateral offset and wave characteristics is fitted using a family of sinusoidal functions. Using the lateral offset sequence given by the encoder as the driving quantity, the lateral offset of the pattern coordinates is paired with it, and the amplitude, period, and phase are obtained by weighted least squares fitting.

[0143] The weights are derived from the product of the encoder's trusted label and the pattern matching confidence level; low-confidence samples are automatically weighted less. The triples of the fitted output constitute a periodic compensation function, which, together with bandwidth constraints, is smoothly updated on the time axis to ensure gradual parameter changes.

[0144] After obtaining the above four types of mappings, they are assembled into a feature mapping function group. Each function includes an input domain definition, calculation scope, output domain, and uncertainty description, and records the applicable sampling rate and scanning speed range. To prevent mutual interference, interpretation priorities are set within the function group: periodic fluctuations are given priority to interpret high-frequency coordinate offsets, vibration mapping handles the remaining short-term misalignments, geometric perspective only applies to global coordinate transformations, and local focusing is limited to the defocused mask area.

[0145] The compensation threshold determination criterion is constructed based on a set of feature mapping functions. The threshold is not fixed as a constant, but is automatically generated by the quantiles of the calibrated samples and managed in batch versioning.

[0146] The vibration trigger threshold is set based on the upper quantile of the residual distribution, the warping threshold is set based on the inflection point of the sharpness curve, the tilt threshold is determined based on the linear interval boundary of the distortion variable, and the start and stop of periodic compensation are given jointly based on the fitting confidence and bandwidth limit. The threshold determination outputs three states: no intervention, mild intervention, and strong intervention, which serve as the entry conditions for the correction rules.

[0147] Feature evaluation functions are designed for various stability parameters. Inter-row displacement is evaluated using a weighted sum of gradient consistency and texture similarity; sharpness is represented by a combination of high-frequency energy and edge sharpness, calculated within a warp mask; curvature is evaluated by normalizing the L2 norm of the reference straight-line residual; and periodicity consistency is evaluated by jointly measuring the variance of the fitted residual and the phase drift rate. The outputs of the evaluation functions are uniformly scaled to the same dimensions to facilitate the synthesis of compensation intensity within the controller.

[0148] Corresponding correction rules are established, specifying triggering conditions, intervention intensity, and execution order. Registration compensation allows only translational degrees of freedom in the mild intervention state, while opening affine degrees of freedom in the strong intervention state. Local focusing sets the step size and upper limit based on the gradient of the evaluation function. Perspective correction enters the iteration with the initial value of the homography matrix; if the quadratic residual remains consistently high, it pauses and backs down. Periodic reverse correction allocates weights according to the fitting confidence level, with only partial intervention in low-confidence segments. The rules also include conflict resolution; when registration and periodic correction produce opposite trends for the same coordinate, the periodic term takes precedence, and its weight is reduced at the registration end.

[0149] Finally, the compensation threshold, evaluation function, and correction rules are integrated into a collaborative compensation model, which is loaded into the system controller as a read-only parameter set and updatable statistics, and marked with version and effective time window. When the controller is running online, it calls the interface of this model to complete the trigger judgment and compensation execution; the evaluation indicators and residuals after execution are written back in the form of samples, which are used to periodically refresh the feature mapping function set and threshold set to maintain the consistency between the model and the operating conditions.

[0150] In one embodiment of the multi-source stability collaborative compensation method for the panel scanning system of this application, it may further include the following:

[0151] Step S601: Read the correction rules of the collaborative compensation model, extract feature points from the continuously acquired image frames, calculate the displacement deviation vector between adjacent image frames, and perform spatial registration transformation on the image based on the displacement deviation vector;

[0152] Step S602: Calculate the sharpness evaluation index of the local area based on the image after spatial registration transformation, compare the sharpness evaluation index with a preset threshold, and drive the focusing mechanism to adjust the focal plane of the area below the threshold.

[0153] This embodiment loads a collaborative compensation model under a unified time base, reads its correction rules and threshold set, and enters the online image sequence registration and focusing linkage process. Image frames enter the processing queue according to the line scan rhythm. When each frame arrives, the controller synchronously obtains the latest snapshot of the vibration amplitude, overall tilt angle, and lateral offset under the corresponding frame index, which is used to filter usable features and limit the registration degrees of freedom. To reduce latency, a two-level buffer is established: a fast loop processes the registration of adjacent frames, and a slow loop maintains sharpness statistics and focus step size estimation.

[0154] Feature point extraction is based on an evaluation function and masking strategy defined by a collaborative compensation model. In the current frame, a low-confidence region mask is first generated based on the local warp height map to avoid using obviously out-of-focus blocks for geometric estimation. In the unmasked regions, several candidate feature points are selected by weighted scoring of gradient consistency and corner response, and sub-pixel localization is performed.

[0155] To improve robustness, point pairs with boundary neighborhoods and duplicate textures are removed from the candidate set, while a uniformly distributed feature set is retained. The scale and orientation of each feature are recorded for subsequent multi-scale matching.

[0156] The displacement deviation vector between adjacent frames is obtained through coarse-to-fine pyramid matching. The coarse layer quickly estimates large displacements using integer pixel correlation, while the fine layer minimizes the joint cost of gradient difference and texture error on the sub-pixel grid to obtain the displacement vector and matching confidence of each feature.

[0157] After matching, robust estimation is used to remove outliers, and a global motion model is calculated. Based on the intervention intensity of the collaborative compensation model, if the vibration triggering is in a mild intervention state, a translational model is fitted; if it is in a strong intervention state and the outlier rate is low, an open affine model is used to absorb slight shearing. The obtained spatial registration transform is then reversed and resampled in the current frame using bicubic interpolation to preserve high-frequency details.

[0158] To avoid redundant compensation with periodic fluctuation correction, a suppression weight is applied to the mid-to-high frequency lateral offset components during the registration stage. This weight comes from the bandwidth decomposition of the upstream lateral offset sequence, limiting registration to absorb only low-frequency components.

[0159] The registration transformation matrix and residual statistics are written into the slow loop, becoming the initial values ​​and quality judgment basis for the next frame. When the residual continues to exceed the quantile threshold given by the model, the affine degrees of freedom are automatically reduced, reverting to pure translation to prevent overfitting from causing jitter at the stitching edges.

[0160] After spatial registration is completed, a local sharpness evaluation index is calculated on the registered image. The sharpness evaluation follows the model's caliber, is represented by a weighted sum of high-frequency energy and edge sharpness, is calculated within a fixed-size sliding window, and is superimposed on a local warp height map to obtain a sharpness heatmap.

[0161] To suppress misjudgments caused by uneven illumination, the mean brightness of each window is adaptively normalized to ensure that the evaluation value primarily reflects image sharpness rather than brightness bias. The evaluation results are bound to the frame index and window coordinates to form structured sharpness entries.

[0162] The sharpness evaluation index is compared with a preset threshold, which is derived from the quantile rules of the calibrated samples and is kept consistent across batches.

[0163] For windows below a threshold, a focal plane adjustment task is generated. The initial adjustment amount references the warp height to focus step size mapping given by the cooperative compensation model, and then micro-step scanning is performed at the center of the window. To avoid frequent reversals of the focusing mechanism, a monotonic acceptance strategy is adopted: a step is confirmed only if the sharpness increases between two adjacent attempts; if it decreases, the step size is reduced and the process is reversed. The upper limit of scanning and the total duration are constrained by the control cycle. Areas not completed within the current frame are marked for delayed processing, and their priority is sorted according to the sharpness gap.

[0164] The focusing mechanism's actions are coupled to adjacent windows. This embodiment utilizes a neighborhood suppression rule to reduce repetitive actions. If a window in the same physical area has already achieved sharpness restoration, adjacent windows that have not reached the threshold are merged and re-evaluated later. After completing one round of focusing, sharpness is quickly reassessed within the calibrated window. If it still has not exceeded the threshold and the improvement trend has stalled, the window is marked as structurally out of focus and processed at a higher level by subsequent perspective correction and periodic reverse correction, avoiding further time consumption on the mechanical side.

[0165] To ensure stable standards for downstream defect detection, parameter snapshots generated during registration and focusing are written to frame-level metadata. This includes registration model type, residual statistics, focusing step size, search count, and final evaluation value. The defect detection module uses this data to filter out low-confidence windows and uses the registration matrix and focusing completion markers during coordinate mapping. The controller maintains parameter version consistency; if abnormal fluctuations occur within a certain window, it automatically reverts to the threshold set and evaluation weights of the previous stable version to maintain continuous output.

[0166] In this process, the inputs are a collaborative compensation model and consecutive image frames, and the outputs are a spatially registered image and a set of regions with improved local sharpness. Intermediate quantities are the displacement deviation vector field and the sharpness heatmap, which support decision-making within each frame and provide samples for subsequent statistical updates. The entire process operates under a fixed time sequence: registration first, then focusing, sharing a unified time base and mask, ensuring that compensation actions do not conflict spatially or overlap temporally.

[0167] In one embodiment of the multi-source stability collaborative compensation method for the panel scanning system of this application, it may further include the following:

[0168] Step S701: Establish a perspective transformation model based on the overall tilt angle, calculate the mapping matrix from image space to physical plane, use the mapping matrix to correct perspective distortion of the image, and generate corrected image data;

[0169] Step S702: Fit a periodic fluctuation function based on the lateral offset sequence, extract the fluctuation amplitude and period parameters, use the fluctuation function to perform reverse compensation calculation on the coordinates of the regular pattern, generate a stable image coordinate sequence, and output the corrected imaging result.

[0170] This embodiment reads two key parameters—overall tilt angle and lateral offset—under a unified time base and coordinate system, and establishes correction links for geometric perspective and periodic fluctuations respectively. Image frames enter this step after registration and local focusing are completed. Relevant metadata includes the two components of the overall tilt angle, the normal estimate of the reference plane, and the observation coordinate sequence of the regular pattern. These quantities will be directly called to avoid redundant calculations.

[0171] A perspective transformation model is established around the overall tilt angle. Based on the normal vector obtained by plane fitting and the pinhole model of imaging geometry, the homography relationship from image space to the physical plane is derived.

[0172] To improve robustness, the degrees of freedom of the homography matrix are constrained by the tilt angle as a priori, retaining only the two main degrees of freedom consistent with pitch and roll, and introducing correction constants for scale and offset to adapt to the pixel coordinate system. The homography solution is validated on the geometric consistency of the feature line. If the residual exceeds the model quantile threshold, incremental iteration is used to correct the tilt angle estimate within a small range, avoiding large steps at once that could lead to overcompensation of distortion.

[0173] The mapping matrix is ​​used to perform perspective distortion correction on the image. To preserve high-frequency details, a reverse mapping strategy is employed, sampling pixel-by-pixel on the physical plane and projecting it back onto the image plane. A bicubic interpolation kernel is used. Cropping is performed at boundaries based on the effective field of view to prevent interpolation from exceeding limits and introducing artifacts.

[0174] The corrected data outputs two types of data: one is the image data itself, and the other is a snapshot of the coordinate transformation, which includes the homography matrix and its confidence index, for subsequent coordinate mapping and defect localization.

[0175] To address the periodic fluctuations caused by the straightness of motion, a periodic fluctuation function is fitted using the lateral offset sequence as the driving force.

[0176] Within the same frame window, the differences between the observed and ideal coordinates of regular patterns are collected to establish a temporal correspondence with the lateral offset. A weighted least squares sinusoidal family model is used, with weights derived from the encoder's trusted labels and the pattern matching confidence level. To avoid parameter jitter, bandwidth constraints are introduced for the amplitude and period parameters, allowing only gradual changes; abrupt changes are suppressed using Kalman updates when necessary. The fitted parameter set serves as the shared state between the current frame and its nearest neighbors, reducing the number of re-estimations.

[0177] Periodic reverse compensation is performed point-by-point. For consistency, this embodiment uses the following relational expression:

[0178] Xcorrection(n) = Xobservation(n) Δ (n),

[0179] Where X is the observation horizontal coordinate of the regular pattern on the calibrated image, and Δ The value of the periodic fluctuation function is obtained based on the lateral offset fitting, and n is the index.

[0180] The Δ Determined by amplitude and period parameters, the phase is estimated from the previous frame and updated within small steps. To avoid overlapping compensation with the previous registration, reverse compensation only applies to mid-to-high frequency components, while low-frequency offset is absorbed by the registration loop. The specific separation is performed according to the bandwidth threshold table.

[0181] After inverse compensation is completed, a stable image coordinate sequence is generated and used together with the homography snapshot obtained from perspective correction for subsequent stitching and defect localization. Stability verification is performed within this frame, and the verification includes whether the short-window variance of the coordinate sequence has decreased to within the quantile threshold, whether the curvature index of the straight line feature has fallen back to the linear range, and whether the phase drift between adjacent frames is continuous. If any of these conditions are not met, the system reverts to the parameter set of the previous version and reduces the weight of this frame to prevent abnormal samples from entering the statistical update.

[0182] To ensure seamless integration between upstream and downstream processes, the corrected image data and stable coordinate sequences are written into frame-level metadata. This metadata records the perspective mapping matrix, periodic function parameters, residual distribution, and confidence labels, and binds them to the input sources and timestamps that generated them. The defect analysis module uses this data to perform inverse coordinate calculations on the physical plane, outputting the defect location and type. The controller monitors the stability of these indicators at the batch level, triggering recalibration of the tilt angle and periodic parameters as needed to maintain consistency between the model and the operating conditions.

[0183] To effectively address the shortcomings of traditional technologies in data processing, model building, and dynamic compensation, and to provide technical support for panel scanning, this application provides an embodiment of a multi-source stability collaborative compensation device for a panel scanning system, which implements all or part of the multi-source stability collaborative compensation method for the panel scanning system. See [link to embodiment]. Figure 2 The multi-source stability collaborative compensation device of the panel scanning system specifically includes the following components:

[0184] The dataset construction module 10 is used to collect vibration acceleration data of the air-floating platform, flatness data of the laser displacement sensor array, and motion straightness data of the guide rail encoder. The vibration acceleration data is analyzed according to three-dimensional components. Based on the flatness data, the local warping height and overall tilt angle parameters of the platform are extracted. The lateral offset between the actual trajectory and the theoretical trajectory of the platform is calculated according to the motion straightness data to construct a multi-source stability parameter dataset.

[0185] The compensation model construction module 20 is used to establish a parameter correlation model based on the multi-source stability parameter dataset, calculate the mapping relationship between vibration amplitude and image pixel offset, local warped area and out-of-focus area, overall tilt angle and perspective distortion, lateral offset and periodic fluctuation of regular pattern, generate a collaborative compensation model including compensation threshold, feature evaluation function and correction rule, and write the collaborative compensation model into the system controller.

[0186] The collaborative compensation module 30 is used to perform dynamic correction based on the collaborative compensation model, perform adjacent frame registration compensation for image misalignment caused by vibration, drive the focusing mechanism to perform local sharpness optimization for the out-of-focus area, calculate the perspective transformation matrix based on the overall tilt angle and perform geometric correction, use the lateral offset to construct a periodic fluctuation function to reverse correct the coordinates of the regular pattern, and output a stable imaging result.

[0187] As described above, the multi-source stability collaborative compensation device for the panel scanning system provided in this application can effectively integrate information through an innovatively designed data processing system, parameter analysis, and feature extraction. It constructs a compensation model mechanism, combining mapping analysis and rule formulation to establish a reliable correction strategy. Dynamic optimization is introduced, and through multi-dimensional correction and collaborative compensation, the stability of the imaging is ensured. This method effectively solves the shortcomings of traditional technologies in data processing, model construction, and dynamic compensation, providing technical support for panel scanning.

[0188] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the multi-source stability collaborative compensation method of the panel scanning system.

[0189] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-source stability collaborative compensation method for the panel scanning system described above.

[0190] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the multi-source stability collaborative compensation method for the panel scanning system described above.

[0191] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0193] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0195] 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 multi-source stability collaborative compensation method for a panel scanning system, characterized in that, The method includes: Vibration acceleration data of the air-floating platform, flatness data of the laser displacement sensor array, and motion straightness data of the guide rail encoder are collected. The vibration acceleration data is analyzed according to three-dimensional components. Based on the flatness data, the local warping height and overall tilt angle parameters of the platform are extracted. The lateral offset between the actual trajectory and the theoretical trajectory of the platform is calculated according to the motion straightness data, and a multi-source stability parameter dataset is constructed. A parameter correlation model is established based on the multi-source stability parameter dataset. The mapping relationship between vibration amplitude and image pixel offset, local warped area and out-of-focus area, overall tilt angle and perspective distortion, and lateral offset and periodic fluctuation of regular pattern is calculated. A collaborative compensation model including compensation threshold, feature evaluation function and correction rule is generated and written into the system controller. Dynamic correction is performed based on the collaborative compensation model. Adjacent frame registration compensation is performed for image misalignment caused by vibration. Local sharpness optimization is performed by driving the focusing mechanism for out-of-focus areas. The perspective transformation matrix is ​​calculated and geometric correction is performed based on the overall tilt angle. The periodic fluctuation function is constructed using the lateral offset to reverse the coordinates of the regular pattern and output a stable imaging result.

2. The multi-source stability collaborative compensation method for the panel scanning system according to claim 1, characterized in that, The vibration acceleration data of the air-floating platform, the flatness data of the laser displacement sensor array, and the motion straightness data of the guide rail encoder are collected. The vibration acceleration data is then analyzed according to three-dimensional components, including: Based on the acquisition of vibration acceleration data stream of air-floating platform, planar height data stream of laser displacement sensor array, and position coordinate data stream of guide rail encoder through multiple sampling channels, the three types of data streams are time-aligned and preprocessed to generate a data sampling sequence under a unified time base; Based on the data sampling sequence, the vibration acceleration data is decomposed into a three-dimensional spatial coordinate system, the relative change of plane height data is calculated, the lateral offset component of position coordinate data is extracted, and a multi-source stability dataset is constructed.

3. The multi-source stability collaborative compensation method for the panel scanning system according to claim 1, characterized in that, The process involves extracting local warping height and overall tilt angle parameters from the platform based on the flatness data, calculating the lateral offset between the platform's actual trajectory and theoretical trajectory based on the motion straightness data, and constructing a multi-source stability parameter dataset, including: The system receives flatness data from a laser displacement sensor and motion trajectory data from a guide rail encoder. Based on the flatness data, it calculates the plane fitting residual, extracts the relative height difference in local areas and the overall surface tilt angle, and generates a platform state feature vector. Based on the comparison between the platform state feature vector and the theoretical uniform linear motion trajectory, the lateral deviation value of the actual motion trajectory is calculated, and a multi-source stability parameter dataset including local warping height, overall tilt angle, and lateral offset is established.

4. The multi-source stability collaborative compensation method for the panel scanning system according to claim 1, characterized in that, The step of establishing a parameter correlation model based on the multi-source stability parameter dataset includes: The multi-source stability parameter dataset is classified and organized according to vibration parameters, flatness parameters, and straightness parameters. A three-dimensional feature space is constructed, and the distribution characteristics of each type of parameter are statistically analyzed to generate a parameter feature mapping matrix. Based on the parameter feature mapping matrix, the correlation coefficients between vibration amplitude and pixel offset, local warping and out-of-focus areas, overall tilt and perspective distortion, and lateral offset and periodic fluctuation are calculated, and a correlation model containing multi-source parameter mapping relationships is established.

5. The multi-source stability collaborative compensation method for the panel scanning system according to claim 1, characterized in that, The calculation of the mapping relationships between vibration amplitude and image pixel offset, local warped and out-of-focus areas, overall tilt angle and perspective distortion, and lateral offset and periodic fluctuations of regular patterns generates a collaborative compensation model including compensation thresholds, feature evaluation functions, and correction rules. This collaborative compensation model is then written into the system controller, including: Dynamically acquire images of the calibration plate, extract the position coordinate sequence of regular patterns, calculate the quantitative correspondence between vibration amplitude and pixel offset, warped area and out-of-focus distribution, tilt angle and perspective distortion, lateral offset and wave characteristics, and generate a set of feature mapping functions. Based on the aforementioned feature mapping function set, a compensation threshold determination criterion is constructed, feature evaluation functions for various stability parameters are designed, corresponding correction rules are formulated, and the compensation threshold, evaluation functions, and correction rules are integrated into a collaborative compensation model, which is then written into the system controller.

6. The multi-source stability collaborative compensation method for the panel scanning system according to claim 1, characterized in that, The dynamic correction based on the collaborative compensation model includes performing adjacent frame registration compensation for image misalignment caused by vibration, and driving the focusing mechanism to perform local sharpness optimization for out-of-focus areas, including: Read the correction rules of the collaborative compensation model, extract feature points from continuously acquired image frames, calculate the displacement deviation vector between adjacent image frames, and perform spatial registration transformation on the image based on the displacement deviation vector; The sharpness evaluation index of the local region is calculated based on the image after spatial registration transformation. The sharpness evaluation index is compared with a preset threshold, and the focusing mechanism is driven to adjust the focal plane of the region below the threshold.

7. The multi-source stability collaborative compensation method for the panel scanning system according to claim 1, characterized in that, The process of calculating the perspective transformation matrix based on the overall tilt angle and performing geometric correction, constructing a periodic fluctuation function using the lateral offset to reverse-correct the coordinates of the regular pattern, and outputting a stable imaging result includes: A perspective transformation model is established based on the overall tilt angle. The mapping matrix from image space to physical plane is calculated. The perspective distortion of the image is corrected using the mapping matrix to generate corrected image data. By fitting a periodic fluctuation function to the lateral offset sequence, the fluctuation amplitude and period parameters are extracted. The fluctuation function is then used to perform reverse compensation calculation on the coordinates of the regular pattern to generate a stable image coordinate sequence, and the corrected imaging result is output.

8. A multi-source stability collaborative compensation device for a panel scanning system, characterized in that, The device includes: The dataset construction module is used to collect vibration acceleration data of the air-floating platform, flatness data of the laser displacement sensor array, and motion straightness data of the guide rail encoder. The vibration acceleration data is analyzed according to three-dimensional components. Based on the flatness data, the local warping height and overall tilt angle parameters of the platform are extracted. The lateral offset between the actual trajectory and the theoretical trajectory of the platform is calculated according to the motion straightness data to construct a multi-source stability parameter dataset. The compensation model construction module is used to establish a parameter correlation model based on the multi-source stability parameter dataset, calculate the mapping relationship between vibration amplitude and image pixel offset, local warped area and out-of-focus area, overall tilt angle and perspective distortion, lateral offset and periodic fluctuation of regular pattern, generate a collaborative compensation model including compensation threshold, feature evaluation function and correction rule, and write the collaborative compensation model into the system controller. The collaborative compensation module is used to perform dynamic correction based on the collaborative compensation model, perform adjacent frame registration compensation for image misalignment caused by vibration, drive the focusing mechanism to perform local sharpness optimization for out-of-focus areas, calculate the perspective transformation matrix based on the overall tilt angle and perform geometric correction, construct a periodic fluctuation function using the lateral offset to reverse correct the coordinates of the regular pattern, and output a stable imaging result.

Images