Picture geometry distortion intelligent control method based on machine vision and image projector

Abstract

The application relates to the technical field of projectors, in particular to a picture geometric distortion intelligent control method based on machine vision and an image projector, the method comprises the following steps: acquiring image data of a projection picture to be corrected, and performing feature analysis on the image data to obtain distortion feature values; matching a correction optimization model corresponding to the distortion feature values to call correction parameters from a preset correction parameter library through the correction optimization model; simulating a correction process corresponding to the projection picture to be corrected according to the correction parameters; adjusting the correction parameters according to the correction process, and correcting the projection picture to be corrected according to the adjusted correction parameters; the application matches the corresponding correction optimization model based on the feature values, can dynamically call adaptive correction parameters from the preset parameter library, changes the limitation of a traditional fixed algorithm, makes the correction strategy highly coupled with an actual distortion scene, and significantly improves correction efficiency and pertinence.

Description

Technical Field

[0001] This invention relates to the field of projector technology, and in particular to a machine vision-based intelligent control method for image geometric distortion and an image projector. Background Technology

[0002] In scenarios where projection display technology is widely used, such as smart conference systems, immersive virtual reality (VR), digital exhibition halls, and home theaters, the diversity of installation positions, projection angles, and screen shapes of projection equipment often leads to geometric distortion of the image, severely affecting the visual effect. Traditional geometric distortion correction methods mostly rely on manual adjustment or automatic correction based on fixed algorithms. The former requires users to have certain operating experience and is time-consuming and labor-intensive, while the latter is difficult to adapt to complex and varied distortion scenarios. For example, in scenarios with different projection distances, curved screens, or multi-projector fusion, fixed correction models and parameter libraries cannot accurately match real-time distortion characteristics, resulting in insufficient correction accuracy or over-correction. Furthermore, the lack of dynamic simulation and parameter optimization mechanisms for the correction process makes it difficult to achieve adaptive high-precision correction in diverse environments.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a machine vision-based intelligent control method for image geometric distortion and an image projector, aiming to solve the technical problem that traditional geometric distortion correction methods mostly rely on manual adjustment or automatic correction based on fixed algorithms. The former requires users to have certain operating experience and is time-consuming and labor-intensive, while the latter is difficult to adapt to complex and ever-changing distortion scenarios.

[0005] To achieve the above objectives, the present invention provides a machine vision-based intelligent control method for image geometric distortion, the method comprising:

[0006] Acquire image data of the projected image to be corrected, and perform feature analysis on the image data to obtain distortion feature values;

[0007] Match the correction optimization model corresponding to the distortion feature value, so as to call the correction parameters from the preset correction parameter library through the correction optimization model;

[0008] Simulate the correction process corresponding to the projected image to be corrected based on the correction parameters;

[0009] The correction parameters are adjusted according to the correction process, and the projected image to be corrected is corrected according to the adjusted correction parameters.

[0010] Optionally, the image data includes edge feature parameters. The step of acquiring image data of the projected image to be corrected and performing feature analysis on the image data to obtain distortion feature values ​​includes:

[0011] Obtain the edge feature parameters of the projection image to be corrected, wherein the edge feature parameters include edge curvature parameters and pixel offset parameters;

[0012] The Hough transform is used to extract the coordinates of key points from the edge feature parameters, and the projection plane equation and distortion mapping equation are constructed based on the coordinates of the key points to obtain the distortion coefficient parameters.

[0013] The distortion vector is obtained based on the distortion coefficient parameters, and the distortion vector is normalized to obtain the distortion feature value, wherein the distortion vector includes curvature, offset and rotation angle.

[0014] Optionally, the step of calling calibration parameters from a preset calibration parameter library through the calibration optimization model includes:

[0015] Obtain the parameter features corresponding to each parameter combination in the correction parameter library, and calculate the similarity between each parameter feature and the distortion feature value through the mapping relationship preset in the correction optimization model;

[0016] When the similarity exceeds a preset threshold, a preset correction and optimization rule is invoked to optimize the current parameter combination and generate the correction parameters corresponding to the optimized parameter combination.

[0017] Optionally, the step of simulating the correction process corresponding to the projected image to be corrected based on the correction parameters includes:

[0018] A virtual projection geometric model is constructed based on the discrete control points of the correction parameters, and the virtual projection geometric model is meshed using quadrilateral mesh elements to obtain a meshed model.

[0019] A pixel mapping equation is established using a bilinear interpolation function, and the deformation field data and brightness field data of the mesh subdivision model are calculated using the pixel mapping equation and a ray tracing solver.

[0020] The feature mapping relationship between the deformation field data and the brightness field data is obtained by extracting features from the deformation field data and the brightness field data through a three-layer convolutional neural network.

[0021] The real-time image data captured by the camera is processed according to the feature mapping relationship to obtain the coordinate offset data and brightness change data of the projected image to be corrected during the correction process.

[0022] By fusing the coordinate offset data and the brightness change data using a particle filter, the actual distortion curve and the actual brightness distribution curve are obtained.

[0023] The reliability of the correction parameters is verified based on the actual distortion curve, the actual brightness distribution curve, the deformation field data, and the brightness field data.

[0024] Optionally, adjusting the correction parameters according to the correction process and correcting the projected image to be corrected according to the adjusted correction parameters includes:

[0025] When the reliability of the correction parameters is verified, a fuzzy evaluator based on confidence intervals is used to calculate the deformation evaluation parameters and brightness evaluation parameters corresponding to the correction parameters, and a sequence of parameters to be adjusted is generated based on the comparison results of the deformation evaluation parameters and the brightness evaluation parameters with preset thresholds respectively.

[0026] A gradient boosting tree regressor is trained based on the sequence of parameters to be adjusted. The parameter adjustment amount is predicted by the gradient boosting tree regressor. The parameter value of the correction parameter is adjusted according to the parameter adjustment amount to obtain the discrete control point coordinates of the updated correction parameter.

[0027] The coordinates of the discrete control points of the updated correction parameters are fitted with NURBS curves, and dynamic speed planning is performed based on the NURBS curves under the constraints of maximum pixel displacement and brightness change rate to obtain the speed planning curve.

[0028] The curvature values ​​at each discrete point of the velocity planning curve are calculated. Transition points are inserted in the curvature abrupt change region by spline interpolation to obtain the adjusted correction parameters. The projected image to be corrected is then corrected according to the adjusted correction parameters.

[0029] Optionally, before acquiring the image data of the projected image to be corrected, the method further includes:

[0030] The material parameters of the projection screen are obtained to construct a geometric distortion static model corresponding to the projection screen and to perform correction analysis on the geometric distortion static model to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to the projection image under each preset initial correction parameter.

[0031] Extract the discrete points of each initial correction parameter in the screen deformation cloud map and the edge gradient in the brightness distribution cloud map, and calculate the correction speed and brightness change rate corresponding to each initial correction parameter at each discrete point to obtain the screen deformation change curve of the projected image during the correction process under different initial correction parameters.

[0032] Based on the screen deformation change curve, select multiple first correction parameters from a plurality of initial correction parameters;

[0033] The first parameter in each of the first correction parameters is optimized according to the preset parameter optimization algorithm to obtain several optimized parameter combinations and save them to the preset correction parameter library.

[0034] Optionally, the material parameters include surface roughness parameters and material refractive index. The process of obtaining the material parameters of the projection screen, constructing a geometric distortion static model corresponding to the projection screen, and performing correction analysis on the geometric distortion static model to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to the projected image under each preset initial correction parameter includes:

[0035] The surface roughness parameters and material refractive index of the projection screen are obtained to calculate the distortion stiffness matrix corresponding to the projection screen. The surface roughness parameters include microstructure depth and distribution density, and the distortion stiffness matrix is ​​determined by the material refractive index, microstructure depth and distribution density.

[0036] The distortion values ​​of pixel nodes are obtained through the distortion stiffness matrix and mapped to the mesh nodes to establish a bicubic spline interpolation equation. The mesh nodes are divided into quadrilateral elements and the mesh size is reduced to one-third of the basic mesh size in the edge region.

[0037] The edge curvature value is calculated based on the bicubic spline interpolation equation to obtain the distortion field calculation equation, i.e., the geometric distortion static model; wherein, the edge curvature value is determined by the pixel gradient vector and the normal vector;

[0038] A set of nonlinear pixel displacement equations is established based on the distortion field calculation equations to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to each preset initial correction parameter of the projected image; wherein, the set of nonlinear pixel displacement equations is solved using the quasi-Newton method until the residual is less than the preset value.

[0039] Optionally, calculating the correction speed and brightness change rate corresponding to each of the initial correction parameters at each of the discrete points to obtain the screen deformation change curve of the projected image during the correction process under different initial correction parameters includes:

[0040] Obtain the parameter curve corresponding to each of the initial correction parameters, and obtain the coordinate dataset of discrete points on the parameter curve; wherein, different initial correction parameters correspond to different curve shapes, curve lengths and curve curvatures;

[0041] The brightness distribution value is obtained by fitting the discrete points with a Bézier curve; wherein, the brightness distribution value is calculated through the mapping relationship between curve parameters and brightness.

[0042] An optical kinematic equation is established based on the brightness distribution value, and the brightness change rate at discrete points is obtained by solving the Euler integral formula; wherein the brightness change rate is calculated by integrating brightness over time.

[0043] A state equation is constructed based on the brightness change rate of the discrete points and the deformation on the screen deformation cloud map, and the deformation time function is obtained by solving it using the least squares method. The screen deformation change curve is obtained based on the deformation time function.

[0044] Optionally, the step of selecting multiple first correction parameters from a plurality of initial correction parameters based on the screen deformation change curve includes:

[0045] The current screen deformation corresponding to each initial correction parameter is determined by the screen deformation change curve. Constraint equations are constructed based on the preset maximum deformation of the projected image and the current screen deformation corresponding to each initial correction parameter, so as to obtain the deformation constraint optimization parameters through conjugate gradient iteration calculation.

[0046] Based on the deformation constraint optimization parameters, a multi-objective optimization function is established for the deformation objective function and the brightness objective function, and the mapping relationship between the parameter vector and the constraint response vector corresponding to each initial correction parameter is obtained through a support vector machine with radial basis function kernel.

[0047] Based on the mapping relationship, a set of constrained optimization equations is constructed using KKT conditions, and the optimal correction parameters and Lagrange multipliers corresponding to each initial correction parameter are obtained by solving the sequential quadratic programming method.

[0048] If the norm difference between the optimal correction parameter and the initial parameter corresponding to the initial correction parameter is less than a preset convergence threshold, then the optimal correction parameter is stored as the first correction parameter.

[0049] Furthermore, to achieve the above objectives, the present invention also provides an image projector, the image projector comprising: a memory, a processor, and a machine vision-based intelligent control program for image geometric distortion stored in the memory and executable on the processor, the machine vision-based intelligent control program for image geometric distortion being configured to implement the steps of the machine vision-based intelligent control method for image geometric distortion as described above.

[0050] This invention provides a machine vision-based intelligent control method for geometric distortion of images. The method acquires image data of the image to be corrected in real time using machine vision and extracts distortion feature values, avoiding the tedious manual adjustments of traditional methods and achieving precise quantification of distortion features. Based on the feature value matching and corresponding correction optimization model, it can dynamically call suitable correction parameters from a preset parameter library, overcoming the limitations of traditional fixed algorithms. This highly couples the correction strategy with the actual distortion scenario, significantly improving correction efficiency and targeting. A correction process simulation mechanism is introduced, which simulates the correction effect in advance by calling the correction parameters. The rationality of the correction scheme can be visually evaluated before actual adjustment, avoiding the losses from multiple trial and error caused by direct hardware adjustments. Simultaneously, the dynamic optimization of correction parameters based on simulation results forms a closed-loop control that can adapt to nonlinear distortion problems in complex scenarios, effectively solving the shortcomings of fixed parameters and lack of feedback optimization in traditional methods, and significantly improving correction accuracy and system robustness. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the image projector structure of the hardware operating environment involved in the embodiments of the present invention;

[0052] Figure 2 This is a flowchart illustrating an embodiment of the intelligent control method for image geometric distortion based on machine vision according to the present invention.

[0053] Figure 3 This is a structural block diagram of an embodiment of the intelligent control system for image geometric distortion based on machine vision according to the present invention.

[0054] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0055] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0056] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of an image projector in the hardware operating environment involved in the embodiments of the present invention.

[0057] like Figure 1As shown, the image projector may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, it may also include a standard wired interface or a wireless interface. In this invention, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0058] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the image projector and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0059] like Figure 1 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a machine vision-based intelligent control program for image geometric distortion.

[0060] exist Figure 1 In the image projector shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to peripherals; the image projector calls the machine vision-based intelligent control program for image geometric distortion stored in the memory 1005 through the processor 1001, and executes the machine vision-based intelligent control method for image geometric distortion provided in this embodiment of the invention.

[0061] Based on the above hardware structure, an embodiment of the intelligent control method for image geometric distortion based on machine vision is proposed.

[0062] Reference Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the intelligent control method for image geometric distortion based on machine vision according to the present invention. An embodiment of the intelligent control method for image geometric distortion based on machine vision according to the present invention is presented.

[0063] In one embodiment, the machine vision-based intelligent control method for image geometric distortion includes the following steps:

[0064] Step S100: Obtain image data of the projected image to be corrected, and perform feature analysis on the image data to obtain distortion feature values.

[0065] Image data refers to the collection of visual information of the image projected onto a screen by a projection device. It is typically stored as a pixel matrix and includes information such as brightness, color, and geometric shape. Its purpose is to provide raw data support for subsequent distortion analysis. Image data can be captured in real time through cameras, sensors, or the visual acquisition module built into the projection device. Examples include capturing the actual display effect on the screen at high resolution using dual-camera stereoscopic imaging or structured light projection technology. Feature analysis is a technical process that extracts and quantifies key geometric or visual features from image data using algorithms. Examples include edge detection, corner recognition, and shape matching. Distortion feature values ​​are a set of parameters that quantify the degree of geometric distortion of the image, such as perspective distortion coefficients, nonlinear distortion amplitude, and edge distortion rate. They are used to characterize the degree of stretching, compression, or bending of the projected image in different directions. Image data is optimized through preprocessing, and then feature extraction algorithms are used to locate key geometric features in the image, such as identifying the curvature of calibration grid lines or detecting the deviation of image edges from the ideal geometry. This process achieves automatic quantification of distortion features, avoiding errors from manual measurement.

[0066] Step S200: Match the correction optimization model corresponding to the distortion feature value, so as to call the correction parameters from the preset correction parameter library through the correction optimization model.

[0067] The calibration optimization model is a mathematical or machine learning-based algorithmic framework used to map distortion feature values ​​to optimal calibration parameters. Examples include support vector machines, random forests, or neural network models such as convolutional neural networks (CNNs). A pre-defined calibration parameter library stores a set of validated calibration parameters for different distortion scenarios. These parameters may include focal length adjustments for the projection lens, pixel coordinate mapping tables, and compensation values ​​for overlapping areas from multi-projector fusion. By inputting the extracted distortion feature values ​​into the calibration optimization model, the model uses a feature matching algorithm to retrieve the closest parameter combination from the library. For example, if the feature values ​​indicate significant barrel distortion, the model will use calibration parameters designed for this type of distortion. This process, by dynamically retrieving data instead of using a fixed algorithm, enhances adaptability to complex scenarios.

[0068] Step S300: Simulate the correction process corresponding to the projected image to be corrected based on the correction parameters.

[0069] The calibration process simulation refers to using computer simulation technology to pre-demonstrate the actual effect of calibration parameters on the image in a virtual environment. Examples include using a 3D modeling engine to simulate the projection ray path or generating a calibrated virtual image using image processing algorithms. Visual evaluation presents the simulation results in an intuitive form, allowing the system or user to verify the rationality of the calibration scheme. When constructing the simulation environment, a virtual scene needs to be built based on the physical parameters of the projection device and the screen's geometric model; for example, if the screen is spherical, ray tracing calculations for curved surface reflections are required. The calibration parameters are applied to simulate the calibrated geometry, and the calibrated virtual image is output. After generating the simulation results, the calibration effect is evaluated using quantitative indicators; if the results are not met, subsequent parameter optimization processes are triggered.

[0070] Step S400: Adjust the calibration parameters according to the calibration process, and then calibrate the projected image to be calibrated according to the adjusted calibration parameters.

[0071] Parameter adjustment refers to iterative optimization of the original correction parameters based on simulation results. Examples include fine-tuning parameter values ​​using gradient descent, genetic algorithms, or artificial neural networks. Dynamic optimization closed-loop refers to forming a cyclical mechanism for simulation evaluation and adjustment, continuously correcting parameters until a preset accuracy threshold is met. By comparing simulation results with the ideal correction target, the correction residual of the current parameters is calculated, for example, locating residual local distortion areas in the image. Parameter values ​​are adjusted according to the error distribution, such as enhancing the pixel mapping coefficient in specific areas or adjusting the local keystone correction intensity of the projection device. Finally, the optimized parameters are sent to the projection device, and geometric distortions of the actual projected image are corrected in real time by adjusting lens shift, digital image processing algorithms, or multi-device synchronous control. For example, barrel distortion is compensated by DLP chip pixel offset or edge blending algorithms are used to eliminate splicing seams.

[0072] This embodiment provides a machine vision-based intelligent control method for geometric distortion correction. By capturing image data of the projected image in real time and extracting geometric distortion feature values, it dynamically matches correction parameters from a preset parameter library using a correction optimization model. Further, through simulation verification and iterative parameter optimization, it achieves precise correction, achieving the following technical effects: Machine vision technology directly quantifies the degree of geometric distortion, avoiding the subjectivity and inefficiency of manual measurement; the feature-value-based dynamic parameter calling mechanism improves adaptability to complex scenarios such as curved screens and multi-projection fusion; the closed-loop mechanism of simulation and parameter optimization reduces trial-and-error costs, and virtual verification avoids the cumulative errors of direct hardware adjustments; dynamic closed-loop control can cope with environmental factors such as changes in projection distance or screen deformation, maintaining the long-term stability of the correction effect; the fully automated process reduces the user's operational burden, while providing an intuitive preview of the correction effect through visual simulation, improving user-friendliness. This method, through the combination of machine vision and intelligent optimization, achieves high precision, adaptability, and efficiency in geometric distortion correction, solving the problems of insufficient precision and lack of flexibility in complex environments found in traditional methods.

[0073] In one embodiment, the image data includes edge feature parameters. Image data of the projected image to be corrected is acquired, and feature analysis is performed on the image data to obtain distortion feature values, including:

[0074] Obtain the edge feature parameters of the projected image to be corrected, including edge curvature parameters and pixel offset parameters;

[0075] Edge feature parameters can be the geometric and positional information of edge regions in the projected image, used to accurately capture the distortion patterns at the edges of the image. Edge curvature parameters can be numerical values ​​describing the degree of edge curvature, such as the radius of curvature or the direction of curvature, and can be extracted using image processing algorithms to extract the curvature variation features of the edge contour. Pixel offset parameters can reflect the displacement of edge pixels relative to their ideal positions, for example, by comparing the edge contour of a calibrated pattern with an ideal coordinate system and recording the pixel coordinate deviation value. For example, edge curvature parameters can include the numerical distribution of local radii of curvature, and pixel offset parameters can include displacement vectors in the X-axis or Y-axis directions.

[0076] The Hough transform is used to extract the coordinates of key points from the edge feature parameters, and the projection plane equation and distortion mapping equation are constructed based on the key point coordinates to obtain the distortion coefficient parameters.

[0077] The Hough transform can be a voting-based image processing algorithm used to locate geometrically regular feature points. For example, the Hough transform can convert the image space into a parameter space by inputting edge feature parameters into the algorithm; for instance, when detecting straight lines, pixels are mapped to curves in the parameter space, and geometric parameters are determined by the intersections of these curves. Keypoint coordinates can be a set of coordinates for edge inflection points or structural intersections, such as the intersection of calibration lines or the center of a circle. The projection plane equation can be a plane equation in a three-dimensional coordinate system or a two-dimensional parameterized equation, used to establish an ideal, distortion-free baseline model. The distortion mapping equation can be a mathematical model describing the relationship between actual distortion and the ideal plane, such as a polynomial function containing radial distortion parameters. In one specific embodiment, the parameters of the projection plane equation are determined by fitting the keypoint coordinates using the least squares method, and then the distortion mapping equation is established by comparing the deviations between the actual coordinates and the ideal plane, thereby obtaining the distortion coefficient parameters.

[0078] The distortion vector is obtained based on the distortion coefficient parameters, and the distortion vector is normalized to obtain the distortion feature values. The distortion vector includes curvature, offset and rotation angle.

[0079] Distortion coefficients can be a set of numerical values ​​that quantify distortion types, such as radial distortion coefficients or perspective distortion parameters. Distortion vectors can be a set of distortion coefficient parameters integrated into a vector form, used to facilitate subsequent calculations and model matching. Curvature can be a numerical value describing the curvature change of an edge or the overall image; offset can be the displacement of a pixel position relative to its ideal position; and rotation angle can be the rotational deviation angle of the entire image or a local area. Normalization can be achieved by standardizing the numerical range of the distortion vectors to a fixed interval using a normalization algorithm. For example, if the original range of curvature is [0,2] and the offset range is [5,5] pixels, then after normalization, all parameters are converted to values ​​within the [0,1] interval, thereby eliminating the impact of dimensional differences on subsequent model matching.

[0080] This embodiment captures the curvature and displacement information of edge regions by acquiring edge feature parameters, extracts key points using Hough transform, and constructs a reference plane and distortion mapping model to quantify distortion coefficient parameters. Finally, it integrates the distortion parameters into a standardized vector to eliminate dimensional differences, thereby improving the accuracy and comprehensiveness of distortion feature extraction and enhancing the adaptive correction capability of complex distortion scenes. This method, through multi-dimensional distortion modeling and parameter standardization, can simultaneously characterize perspective distortion, nonlinear distortion, and rotational deviation. Furthermore, it reduces noise interference and model matching errors through robust key point localization and normalization algorithms, ultimately optimizing the dynamic matching efficiency and geometric restoration accuracy of the subsequent correction model.

[0081] In one embodiment, the calibration parameters are called from a preset calibration parameter library through a calibration optimization model, including: obtaining the parameter features corresponding to each parameter combination in the calibration parameter library, and calculating the similarity between each parameter feature and the distortion feature value through a preset mapping relationship in the calibration optimization model; when the similarity exceeds a preset threshold, the preset calibration optimization rule is called to optimize the current parameter combination, and the calibration parameters corresponding to the optimized parameter combination are generated.

[0082] The parameter features can be a set of metadata describing the characteristics of the combination of correction parameters. Their purpose is to provide a quantifiable benchmark for matching parameters with distortion features. Parameter features can be quickly retrieved using the indexing mechanism of the parameter library. For example, parameter features may include the applicable distortion type, the effective projection distance range, and the screen curvature radius matching value. The mapping relationship can be a pre-defined mathematical or logical rule within the correction optimization model, used to quantify the correlation between distortion feature values ​​and parameter features. For example, the mapping relationship may be a non-linear mapping established by the fully connected layers of a neural network, or a matching priority between parameter features and distortion features defined by a rule engine.

[0083] Similarity can be calculated by comparing distortion feature values ​​with parameter features numerically or logically based on mapping relationships. For example, when the distortion feature value indicates that the current image has 15% barrel distortion, the model can use cosine similarity calculation to evaluate the matching degree of parameter combinations marked "applicable to barrel distortion 10% to 20%". The technical effect of this operation is to achieve quantitative matching between parameters and distortion scenes through multi-dimensional weighted calculation (such as assigning weights to features such as distortion type, amplitude, and screen shape and then comprehensively scoring), thereby avoiding the inefficiency of full traversal.

[0084] When the similarity exceeds a preset threshold, a preset correction and optimization rule is invoked to optimize the current parameter combination. The threshold can be a preset numerical benchmark (e.g., 0.8) used to determine whether the matching degree between the parameter combination and the current distortion scene meets the conditions for triggering optimization. The correction and optimization rule can be a preset parameter adjustment algorithm or strategy, which, for example, may include gradient descent iteration, parameter weighted fusion, parameter boundary constraints, etc. The technical effect of this operation is to refine the parameter combination through parameter fine-tuning algorithms. For example, when the focal length adjustment of the original parameter combination is 2mm but the simulation still shows edge distortion, the optimization rule can increase the focal length adjustment to 2.5mm and recalculate the pixel mapping table, thereby improving the matching accuracy.

[0085] The optimized parameter combination and its corresponding correction parameters can be generated through a process that ensures compatibility between parameters. For example, if the original parameters only include lens shift parameters, the optimized parameters may include a brightness compensation coefficient for multi-projector fusion. The technical effect of this operation is to form a composite correction scheme, such as simultaneously solving geometric alignment and color consistency issues in multi-projector fusion scenarios, while avoiding conflicts in other dimensions caused by adjusting a single parameter.

[0086] This embodiment achieves rapid screening of potential candidate parameters through quantitative matching of parameter features and distortion features. It then uses threshold-triggered correction optimization rules to fine-tune parameter combinations and generates final corrected parameters through parameter compatibility verification. This approach improves the accuracy and efficiency of parameter retrieval. Furthermore, the hierarchical strategy of mapping relationships and similarity calculation enhances parameter adaptation accuracy while maintaining processing speed. The structured storage and dynamic computation characteristics of parameter features support flexible system expansion, enabling adaptation to future additions of distortion types or hardware configuration requirements.

[0087] In one embodiment, the calibration process corresponding to the projected image to be calibrated is simulated according to calibration parameters, including:

[0088] A virtual projection geometric model is constructed based on the discrete control points of the correction parameters, and the virtual projection geometric model is meshed using quadrilateral mesh elements to obtain a meshed model.

[0089] Discrete control points can be sets of coordinates defining key deformation nodes in the projected image, such as the four corner points or grid intersections of the projected image. These can be obtained through preset lens offsets or pixel mapping coordinates in the correction parameters. Quadrilateral grid cells can be grid structures that divide the projected image into multiple quadrilateral regions, each grid cell consisting of four vertices. For example, this can be achieved through uniform grid partitioning or adaptive grid generation algorithms. The mesh partitioning model can be a digital representation of a continuous geometric model divided into discrete quadrilateral cells. Its generation process is completed using geometric modeling tools such as OpenGL or MATLAB's mesh generation library. By dynamically associating discrete control points with the vertex coordinates of grid cells, the uniformity or adaptability of the mesh partitioning directly affects the accuracy and efficiency of subsequent calculations. Technical operations implement mesh partitioning through geometric modeling tools. For example, in a projected image with barrel distortion, the algorithm can adaptively increase the grid density in the distorted region to improve local modeling accuracy, thereby providing a structured data foundation for subsequent calculations of the deformation field and brightness field.

[0090] A pixel mapping equation is established using a bilinear interpolation function, and the deformation field data and brightness field data of the mesh subdivision model are obtained by calculating the pixel mapping equation and the ray tracing solver.

[0091] The bilinear interpolation function can be an estimation method based on a weighted average of the vertex values ​​of a two-dimensional mesh, for example, used for calculating the grayscale or color values ​​of pixels with non-integer coordinates. The ray tracing solver can be a tool for simulating the propagation path of light from a projection device to a screen, for example, implemented through a path tracing algorithm. The deformation field data can be a set describing the vertex coordinate offsets of mesh cells, such as X / Y axis displacements; the brightness field data can be distribution data recording changes in the illumination intensity of mesh cells, including light attenuation or superposition effects caused by correction. The technical operation establishes pixel coordinate mapping relationships through the bilinear interpolation function, for example, using vertex coordinate weighting to calculate the deformation mapping of internal pixels when a mesh cell vertex is offset. Combined with the ray tracing solver, the algorithm can simulate the interaction between the ray path and the mesh model, quantifying the deformation field and brightness field data of each mesh cell through numerical integration or ray tracing techniques, thereby accurately predicting the impact of correction parameters on geometry and brightness distribution.

[0092] A three-layer convolutional neural network is used to extract features from deformation field data and brightness field data to obtain the feature mapping relationship between deformation field data and brightness field data.

[0093] The three-layer convolutional neural network can be a neural network architecture containing an input layer, two hidden convolutional layers, and an output layer. For example, a lightweight network such as LeNet5 can be used. The feature mapping relationship can be a description of the correlation between the deformation field and the brightness field learned by the neural network, such as the degree of coupling between brightness changes and geometric deformation. The technical operation involves inputting the deformation field and brightness field data into the network, using convolutional kernels to extract local features. For example, the first convolutional layer extracts the subtle deformation patterns of grid cells, the second convolutional layer fuses the joint features of deformation and brightness, and finally, the output layer generates the feature mapping relationship. This process optimizes the network parameters through the backpropagation algorithm, enabling the model to capture the nonlinear correlation between deformation and brightness changes. For example, edge stretching caused by correction may be accompanied by local overexposure or underexposure, thus providing a feature correlation model for subsequent real-time data processing.

[0094] The real-time image data captured by the camera is processed according to the feature mapping relationship to obtain the coordinate offset data and brightness change data of the projected image to be corrected during the correction process.

[0095] The real-time image data can be the current state information of the projected image continuously captured by a camera, including uncorrected distortion features. Coordinate offset data can be a set of values ​​quantifying the difference in pixel or grid unit coordinates before and after geometric correction, and brightness change data can be a set of values ​​describing the distribution of brightness gain or attenuation caused by correction. The technical operation involves inputting the real-time image into a trained feature mapping model and using a forward propagation algorithm to predict the coordinate offset and brightness change in the current distortion scene. For example, if there is a shadow in the lower left corner of the real-time display screen, the model can combine the feature mapping relationship to infer the brightness attenuation trend that the correction parameters in that area might cause, and output adjustment suggestions to balance geometric correction and brightness uniformity, thus providing data support for subsequent fusion.

[0096] By fusing coordinate offset data and brightness change data using a particle filter, the actual distortion curve and the actual brightness distribution curve are obtained.

[0097] The particle filter can be a state estimation technique based on the Monte Carlo method; for example, a sequential Monte Carlo method can be used. The actual distortion curve can be a quantitative description reflecting the degree of residual distortion of the corrected image as a function of position, while the actual brightness distribution curve can be a regular expression describing the regularity of brightness changes with spatial position after correction. The technical operation generates a large number of particles, each representing a set of possible combinations of coordinate offsets and brightness changes, and calculates particle weights based on real-time data and simulation data. For example, if the camera displays abnormal brightness in a certain area while the simulation predicts reduced distortion, the algorithm can fuse the two types of data through weighted averaging, ultimately outputting actual distortion and brightness distribution curves with higher confidence, thereby effectively reducing noise interference from a single data source.

[0098] The reliability of the correction parameters was verified based on the actual distortion curve, actual brightness distribution curve, deformation field data, and brightness field data.

[0099] Reliability verification can be achieved by comparing simulation predictions with actual measurement data to evaluate the balance between geometric correction and brightness fidelity. Technically, the actual distortion curve output by the particle filter is compared with the deformation field data; if the difference is within a preset error range, the geometric correction is considered effective. Simultaneously, the actual brightness distribution curve is compared with the brightness field data; if the fluctuations meet human eye comfort standards, color fidelity is considered satisfactory. If verification fails, the system can adjust the mesh density or feature extraction parameters, iterating the simulation and verification process again. For example, increasing the mesh density in the distortion region or optimizing the CNN convolution kernel parameters can improve model accuracy.

[0100] This embodiment achieves high-precision deformation modeling by constructing a virtual projection geometric model and performing mesh generation. It combines bilinear interpolation and ray tracing to calculate deformation and brightness field data, utilizes convolutional neural networks to extract feature mapping relationships and process real-time image data, and fuses multi-source data using a particle filter to generate actual distortion and brightness distribution curves. Finally, it verifies the reliability of correction parameters based on simulation and measured data. This approach improves the simulation accuracy of correction parameters in complex scenes, reduces data silos, and supports dynamic parameter optimization. Through multi-technology collaboration, this method overcomes the shortcomings of traditional methods in local distortion modeling, multimodal data fusion, and real-time verification, achieving a dynamic balance between geometric correction and brightness fidelity.

[0101] In one embodiment, the calibration parameters are adjusted according to the calibration process, and the projected image to be calibrated is calibrated according to the adjusted calibration parameters, including:

[0102] When the verification of the calibration parameters is reliable, a fuzzy evaluator based on confidence intervals is used to calculate the deformation evaluation parameters and brightness evaluation parameters corresponding to the calibration parameters, and a sequence of parameters to be adjusted is generated based on the comparison results of the deformation evaluation parameters and brightness evaluation parameters with preset thresholds.

[0103] The confidence interval-based fuzzy evaluator can be an evaluation tool combining statistical confidence intervals and fuzzy logic to quantify parameter reliability. This tool generates multiple possible parameter values ​​using the Monte Carlo method, calculates their statistical distribution, and thus reflects the uncertainty range of the parameter estimation. Deformation evaluation parameters can be indicators that quantify the geometric correction effect, such as the standard deviation of distortion residuals or edge alignment error, mapped to fuzzy sets such as high distortion through the statistical characteristics of the parameter confidence intervals and fuzzy membership functions. Brightness evaluation parameters can be indicators that assess brightness uniformity, such as the inter-region brightness difference coefficient or the proportion of overexposed / underexposed pixels, mapped to fuzzy sets such as "low brightness" through similarity methods. The preset threshold can be a correction target value set according to human visual comfort, such as a maximum allowable distortion residual of 0.5 pixels or a brightness difference not exceeding 10%. Technical operations can be implemented through fuzzy inference rules, such as "if the brightness difference is high and the confidence level is low, then adjust the gain parameter first." A sequence of parameters to be adjusted is generated by comparing whether the lower limit of the confidence intervals of the deformation evaluation parameters and the brightness evaluation parameters exceeds the threshold. This process prioritizes parameters that have the greatest impact on the correction effect, such as lens focal length, rather than edge fine-tuning, thereby determining the order and weight of parameter adjustments.

[0104] A gradient boosting tree regressor is trained based on the sequence of parameters to be adjusted. The parameter adjustment amount is predicted by the gradient boosting tree regressor, and the parameter values ​​of the correction parameters are adjusted according to the parameter adjustment amount to obtain the discrete control point coordinates of the updated correction parameters.

[0105] The gradient boosting tree regressor can be an ensemble learning model, such as a regression model implemented using the XGBoost or LightGBM framework, which gradually reduces prediction error by iteratively stacking decision tree weak learners. The parameter adjustment amount can be the increment or proportion to be applied to the original correction parameters, such as adjusting the lens focal length from 5mm to 5.2mm. The discrete control point coordinates can be the set of coordinates defining key deformation nodes in the projected image within the correction parameters, such as the positions of grid vertices or control handles. Technical operations can be performed by constructing a training set containing historical correction data (such as distortion residuals and brightness differences corresponding to different parameter values) and using a forward stepwise algorithm to fit the residuals between the current model and the target value. For example, if the current brightness difference is 15%, the model predicts that the brightness gain parameter needs to be reduced by 8%. The adjustment amount updates the parameter values ​​using backpropagation gradient descent, ultimately outputting the updated discrete control point coordinates, such as the new coordinate positions of grid vertices. This process ensures the priority adjustment of key parameters (such as master control point coordinates) through model iteration and feature importance analysis.

[0106] The coordinates of discrete control points for the updated correction parameters are fitted using NURBS curves, and dynamic velocity planning is performed based on the NURBS curves under the constraints of maximum pixel displacement and brightness change rate to obtain the velocity planning curve.

[0107] The NURBS curve can be a non-uniform rational B-spline curve, precisely describing complex geometries by controlling vertices, node vectors, and weight parameters. Maximum pixel displacement constraints limit the maximum movement distance of a single pixel during correction, for example, no more than 3 pixels to prevent screen tearing. Brightness change rate constraints control the slope of brightness changes caused by correction, for example, no more than 20% per second to avoid visual flicker. Dynamic velocity planning generates a velocity curve based on path constraints, such as reducing the adjustment speed in sensitive areas to improve accuracy. Technical operations can use discrete control point coordinates as control vertices of the NURBS curve, combined with non-uniform node vector fitting to smooth the path. For example, if a region needs to move 5 pixels but is constrained by 3 pixels / second, a quadratic programming algorithm is used to generate the velocity curve in stages. This curve also considers the physical hardware's motion capabilities, ensuring the adjustment process is reliably executed within mechanical performance boundaries.

[0108] The curvature values ​​at each discrete point of the velocity planning curve are calculated. Transition points are inserted in the curvature abrupt change region by spline interpolation to obtain the adjusted correction parameters. The projected image to be corrected is then corrected based on the adjusted correction parameters.

[0109] The curvature value can be a geometric quantity describing the degree of curvature of a curve at a point, such as the reciprocal of a pixel. Spline interpolation can generate a smooth transition curve by fitting discrete data points with a piecewise polynomial function; for example, cubic spline interpolation ensures the continuity of the first and second derivatives. Transition points can be newly added control points at points of curvature abrupt change, such as inserting intermediate control points at curvature peaks to mitigate drastic path changes. Technical operations can calculate the curvature of each point on the curve. If the curvature of a region exceeds a threshold, such as 1 / pixel, transition points are generated near the abrupt change point using cubic spline interpolation. For example, if the NURBS curve indicates that a region needs to be shifted upwards by 2 pixels, the drive system adjusts step-by-step according to the speed planning curve, and finally sends the coordinate set containing the original control points and transition points to the projection device for correction.

[0110] This embodiment quantifies parameter reliability and generates adjustment sequences using a confidence interval-based fuzzy evaluator. It then combines this with a gradient boosting tree regressor to predict parameter adjustments and optimize discrete control point coordinates. NURBS curves and dynamic velocity planning ensure the adjustment path remains smooth and continuous under physical constraints. Finally, curvature analysis and spline interpolation eliminate path abrupt changes, significantly improving the precision and stability of parameter adjustment. This method, through multi-level optimization techniques, transforms parameter adjustment from coarse-grained to precise and controllable, reducing visual persistence and hardware consumption in complex scenarios while shortening convergence time, achieving high-precision, low-energy real-time correction.

[0111] In one embodiment, before acquiring the image data of the projected image to be corrected, the method further includes:

[0112] Obtain the material parameters of the projection screen to construct a geometric distortion static model corresponding to the projection screen and perform correction analysis on the geometric distortion static model to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to each preset initial correction parameter.

[0113] The material parameters of the projection screen can be a set of numerical values ​​describing the screen's physical properties, obtained through sensor measurement or manual input. Examples include the bending stiffness of a flexible screen, the radius of curvature of a curved screen, or surface roughness. The geometric distortion static model can be a mathematical model based on materials mechanics and structural analysis, used to quantify the influence of different correction parameters on screen deformation. For example, a meshed mechanical model can be established through finite element analysis (FEA). The screen deformation cloud map can be a deformation distribution map in the form of a two-dimensional heat map. For example, it can show the displacement of different areas of a curved screen or the local unevenness of a flat screen. The brightness distribution cloud map can be a distribution map describing the brightness gradient of the projected image. For example, it can reflect the brightness attenuation characteristics of edge areas due to material absorption or diffuse reflection.

[0114] A geometric distortion static model is constructed and corrective analysis is performed. For example, material parameters can be input into the model to simulate the impact of different initial correction parameters (such as lens offset or digital correction coefficients) on screen deformation and brightness. For instance, stress distribution under a specific focal length adjustment can be calculated using finite element simulation to generate a deformation cloud map, or a brightness distribution cloud map can be generated by simulating the reflection path using a ray tracing algorithm. The multi-parameter traversal process performs the above analysis on each preset set of initial correction parameters, generating deformation and brightness cloud maps for the corresponding parameters. This process is completed offline, providing basic data for subsequent screening.

[0115] Extract the discrete points of each initial correction parameter in the screen deformation cloud map and the edge gradient in the brightness distribution cloud map, and calculate the correction speed and brightness change rate corresponding to each initial correction parameter at each discrete point to obtain the screen deformation change curve of the projected image during the correction process under different initial correction parameters.

[0116] In this model, discrete points can be key sampling points in the deformation cloud map used to quantify local deformation, typically distributed at screen edges, corners, or areas prone to distortion. Edge gradients can be the slope of brightness change in edge regions of the brightness distribution cloud map, typically obtained by detecting brightness differences between edge pixels using the Sobel operator. Correction speed can be the rate of change of screen deformation per unit parameter adjustment, typically calculated using the time derivative or parameter derivative of the deformation data. Brightness change rate can be the ratio of edge gradient adjustment to correction parameters, typically reflecting the sensitivity of brightness attenuation to parameter changes. Deformation change curves can be a visual model of the parameter deformation relationship, typically generated by fitting correction speed and deformation data under different initial correction parameters.

[0117] The process involves extracting discrete points and edge gradients and calculating dynamic parameters. For example, this includes selecting predefined discrete points in the deformation contour map to record deformation values ​​and calculating gradient values ​​in the edge region in the brightness contour map. Dynamic parameter calculation combines deformation and brightness data, such as obtaining the correction rate and brightness change rate through derivative operations. Deformation change curve generation involves fitting deformation data with different initial parameters into a curve, forming a visual model of the relationship between parameters and deformation; for example, curve peaks may indicate correction critical points.

[0118] Based on the screen deformation change curve, select multiple first correction parameters from several initial correction parameters;

[0119] The first correction parameter can be a set of candidate parameters that meet indicators such as deformation convergence rate or residual distortion rate. The screening process, for example, includes setting a convergence threshold for the deformation change curve, such as a deformation rate below 0.1 mm / s or residual distortion less than 1%, to eliminate initial parameters that do not meet the conditions. Multi-objective optimization uses Pareto front analysis or a weighted scoring method to comprehensively consider indicators such as deformation and brightness change rate, selecting parameters that achieve a balance between deformation correction and brightness preservation. For example, a parameter that causes a sudden drop in edge brightness may be excluded.

[0120] The first parameter in each first correction parameter is optimized according to the preset parameter optimization algorithm to obtain several optimized parameter combinations and save them to the preset correction parameter library.

[0121] The parameter optimization algorithm can be an algorithm that iteratively improves the performance of candidate parameters, such as particle swarm optimization (PSO), genetic algorithms, or gradient descent. The parameter combination can be a vector containing multi-dimensional correction parameters, such as focal length, trapezoidal correction angle, or edge compensation coefficients. The correction parameter library can be a database storing the optimized parameter set, supporting the overwriting or supplementation of existing parameters to improve adaptability.

[0122] The process of optimizing parameter combinations can be exemplified by dynamically adjusting sub-parameters, such as lens displacement or digital correction coefficient weights, through algorithms to minimize deformation residuals or maximize brightness uniformity. For instance, genetic algorithms generate new parameter combinations through crossover mutation and verify their performance through simulation. Parameter combination generation must ensure synergy between sub-parameters; for example, adjusting the focal length requires coordination with pixel mapping parameters to avoid edge distortion. Parameter library updates store the optimized parameter combinations in a database, forming a self-evolving parameter library to adapt to new screen materials or complex scene requirements.

[0123] This embodiment obtains the material parameters of the projection screen and constructs a geometric distortion static model to generate deformation and brightness cloud maps corresponding to different initial correction parameters; extracts key points and gradient data to calculate dynamic parameters and generate deformation change curves; filters candidate parameters based on the curves and performs multi-objective optimization; finally, it generates parameter combinations and updates the correction parameter library through parameter optimization algorithms, achieving the following technical effects: Material-driven model construction deeply binds the correction schemes in the parameter library to the physical properties of the screen, reducing correction deviations caused by material differences; parameter selection using composite indicators such as deformation change curves and brightness gradients improves the robustness and environmental adaptability of parameter combinations; pre-generated parameter libraries reduce the computational load of real-time correction, enabling subsequent steps to quickly respond to environmental changes; the dynamic parameter library update mechanism supports adaptive expansion for new screen materials or complex scenarios, improving the overall flexibility and efficiency of the solution.

[0124] In one embodiment, the material parameters include surface roughness parameters and material refractive index. The material parameters of the projection screen are obtained, a geometric distortion static model corresponding to the projection screen is constructed, and the geometric distortion static model is corrected and analyzed to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to the projected image under each preset initial correction parameter, including:

[0125] The surface roughness parameters and material refractive index of the projection screen are obtained to calculate the distortion stiffness matrix corresponding to the projection screen. The surface roughness parameters include the microstructure depth and distribution density, and the distortion stiffness matrix is ​​determined by the material refractive index, microstructure depth and distribution density.

[0126] Surface roughness parameters can be quantitative indicators describing the microstructural characteristics of the screen surface, and can be measured by three-dimensional microscopic imaging or white light interferometry. For example, microstructure depth can be the vertical height of the fiber indentation in a fabric screen or the depth of the etched groove in a metal screen, and distribution density can be the number or area ratio of microstructures per unit area. Material refractive index can be the relative change of the screen material with respect to the speed of light propagation, and can be obtained by spectrophotometry. Distortion stiffness matrix can be a coupled mathematical expression of mechanical and optical properties, and its element values ​​are jointly determined by the above parameters, used to characterize the deformation response characteristics of the screen under different correction parameters. In technical operation, the refractive index, microstructure depth, and distribution density are substituted into the mechanical-optical coupling model. Through parameter quantification and matrix construction processes, such as increasing the local stiffness coefficient by increasing the microstructure depth, or adjusting the light path by using high refractive index materials to affect the deformation distribution, a distortion stiffness matrix that can reflect the material properties is established.

[0127] The distortion values ​​of pixel nodes are obtained by the distortion stiffness matrix and mapped to the mesh nodes to establish a bicubic spline interpolation equation. The mesh nodes are divided into quadrilateral elements and the mesh size is reduced to one-third of the basic mesh size in the edge region.

[0128] Pixel node distortion values ​​can be the deformation displacement of discrete points on the screen surface, which can be obtained through mechanical analysis of the stiffness matrix. Quadrilateral element partitioning can be a discretization method that divides the screen surface into rectangular mesh elements; for example, the base mesh size can be set to 5mm. Reducing the mesh size in edge regions can be achieved by increasing the local mesh density to one-third of the base size, for example, adjusting the mesh spacing in edge regions to 1.67mm. The bicubic spline interpolation equation can be a high-order continuous surface fitting model based on polynomial functions, and its expression can be represented as... ,in This represents the spatial coordinates of the projection screen surface in a two-dimensional coordinate system, typically with the bottom left corner of the screen as the origin, x representing the horizontal direction and y representing the vertical direction. This represents the deformation displacement at a point (x, y) on the screen surface, that is, the positional shift of this point due to material properties during the projection image correction process. The coefficients of the bicubic polynomial, totaling 16 (4×4), are determined by fitting the distortion values ​​of discrete grid nodes using the least squares method, reflecting the surface deformation morphology of the screen. In the technical operation, the screen surface is discretized through mesh generation, and the distortion values ​​of each pixel node are mapped to the nearest grid node to form a discrete dataset. The coefficients of the equation are then determined using the least squares method, thereby establishing a smooth and continuous deformation field function. For example, using a finer mesh in edge regions can improve the accuracy of local deformation simulation, and bicubic interpolation can more accurately describe the distortion morphology of surface transition regions compared to linear interpolation.

[0129] The edge curvature value is calculated based on the bicubic spline interpolation equation to obtain the distortion field calculation equation, i.e., the geometric distortion static model; where the edge curvature value is determined by the pixel gradient vector and the normal vector;

[0130] Edge curvature values ​​can be parameters representing the degree of bending of the edge region of the screen surface, and their values ​​can be obtained by solving the second derivative of the interpolation equation. Pixel gradient vectors can be vector representations of the direction and magnitude of pixel grayscale changes; for example, they can be calculated using the Sobel operator. Normal vectors can be the vertical vectors at a point on the screen surface, and their direction is determined by the surface curvature. In practice, the radius of curvature of the edge region is calculated by combining the angle between the pixel gradient vector and the normal vector—for example, when the gradient vector and normal vector are nearly perpendicular, it indicates significant bending. The curvature values ​​are then substituted into the distortion field model to form a comprehensive equation encompassing deformation, curvature, and optical properties, thereby establishing a geometric distortion static model that reflects the optical and mechanical properties of the edge region.

[0131] A set of nonlinear pixel displacement equations is established based on the distortion field calculation equations to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to each preset initial correction parameter of the projected image; among them, the nonlinear pixel displacement equations are solved by the quasi-Newton method until the residual is less than the preset value.

[0132] A nonlinear pixel displacement equation set can be a set of equations describing the spatial displacement relationship of pixels, containing nonlinear terms to reflect the effects of material plastic deformation or large deformation. Quasi-Newton methods can be iterative optimization algorithms that accelerate the convergence process by approximating the Hessian matrix. In technical operation, the distortion field model is transformed into pixel-level displacement equations, for example, the expression is... ,in, These are the original coordinates of the pixels in the projected image, corresponding to their ideal positions when the screen is not distorted. , These are the actual coordinates of the pixels after screen deformation, reflecting the positional shift caused by material mechanical deformation and optical properties. , The x and y displacements of a pixel are represented by the refractive index, which is the relative change in the speed of light propagation of the screen material. Measured by a spectrophotometer, it affects the light path and reflection / scattering characteristics, thus influencing the calculation of the brightness distribution cloud map. The microstructure depth is the vertical height of the microstructures on the screen surface, such as the depth of a fabric fiber indentation or the depth of a metal etching groove. It is a surface roughness parameter that affects the local stiffness of the material; the greater the depth, the higher the local stiffness. as well as The deformation functions in the x and y directions are respectively used. A continuous deformation field generated based on bicubic interpolation equations describes the displacement distribution on the screen surface. The equation system is solved using iterative methods such as the Broyden update formula, and a deformation error threshold (e.g., less than 0.01 pixels) is set as the termination condition. At the same time, the refractive index and curvature values ​​are combined to simulate the light reflection path and generate a brightness distribution cloud map. For example, the brightness attenuation in high curvature areas is due to light scattering.

[0133] This embodiment constructs a distortion stiffness matrix by quantifying surface microstructure parameters and optical properties, establishes a deformation field model using adaptive mesh generation and high-order interpolation algorithms, and generates high-precision deformation and brightness cloud maps by combining edge curvature analysis and solving nonlinear equations. This achieves the technical effects of improving the deformation prediction accuracy of flexible screens, optimizing the distortion compensation effect in edge regions, and reducing the computational complexity of large-scale parameter traversal. This method, through the synergistic effect of coupled micro- and macro-parameter modeling, local high-density mesh optimization, and efficient numerical algorithms, solves the problems of insufficient simulation accuracy and low computational efficiency of traditional methods in complex deformation and optical coupling scenarios.

[0134] In one embodiment, the correction rate and brightness change rate corresponding to each initial correction parameter at each discrete point are calculated to obtain the screen deformation change curve of the projected image during the correction process under different initial correction parameters, including:

[0135] Obtain the parameter curve corresponding to each of the initial correction parameters, and obtain the coordinate dataset of discrete points on the parameter curve; wherein, different initial correction parameters correspond to different curve shapes, curve lengths and curve curvatures;

[0136] A parametric curve is a mathematical function describing the mapping relationship between correction parameters and screen response. Its geometric characteristics are determined by the physical properties of the initial correction parameters. For example, a focus adjustment parameter might correspond to a parabolic curve, while a trapezoidal correction parameter might exhibit a linear or exponential curve. This curve can be constructed through experimental measurements or numerical simulations, such as by adjusting lens displacement or digital correction coefficients and recording screen deformation or brightness response data under different parameter values. A discrete point dataset is a set of coordinates obtained by uniformly or non-uniformly sampling along the parametric curve, containing paired data of parameter values ​​and corresponding response values. The differences in curves for different initial correction parameters reflect the nonlinear characteristics of their impact on the screen; for example, a high curvature curve may indicate that the parameter is more sensitive to deformation.

[0137] The brightness distribution value is obtained by fitting the discrete points with a Bézier curve; wherein, the brightness distribution value is calculated through the mapping relationship between curve parameters and brightness.

[0138] Bézier curves are a parametric curve fitting method that defines a smooth curve shape using control points. They are suitable for interpolating discrete data points and generating continuous functions. The brightness distribution value is a continuous function value calculated from the parameters fitted by the Bézier curve and the brightness mapping relationship. This mapping relationship may include optical parameters such as reflectivity and diffuse reflection coefficient. For example, a cubic Bézier curve can fit a naturally transitioning brightness variation trend by adjusting the positions of four control points. The calculation of brightness values ​​requires combining the parameters of the fitted curve with an optical model. For instance, when the curve parameters represent lens offset, the impact of the change in the incident angle caused by the offset on the brightness distribution must be considered.

[0139] An optical kinematic equation is established based on the brightness distribution value, and the brightness change rate at discrete points is obtained by solving the Euler integral formula; wherein the brightness change rate is calculated by integrating brightness over time.

[0140] Optical kinematic equations are dynamic models describing the change of brightness over time, including parameters such as brightness gradient and diffusion coefficient, used to simulate the evolution of brightness distribution during correction. For example, if brightness change is affected by heat conduction, the equations may take the form of thermal equations. By substituting the brightness distribution values ​​into this equation, a differential equation for brightness change over time is established. The Euler integral formula serves as a numerical solution method, estimating the function value at the next time step using the derivative value at the current time step. For example, if the current brightness gradient is ΔL / Δt, then the brightness value at the next time step is L(t+Δt)=L(t)+ΔL. This process transforms the static brightness distribution into a dynamic rate of change, providing a time dimension parameter for deformation analysis.

[0141] A state equation is constructed based on the brightness change rate of the discrete points and the deformation on the screen deformation cloud map, and the deformation time function is obtained by solving it using the least squares method. The screen deformation change curve is obtained based on the deformation time function.

[0142] The state equations are a set of mathematical equations that combine the theoretical rate of change of brightness with the deformation, describing the coupling relationship between the two in the physical process. For example, the deformation rate may be proportional to the brightness gradient. The least squares method solves for the parameters of the deformation-time function by minimizing the sum of squared errors between the theoretical prediction and the actual measurement. The resulting deformation-time function is then visualized as a screen deformation curve, showing the evolution of the deformation over time under different initial correction parameters, such as exponential decay or oscillatory convergence trends.

[0143] This embodiment acquires the geometric features of the parametric curve and samples discrete point data. It then combines this with Bézier curve fitting to generate a continuous brightness distribution model. The dynamic rate of change of brightness is quantified using optical kinematic equations and Euler integrals. Furthermore, deformation and brightness parameters are coupled through state equations, and the deformation time function is optimized using the least squares method, ultimately generating the deformation variation curve. This method, through a deep integration of mathematical modeling and numerical computation, transforms discrete, static parameter analysis into continuous, dynamic multiphysics coupled modeling. This significantly improves the accuracy of deformation variation curve generation and the scientific rigor of parameter selection. Simultaneously, by eliminating discrete data interpolation errors, quantifying dynamic response characteristics, integrating multi-field coupling effects, and optimizing solution efficiency, it enhances the quality of the calibration parameter library and provides a reliable predictive model for real-time calibration in complex dynamic scenarios.

[0144] In one embodiment, a plurality of first correction parameters are selected from a plurality of initial correction parameters based on the screen deformation change curve, including:

[0145] The current screen deformation corresponding to each initial correction parameter is determined by the screen deformation change curve. Constraint equations are constructed based on the preset maximum deformation of the projected image and the current screen deformation corresponding to each initial correction parameter. Deformation constraint optimization parameters are obtained by iterative calculation of conjugate gradient.

[0146] The screen deformation curve can be a quantitative model describing the change in the degree of physical deformation of the screen over time or the magnitude of parameter adjustment under different initial correction parameters. It can be generated by collecting screen surface displacement data using machine vision sensors and fitting the data. For example, the curve can include time-series data of deformation displacement obtained based on strain gauge measurements or optical tracking technology. The preset maximum deformation can be an upper limit of allowable deformation set by the projection equipment manufacturer based on the screen material properties and service life requirements. For example, this value can include a safety deformation threshold for flexible screens under high-frequency vibration. The constraint equation can be a mathematical expression comparing the current deformation with the maximum allowable value, such as deformation ≤ preset maximum deformation. Conjugate gradient iteration can be a numerical optimization method that gradually approximates the optimal solution by constructing a sequence of conjugate directions. Unlike traditional gradient descent, it accelerates the convergence process by reducing the correlation of search directions. In a specific embodiment, when the current deformation corresponding to a certain initial correction parameter exceeds the threshold, conjugate gradient iteration will adjust the parameter value in the reverse direction until the deformation converges within the constraint range, thereby selecting candidate parameters that meet the physical deformation limits.

[0147] Based on the deformation constraint optimization parameters, a multi-objective optimization function is established for the deformation objective function and the brightness objective function. The mapping relationship between the parameter vector and the constraint response vector corresponding to each initial correction parameter is obtained through a support vector machine with radial basis function kernel.

[0148] The deformation objective function can be a mathematical expression quantifying the degree of residual screen deformation, such as a function defined as minimizing the sum of squares of deformation variables. The brightness objective function can be an indicator characterizing the uniformity of the projected image, such as a function minimizing the standard deviation of edge gradients. The multi-objective optimization function can be a comprehensive optimization objective formed by combining the above two objective functions through weighted coefficients. For example, the weight coefficients can be dynamically adjusted according to the application scenario requirements. The radial basis function kernel function can be a nonlinear mapping tool in the form of a Gaussian function, used to map the complex relationship between the initial correction parameter vector and the constraint response vector to a high-dimensional feature space. The support vector machine can be a supervised learning model that maximizes the classification margin or minimizes the regression error. Combined with the radial basis function kernel function, it can effectively capture the nonlinear correlation between parameters and responses, such as modeling the nonlinear response relationship between material deformation characteristics and correction parameters in flexible screen scenarios.

[0149] Based on the mapping relationship, a set of constrained optimization equations is constructed using KKT conditions, and the optimal correction parameters and Lagrange multipliers corresponding to each initial correction parameter are obtained by solving the sequential quadratic programming method.

[0150] KKT conditions can be necessary conditions to ensure that the optimal solution simultaneously satisfies feasibility and optimality, including the satisfaction of inequality constraints such as deformation constraint ≤ maximum allowable value and brightness gradient constraint ≥ minimum threshold. A constrained optimization equation set can transform a multi-objective optimization problem into a set of equations containing objective functions, constraints, and Lagrange multipliers. For example, deformation constraints and brightness constraints can be simultaneously incorporated as hard constraints into the equation system. Sequential quadratic programming can be an iterative solution method that decomposes the original optimization problem into a series of quadratic programming subproblems. Each subproblem contains a linearized objective function and constraints. For example, this method can gradually approach the global optimum by alternately optimizing the parameter vector and Lagrange multipliers. In a specific embodiment, when deformation constraints and brightness constraints conflict, sequential quadratic programming achieves dynamic balance among multiple objectives by adjusting weight coefficients or priority parameters.

[0151] If the norm difference between the optimal correction parameter and the initial parameter corresponding to the initial correction parameter is less than the preset convergence threshold, then the optimal correction parameter is stored as the first correction parameter.

[0152] The norm difference can be a mathematical metric measuring the degree of difference between the optimal correction parameters and the initial parameters. For example, it can be calculated using Euclidean distance or L2 norm to measure the difference between two vectors. The preset convergence threshold can be a termination condition set based on a combination of parameter adjustment accuracy and computational efficiency. For example, this threshold can include a numerical parameter such as 0.01. This convergence criterion can effectively avoid infinite iteration of the algorithm in local optima or oscillating states, ensuring that the selection process terminates and outputs a stable solution under limited computational resources. When the difference is lower than the threshold, it indicates that the parameter adjustment has converged to a stable state that satisfies the constraints and is close to the initial parameters. At this point, storing the optimal parameters as the first correction parameters can ensure the real-time performance and reliability of subsequent correction processes.

[0153] This embodiment extracts real-time deformation data from the screen deformation change curve and constructs constraint equations. It then achieves parameter convergence under constraints through conjugate gradient iteration. By modeling a nonlinear mapping between a multi-objective optimization function and a support vector machine, it balances the conflict between deformation correction and image quality. The multi-objective problem is transformed into a solvable set of constraint optimization equations using KKT conditions and sequential quadratic programming, and the iteration process is terminated by norm difference determination. Finally, a complete closed loop is formed from initial parameter selection to optimal parameter storage. This method, through deep integration of mathematical optimization algorithms and machine learning models, ensures the feasibility of correction parameters under physical deformation constraints, improves the adaptability of parameter selection to nonlinear scenarios, and reduces computational redundancy through convergence condition control. This significantly improves the quality and environmental adaptability of the correction parameter library and reduces the risk of correction failure due to improper parameter selection.

[0154] Furthermore, this embodiment of the invention also proposes a storage medium storing a machine vision-based intelligent control program for image geometric distortion. When the machine vision-based intelligent control program for image geometric distortion is executed by a processor, it implements the steps of the machine vision-based intelligent control method for image geometric distortion described above.

[0155] In addition, refer to Figure 3 This invention also proposes a machine vision-based intelligent control system for image geometric distortion, which includes:

[0156] The data acquisition module 10 is used to acquire image data of the projected image to be corrected, and to perform feature analysis on the image data to obtain distortion feature values;

[0157] The model matching module 20 is used to match the correction optimization model corresponding to the distortion feature value, so as to call the correction parameters from the preset correction parameter library through the correction optimization model;

[0158] The calibration simulation module 30 is used to simulate the calibration process corresponding to the projection image to be calibrated according to the calibration parameters;

[0159] The parameter adjustment module 40 is used to adjust the correction parameters according to the correction process, and to correct the projected image to be corrected according to the adjusted correction parameters.

[0160] Other embodiments or specific implementations of the machine vision-based intelligent control system for image geometric distortion described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0161] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0162] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the module claims listing several systems, several of these systems may be specifically embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as names.

[0163] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal user device (which may be a mobile phone, computer, server, air conditioner, or network user device, etc.) to execute the methods described in the various embodiments of the present invention.

[0164] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A machine vision-based intelligent control method for image geometric distortion, characterized in that, The method includes: Acquire image data of the projected image to be corrected, and perform feature analysis on the image data to obtain distortion feature values; Match the correction optimization model corresponding to the distortion feature value, so as to call the correction parameters from the preset correction parameter library through the correction optimization model; Simulate the correction process corresponding to the projected image to be corrected based on the correction parameters; The correction parameters are adjusted according to the correction process, and the projected image to be corrected is corrected according to the adjusted correction parameters. The step of simulating the correction process corresponding to the projected image to be corrected based on the correction parameters includes: A virtual projection geometric model is constructed based on the discrete control points of the correction parameters, and the virtual projection geometric model is meshed using quadrilateral mesh elements to obtain a meshed model. A pixel mapping equation is established using a bilinear interpolation function, and the deformation field data and brightness field data of the mesh subdivision model are calculated using the pixel mapping equation and a ray tracing solver. The feature mapping relationship between the deformation field data and the brightness field data is obtained by extracting features from the deformation field data and the brightness field data through a three-layer convolutional neural network. The real-time image data captured by the camera is processed according to the feature mapping relationship to obtain the coordinate offset data and brightness change data of the projected image to be corrected during the correction process. By fusing the coordinate offset data and the brightness change data using a particle filter, the actual distortion curve and the actual brightness distribution curve are obtained. The reliability of the correction parameters is verified based on the actual distortion curve, the actual brightness distribution curve, the deformation field data, and the brightness field data.

2. The intelligent control method for image geometric distortion based on machine vision as described in claim 1, characterized in that, The image data includes edge feature parameters. The process of acquiring image data of the projected image to be corrected and performing feature analysis on the image data to obtain distortion feature values ​​includes: Obtain the edge feature parameters of the projected image to be corrected, wherein the edge feature parameters include edge curvature parameters and pixel offset parameters; The Hough transform is used to extract the coordinates of key points from the edge feature parameters, and the projection plane equation and distortion mapping equation are constructed based on the coordinates of the key points to obtain the distortion coefficient parameters. The distortion vector is obtained based on the distortion coefficient parameters, and the distortion vector is normalized to obtain the distortion feature value, wherein the distortion vector includes curvature, offset and rotation angle.

3. The intelligent control method for image geometric distortion based on machine vision as described in claim 1, characterized in that, The step of calling correction parameters from a preset correction parameter library through the correction optimization model includes: Obtain the parameter features corresponding to each parameter combination in the correction parameter library, and calculate the similarity between each parameter feature and the distortion feature value through the mapping relationship preset in the correction optimization model; When the similarity exceeds a preset threshold, a preset correction and optimization rule is invoked to optimize the current parameter combination and generate the correction parameters corresponding to the optimized parameter combination.

4. The intelligent control method for image geometric distortion based on machine vision as described in claim 1, characterized in that, The step of adjusting the correction parameters according to the correction process and correcting the projected image to be corrected according to the adjusted correction parameters includes: When the reliability of the correction parameters is verified, a fuzzy evaluator based on confidence intervals is used to calculate the deformation evaluation parameters and brightness evaluation parameters corresponding to the correction parameters, and a sequence of parameters to be adjusted is generated based on the comparison results of the deformation evaluation parameters and the brightness evaluation parameters with preset thresholds respectively. A gradient boosting tree regressor is trained based on the sequence of parameters to be adjusted. The parameter adjustment amount is predicted by the gradient boosting tree regressor. The parameter value of the correction parameter is adjusted according to the parameter adjustment amount to obtain the discrete control point coordinates of the updated correction parameter. The coordinates of the discrete control points of the updated correction parameters are fitted with NURBS curves, and dynamic speed planning is performed based on the NURBS curves under the constraints of maximum pixel displacement and brightness change rate to obtain the speed planning curve. The curvature values ​​at each discrete point of the velocity planning curve are calculated. Transition points are inserted in the curvature abrupt change region by spline interpolation to obtain the adjusted correction parameters. The projected image to be corrected is then corrected according to the adjusted correction parameters.

5. The intelligent control method for image geometric distortion based on machine vision as described in claim 1, characterized in that, Before acquiring the image data of the projected image to be corrected, the process also includes: The material parameters of the projection screen are obtained to construct a geometric distortion static model corresponding to the projection screen and to perform correction analysis on the geometric distortion static model to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to the projection image under each preset initial correction parameter. Extract the discrete points of each initial correction parameter in the screen deformation cloud map and the edge gradient in the brightness distribution cloud map, and calculate the correction speed and brightness change rate corresponding to each initial correction parameter at each discrete point to obtain the screen deformation change curve of the projected image during the correction process under different initial correction parameters. Based on the screen deformation change curve, select multiple first correction parameters from a plurality of initial correction parameters; The first parameter in each of the first correction parameters is optimized according to the preset parameter optimization algorithm to obtain several optimized parameter combinations and save them to the preset correction parameter library.

6. The intelligent control method for image geometric distortion based on machine vision as described in claim 5, characterized in that, The material parameters include surface roughness parameters and material refractive index. The process of obtaining the material parameters of the projection screen, constructing a geometric distortion static model corresponding to the projection screen, and performing correction analysis on the geometric distortion static model to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to the projected image under each preset initial correction parameter includes: The surface roughness parameters and material refractive index of the projection screen are obtained to calculate the distortion stiffness matrix corresponding to the projection screen. The surface roughness parameters include microstructure depth and distribution density, and the distortion stiffness matrix is ​​determined by the material refractive index, microstructure depth and distribution density. The distortion values ​​of pixel nodes are obtained through the distortion stiffness matrix and mapped to the mesh nodes to establish a bicubic spline interpolation equation. The mesh nodes are divided into quadrilateral elements and the mesh size is reduced to one-third of the basic mesh size in the edge region. The edge curvature value is calculated based on the bicubic spline interpolation equation to obtain the distortion field calculation equation, i.e., the geometric distortion static model; wherein, the edge curvature value is determined by the pixel gradient vector and the normal vector; A set of nonlinear pixel displacement equations is established based on the distortion field calculation equations to obtain the screen deformation cloud map and brightness distribution cloud map corresponding to each preset initial correction parameter of the projected image; wherein, the set of nonlinear pixel displacement equations is solved using the quasi-Newton method until the residual is less than the preset value.

7. The intelligent control method for image geometric distortion based on machine vision as described in claim 6, characterized in that, The calculation of the correction speed and brightness change rate corresponding to each of the initial correction parameters at each discrete point, to obtain the screen deformation change curve of the projected image during the correction process under different initial correction parameters, includes: Obtain the parameter curve corresponding to each of the initial correction parameters, and obtain the coordinate dataset of discrete points on the parameter curve; wherein, different initial correction parameters correspond to different curve shapes, curve lengths and curve curvatures; The brightness distribution value is obtained by fitting the discrete points with a Bézier curve; wherein, the brightness distribution value is calculated through the mapping relationship between curve parameters and brightness. An optical kinematic equation is established based on the brightness distribution value, and the brightness change rate at discrete points is obtained by solving the Euler integral formula; wherein the brightness change rate is calculated by integrating brightness over time. A state equation is constructed based on the brightness change rate of the discrete points and the deformation on the screen deformation cloud map, and the deformation time function is obtained by solving it using the least squares method. The screen deformation change curve is obtained based on the deformation time function.

8. The intelligent control method for image geometric distortion based on machine vision as described in claim 5, characterized in that, The step of selecting multiple first correction parameters from a plurality of initial correction parameters based on the screen deformation change curve includes: The current screen deformation corresponding to each initial correction parameter is determined by the screen deformation change curve. Constraint equations are constructed based on the preset maximum deformation of the projected image and the current screen deformation corresponding to each initial correction parameter, so as to obtain the deformation constraint optimization parameters through conjugate gradient iteration calculation. Based on the deformation constraint optimization parameters, a multi-objective optimization function is established for the deformation objective function and the brightness objective function, and the mapping relationship between the parameter vector and the constraint response vector corresponding to each initial correction parameter is obtained through a support vector machine with radial basis function kernel. Based on the mapping relationship, a set of constrained optimization equations is constructed using KKT conditions, and the optimal correction parameters and Lagrange multipliers corresponding to each initial correction parameter are obtained by solving the sequential quadratic programming method. If the norm difference between the optimal correction parameter and the initial parameter corresponding to the initial correction parameter is less than a preset convergence threshold, then the optimal correction parameter is stored as the first correction parameter.

9. An image projector, characterized in that, The image projector includes: a memory, a processor, and a machine vision-based intelligent control program for image geometric distortion stored in the memory and executable on the processor, wherein the machine vision-based intelligent control program for image geometric distortion is configured to implement the steps of the machine vision-based intelligent control method for image geometric distortion as described in any one of claims 1 to 8.

Images