An AI-driven digital image deformation detection system

By using an AI-driven digital image deformation detection system that combines Fourier feature coding and physical information constraints, the shortcomings of traditional DIC technology in terms of measurement accuracy and physical rationality are solved, achieving high-precision and robust deformation measurement.

CN120599018BActive Publication Date: 2026-11-17CHONGQING UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510709211.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-05-29
Publication Date
2026-11-17
Estimated Expiration
2045-05-29

AI Technical Summary

Technical Problem

Existing digital image correlation technologies have shortcomings in terms of measurement accuracy and physical plausibility, especially in cases of large deformation or discontinuous deformation, where it is difficult to guarantee the accuracy and reliability of measurements. Furthermore, the application of intelligent DIC methods, which rely on a large amount of labeled data, is limited in practical engineering.

Method used

An AI-driven digital image deformation detection system generates multi-scale prediction models of displacement and strain fields through a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, and a physical information constraint unit. Combined with image consistency and physical compatibility constraints, it achieves self-supervised learning.

Benefits of technology

It improves the accuracy and robustness of deformation measurement, ensures the physical compatibility between the displacement field and the strain field, avoids the error amplification problem in traditional methods, requires no large amount of labeled data, and is adaptable to complex deformation modes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120599018B_ABST
    Figure CN120599018B_ABST
Patent Text Reader

Abstract

This invention discloses an artificial intelligence-driven digital image deformation detection system, including a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, a physical information constraint unit, and a result output unit. This invention achieves synchronous and high-precision measurement of displacement field and strain field, avoiding the error amplification problem caused by numerical differentiation in traditional methods.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an artificial intelligence-driven digital image deformation detection system. Background Technology

[0002] Digital image correlation (DIC) is a non-contact optical measurement technique widely used in materials science, mechanical engineering, and civil engineering to measure the displacement and strain fields of object surfaces. Traditional DIC techniques typically employ sub-region correlation-based methods, dividing the images before and after deformation into multiple sub-regions and calculating the displacement information of each sub-region using a matching algorithm. However, this method has several limitations: measurement accuracy heavily depends on the choice of sub-region size; secondly, traditional methods struggle to guarantee the physical compatibility between the displacement and strain fields; and under conditions of large or discontinuous deformation, matching accuracy often decreases significantly.

[0003] In recent years, with the development of deep learning technology, it has become possible to combine artificial intelligence with DIC technology. However, most existing intelligent DIC methods rely on a large amount of labeled data for training, making it difficult to adapt to the data-scarce situation in actual engineering. At the same time, existing methods often lack effective integration of physical constraints, which may lead to measurement results that violate the basic principles of continuum mechanics.

[0004] Therefore, developing an intelligent DIC technology that can simultaneously guarantee measurement accuracy and physical rationality is of great significance for improving the accuracy and reliability of deformation measurement. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-driven digital image deformation detection system, including a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, a physical information constraint unit, and a result output unit;

[0006] The reference image and deformed image processing unit reads the reference image and deformed image data of the surface of the object under test and performs preprocessing to generate a pixel coordinate grid of the reference image and deformed image.

[0007] The Fourier feature coding unit maps the image coordinates in the pixel coordinate grid to a high-dimensional feature space to obtain the feature data of the reference image and the deformed image.

[0008] The physical information constraint unit generates physical information constraints between the displacement field and the strain field;

[0009] The intelligent neural network model unit constructs a multi-scale displacement field and strain field prediction model that satisfies physical information constraints.

[0010] The intelligent neural network model unit inputs the spatial coordinates of the reference image and the deformed image into the multi-scale displacement field and strain field prediction model to obtain the displacement field and strain field analysis results of the object under test.

[0011] The result output unit is used to visualize and export the displacement field and strain field analysis results.

[0012] Furthermore, the reference image and deformed image processing unit includes an image reading module, an image normalization module, and a coordinate grid generation module;

[0013] The image reading module is used to read reference image and deformed image data of the object's surface; the reference image of the object's surface is the image of the object before deformation.

[0014] The image normalization module standardizes the grayscale values ​​of the reference image and the deformed image data.

[0015] The coordinate grid generation module generates a regularized pixel coordinate grid based on the standardized grayscale values.

[0016] Furthermore, the Fourier feature coding unit maps image coordinates in the pixel coordinate grid to a high-dimensional feature space through a multi-scale Fourier feature coding method. The steps include: random frequency matrix generation, coordinate projection transformation, feature scaling, and trigonometric function mapping.

[0017] Furthermore, the intelligent neural network model unit is a multi-branch parallel structure, including multiple parallel processing branches and a fusion output layer;

[0018] Each processing branch corresponds to a Fourier feature at a specific scale, consisting of a multi-layer fully connected network and a nonlinear activation function; the fusion output layer combines the outputs of each branch and outputs the displacement field and strain field predictions simultaneously.

[0019] Furthermore, the physical information constraints include image consistency constraints and physical compatibility constraints;

[0020] The image consistency constraint predicts the displacement field through a neural network, then uses the displacement field to perform pixel-level remapping interpolation on the original reference image to generate a reconstructed deformed image. Then, this reconstructed image is compared pixel by pixel with the actual deformed image captured, and the difference between the two gray values ​​is calculated. This difference measure is used as a constraint term of the loss function.

[0021] This difference value serves as a constraint term in the loss function, guiding the neural network to optimize the accuracy of displacement field prediction and ensuring that the predicted displacement field can accurately transform the reference image into a result consistent with the actual deformed image.

[0022] The physical compatibility constraint is measured by the relationship between displacement gradient and strain prediction value;

[0023] The physical compatibility constraints are as follows:

[0024]

[0025]

[0026] Where u and v are the horizontal and vertical displacements, respectively, and εxx, εyy, and εxy are the normal strain and shear strain components, respectively. (x, y) represents the coordinates.

[0027] Furthermore, the result output unit includes a displacement field visualization module, a strain field visualization module, a vector field visualization module, an equivalent strain calculation module, and a data export module;

[0028] The displacement field visualization module is used to display the horizontal and vertical displacement distribution;

[0029] The strain field visualization module is used to display the distribution of normal strain and shear strain;

[0030] The vector field visualization module is used to display the direction and magnitude of the displacement;

[0031] The equivalent strain calculation module is used to calculate the comprehensive strain index equivalent_strain, that is:

[0032] equivalent_strain=sqrt(2 / 3(exx 2 +eyy 2 +2exy 2 ))(4)

[0033] Where exx and eyy are normal strains, exy is shear strain, and sqrt is the square root function;

[0034] The data export module is used to save the analysis results.

[0035] Furthermore, the objects to be tested include, but are not limited to, metal specimens, engineering composite materials, concrete structures, and rock samples.

[0036] A method for using an AI-driven surface image deformation detection system includes the following steps:

[0037] Step 1: Load the reference image and the deformed image, perform preprocessing, and create a normalized coordinate grid;

[0038] Step 2: Generate multi-scale Fourier features based on preset scale parameters, and map the coordinate information to a high-dimensional feature space;

[0039] Step 3: Construct a multi-branch neural network model and set the number of layers and nodes;

[0040] Step 4: Define a composite loss function that includes image consistency constraints and physical compatibility constraints;

[0041] Step 5: Iterate through the training of the model and update the network parameters until the loss function converges;

[0042] Step 6: Use the trained model to predict the displacement field and strain field, and calculate the equivalent strain;

[0043] Step 7: Visualize the analysis results and save the displacement field and strain field data.

[0044] Furthermore, the Fourier feature generation in step 2 includes: generating a random frequency matrix for each scale, projecting the normalized coordinates onto the frequency matrix, and applying sine and cosine functions for nonlinear mapping to obtain a high-dimensional feature representation.

[0045] Furthermore, the physical compatibility constraint in step 4 is achieved through the following formula:

[0046] diff_exx=u_x(5)

[0047] diff_eyy=v_y(6)

[0048] diff_exy = 0.5 * (u_y + v_x) (7)

[0049] Where u_x and u_y are the partial derivatives of the horizontal displacement u with respect to x and y, respectively; v_x and v_y are the partial derivatives of the vertical displacement v with respect to x and y, respectively; and diff_exx, diff_eyy, and diff_exy are the normal strain and shear strain calculated by the displacement gradient, respectively.

[0050] Furthermore, the equivalent strain in step 6 is as follows:

[0051] equivalent_strain=sqrt(2 / 3(exx 2 +eyy 2 +2exy 2 ))(8)

[0052] Where exx and eyy are normal strains, and exy is shear strain.

[0053] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0054] 1) This invention uses the Fourier feature encoding method, which significantly improves the expressive power and accuracy of the model compared to the traditional direct input of coordinates, and is particularly suitable for capturing complex deformation patterns;

[0055] 2) This invention, through a multi-scale feature extraction strategy, can simultaneously focus on the local details and global features of the deformation field, thereby improving the robustness of the measurement;

[0056] 3) This invention incorporates physical information constraints to ensure physical compatibility between the displacement field and the strain field, thus avoiding unreasonable predictions that violate the principles of mechanics;

[0057] 4) This invention does not require pre-labeled training data, but relies solely on image pairs before and after deformation and physical constraints for self-supervised learning, which greatly improves the practicality of the method;

[0058] 5) This invention achieves synchronous and high-precision measurement of displacement field and strain field, avoiding the error amplification problem caused by numerical differentiation in traditional methods. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the system architecture of the artificial intelligence-driven surface image deformation detection system of the present invention;

[0060] Figure 2 This is a schematic diagram of the multi-scale Fourier feature coding method of the present invention;

[0061] Figure 3 This is a schematic diagram of the intelligent neural network model structure of the present invention;

[0062] Figure 4 This is a schematic diagram illustrating the working principle of the physical information constraint unit of the present invention;

[0063] Figure 5 This is a schematic diagram comparing the displacement field reconstruction and image consistency of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0065] Example 1:

[0066] See Figures 1 to 5 An artificial intelligence-driven digital image deformation detection system includes a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, a physical information constraint unit, and a result output unit.

[0067] The reference image and deformed image processing unit reads the reference image and deformed image data of the surface of the object under test and performs preprocessing to generate a pixel coordinate grid of the reference image and deformed image.

[0068] The Fourier feature coding unit maps the image coordinates in the pixel coordinate grid to a high-dimensional feature space to obtain the feature data of the reference image and the deformed image.

[0069] The physical information constraint unit generates physical information constraints between the displacement field and the strain field;

[0070] The intelligent neural network model unit constructs a multi-scale displacement field and strain field prediction model that satisfies physical information constraints.

[0071] The intelligent neural network model unit inputs the spatial coordinates of the reference image and the deformed image into the multi-scale displacement field and strain field prediction model to obtain the displacement field and strain field analysis results of the object under test.

[0072] The result output unit is used to visualize and export the displacement field and strain field analysis results.

[0073] The reference image and deformed image processing unit includes an image reading module, an image normalization module, and a coordinate grid generation module;

[0074] The image reading module is used to read reference image and deformed image data of the object's surface; the reference image of the object's surface is the image of the object before deformation.

[0075] The image normalization module standardizes the grayscale values ​​of the reference image and the deformed image data.

[0076] The coordinate grid generation module generates a regularized pixel coordinate grid based on the standardized grayscale values.

[0077] The Fourier feature coding unit maps image coordinates in the pixel coordinate grid to a high-dimensional feature space through a multi-scale Fourier feature coding method. The steps include: random frequency matrix generation, coordinate projection transformation, feature scaling, and trigonometric function mapping.

[0078] The intelligent neural network model unit is a multi-branch parallel structure, including multiple parallel processing branches and a fusion output layer;

[0079] Each processing branch corresponds to a Fourier feature at a specific scale, consisting of a multi-layer fully connected network and a nonlinear activation function; the fusion output layer combines the outputs of each branch and outputs the displacement field and strain field predictions simultaneously.

[0080] The physical information constraints include image consistency constraints and physical compatibility constraints;

[0081] The image consistency constraint predicts the displacement field through a neural network, then uses the displacement field to perform pixel-level remapping interpolation on the original reference image to generate a reconstructed deformed image. Then, this reconstructed image is compared pixel by pixel with the actual deformed image captured, and the difference between the two gray values ​​is calculated. This difference measure is used as a constraint term of the loss function.

[0082] This difference value serves as a constraint term in the loss function, guiding the neural network to optimize the accuracy of displacement field prediction and ensuring that the predicted displacement field can accurately transform the reference image into a result consistent with the actual deformed image.

[0083] The physical compatibility constraint is measured by the relationship between displacement gradient and strain prediction value;

[0084] The physical compatibility constraints are as follows:

[0085]

[0086] Where u and v are the horizontal and vertical displacements, respectively, and εxx, εyy, and εxy are the normal strain and shear strain components, respectively. (x, y) represents the coordinates.

[0087] The result output unit includes a displacement field visualization module, a strain field visualization module, a vector field visualization module, an equivalent strain calculation module, and a data export module.

[0088] The displacement field visualization module is used to display the horizontal and vertical displacement distribution;

[0089] The strain field visualization module is used to display the distribution of normal strain and shear strain;

[0090] The vector field visualization module is used to display the direction and magnitude of the displacement;

[0091] The equivalent strain calculation module is used to calculate the comprehensive strain index equivalent_strain, that is:

[0092] equivalent_strain=sqrt(2 / 3(exx 2 +eyy 2 +2exy 2 ))(4)

[0093] Where exx and eyy are normal strains, exy is shear strain, and sqrt is the square root function;

[0094] The data export module is used to save the analysis results.

[0095] The objects to be tested include, but are not limited to, metal specimens, engineering composite materials, concrete structures, and rock samples.

[0096] Example 2:

[0097] A method for using an AI-driven surface image deformation detection system includes the following steps:

[0098] Step 1: Load the reference image and the deformed image, perform preprocessing, and create a normalized coordinate grid;

[0099] Step 2: Generate multi-scale Fourier features based on preset scale parameters, and map the coordinate information to a high-dimensional feature space;

[0100] Step 3: Construct a multi-branch neural network model and set the number of layers and nodes;

[0101] Step 4: Define a composite loss function that includes image consistency constraints and physical compatibility constraints;

[0102] Step 5: Iterate through the training of the model and update the network parameters until the loss function converges;

[0103] Step 6: Use the trained model to predict the displacement field and strain field, and calculate the equivalent strain;

[0104] Step 7: Visualize the analysis results and save the displacement field and strain field data.

[0105] The Fourier feature generation in step 2 includes: generating a random frequency matrix for each scale, projecting the normalized coordinates onto the frequency matrix, and applying sine and cosine functions for nonlinear mapping to obtain a high-dimensional feature representation.

[0106] The physical compatibility constraint in step 4 is achieved by the following formula:

[0107] diff_exx=u_x(5)

[0108] diff_eyy=v_y(6)

[0109] diff_exy = 0.5 * (u_y + v_x) (7)

[0110] Where u_x and u_y are the partial derivatives of the horizontal displacement u with respect to x and y, respectively; v_x and v_y are the partial derivatives of the vertical displacement v with respect to x and y, respectively; and diff_exx, diff_eyy, and diff_exy are the normal strain and shear strain calculated by the displacement gradient, respectively.

[0111] The equivalent strain in step 6 is shown below:

[0112] equivalent_strain=sqrt(2 / 3(exx 2 +eyy 2 +2exy 2 ))(8)

[0113] Where exx and eyy are normal strains, and exy is shear strain.

[0114] Example 3:

[0115] An artificial intelligence-driven digital image deformation detection system includes a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, a physical information constraint unit, and a result output unit.

[0116] The reference image and deformed image processing unit reads the reference image and deformed image data of the surface of the object under test and performs preprocessing to generate a pixel coordinate grid of the reference image and deformed image.

[0117] The Fourier feature coding unit maps the image coordinates in the pixel coordinate grid to a high-dimensional feature space to obtain the feature data of the reference image and the deformed image.

[0118] The physical information constraint unit generates physical information constraints between the displacement field and the strain field;

[0119] The intelligent neural network model unit constructs a multi-scale displacement field and strain field prediction model that satisfies physical information constraints.

[0120] The intelligent neural network model unit inputs the spatial coordinates of the reference image and the deformed image into the multi-scale displacement field and strain field prediction model to obtain the displacement field and strain field analysis results of the object under test.

[0121] The result output unit is used to visualize and export the displacement field and strain field analysis results.

[0122] Example 4:

[0123] An artificial intelligence-driven digital image deformation detection system, with the same technical content as in Embodiment 3, further comprising an image reading module, an image normalization module, and a coordinate grid generation module;

[0124] The image reading module is used to read reference image and deformed image data of the object's surface; the reference image of the object's surface is the image of the object before deformation.

[0125] The image normalization module standardizes the grayscale values ​​of the reference image and the deformed image data.

[0126] The coordinate grid generation module generates a regularized pixel coordinate grid based on the standardized grayscale values.

[0127] Example 5:

[0128] An artificial intelligence-driven digital image deformation detection system, with the same technical content as any one of embodiments 3-4, further wherein the Fourier feature coding unit maps the image coordinates in the pixel coordinate grid to a high-dimensional feature space through a multi-scale Fourier feature coding method, the steps of which include: random frequency matrix generation, coordinate projection transformation, feature scaling and trigonometric function mapping.

[0129] Example 6:

[0130] An artificial intelligence-driven digital image deformation detection system, with the same technical content as any one of embodiments 3-5, further wherein the intelligent neural network model unit is a multi-branch parallel structure, including multiple parallel processing branches and a fusion output layer;

[0131] Each processing branch corresponds to a Fourier feature at a specific scale, consisting of a multi-layer fully connected network and a nonlinear activation function; the fusion output layer combines the outputs of each branch and outputs the displacement field and strain field predictions simultaneously.

[0132] Example 7:

[0133] An artificial intelligence-driven digital image deformation detection system, with the same technical content as any one of embodiments 3-6, further wherein the physical information constraints include image consistency constraints and physical compatibility constraints;

[0134] The image consistency constraint predicts the displacement field through a neural network, then uses the displacement field to perform pixel-level remapping interpolation on the original reference image to generate a reconstructed deformed image. Then, this reconstructed image is compared pixel by pixel with the actual deformed image captured, and the difference between the two gray values ​​is calculated. This difference measure is used as a constraint term of the loss function.

[0135] This difference value serves as a constraint term in the loss function, guiding the neural network to optimize the accuracy of displacement field prediction and ensuring that the predicted displacement field can accurately transform the reference image into a result consistent with the actual deformed image.

[0136] The physical compatibility constraint is measured by the relationship between displacement gradient and strain prediction value;

[0137] The physical compatibility constraints are as follows:

[0138]

[0139] Where u and v are the horizontal and vertical displacements, respectively, and εxx, εyy and εxy are the normal strain and shear strain components, respectively.

[0140] Example 8:

[0141] An artificial intelligence-driven digital image deformation detection system, with the same technical content as any one of embodiments 3-7, further comprising a displacement field visualization module, a strain field visualization module, a vector field visualization module, an equivalent strain calculation module, and a data export module;

[0142] The displacement field visualization module is used to display the horizontal and vertical displacement distribution;

[0143] The strain field visualization module is used to display the distribution of normal strain and shear strain;

[0144] The vector field visualization module is used to display the direction and magnitude of the displacement;

[0145] The equivalent strain calculation module is used to calculate the comprehensive strain index, namely:

[0146] equivalent_strain=sqrt(2 / 3(exx 2 +eyy 2 +2exy 2 ))(4)

[0147] Where exx and eyy are normal strains, and exy is shear strain;

[0148] The data export module is used to save the analysis results.

[0149] Example 9:

[0150] An artificial intelligence-driven digital image deformation detection system, with the same technical content as any one of embodiments 3-8, further wherein the object to be tested includes, but is not limited to, engineering materials and components such as metal specimens, engineering composite materials, concrete structures, and rock specimens.

[0151] This system measures the surface deformation of the object under test, thereby evaluating the mechanical properties of the material (flexural strength, yield strength, etc.), ensuring structural safety, providing a basis for material research and development, structural design, manufacturing process optimization and engineering safety, and significantly improving reliability and economy.

[0152] Example 10:

[0153] A method of using the AI-driven surface image deformation detection system according to any one of embodiments 3-9 includes the following steps:

[0154] Step 1: Load the reference image and the deformed image, perform preprocessing, and create a normalized coordinate grid;

[0155] Step 2: Generate multi-scale Fourier features based on preset scale parameters, and map the coordinate information to a high-dimensional feature space;

[0156] Step 3: Construct a multi-branch neural network model and set the number of layers and nodes;

[0157] Step 4: Define a composite loss function that includes image consistency constraints and physical compatibility constraints;

[0158] Step 5: Iterate through the training of the model and update the network parameters until the loss function converges;

[0159] Step 6: Use the trained model to predict the displacement field and strain field, and calculate the equivalent strain;

[0160] Step 7: Visualize the analysis results and save the displacement field and strain field data.

[0161] Example 11:

[0162] A method for using the AI-driven surface image deformation detection system described in any one of Embodiments 3-9, with the same technical content as Embodiment 10, further comprising the following steps: generating a random frequency matrix for each scale, performing a projection transformation between the normalized coordinates and the frequency matrix, applying sine and cosine functions for nonlinear mapping, and obtaining a high-dimensional feature representation.

[0163] Example 12:

[0164] A method for using the AI-driven surface image deformation detection system described in any one of Embodiments 3-9, with the same technical content as Embodiments 10-11, further wherein the physical compatibility constraint in step 4 is achieved by the following formula:

[0165] diff_exx=u_x(5)

[0166] diff_eyy=v_y(6)

[0167] diff_exy = 0.5 * (u_y + v_x) (7)

[0168] Where u_x and u_y are the partial derivatives of the horizontal displacement u with respect to x and y, respectively; v_x and v_y are the partial derivatives of the vertical displacement v with respect to x and y, respectively; and diff_exx, diff_eyy, and diff_exy are the normal strain and shear strain calculated by the displacement gradient, respectively.

[0169] Example 13:

[0170] A method for using the AI-driven surface image deformation detection system described in any one of Embodiments 3-9, with the same technical content as Embodiments 10-12, further wherein the equivalent strain in step 6 is as follows:

[0171] equivalent_strain=sqrt(2 / 3(exx 2 +eyy 2 +2exy 2 ))(8)

[0172] Where exx and eyy are normal strains, and exy is shear strain.

[0173] Example 14:

[0174] An artificial intelligence-driven surface image deformation detection system includes a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, a physical information constraint unit, and a result output unit.

[0175] The reference image and deformed image processing unit is responsible for reading and preprocessing the measured images. For example... Figure 1 As shown, this unit loads grayscale images of the reference state and the deformed state through the image reading module, performs image standardization to normalize the grayscale value range, and then generates a regularized pixel coordinate grid to lay the foundation for subsequent analysis. In actual operation, the image processing unit first reads the grayscale image, obtains the image size H×L, and then creates a normalized coordinate grid, normalizing the coordinate values ​​to the range of [0,1] to ensure the stability and consistency of subsequent processing.

[0176] Fourier feature coding unit such as Figure 2 As shown, a multi-scale Fourier feature mapping method is employed. This method consists of four main steps: First, a random frequency matrix B is generated for each scale; second, the normalized coordinates are projected onto the frequency matrix to obtain projected values; then, sine and cosine functions are applied to the projected values ​​for nonlinear mapping; finally, features from different scales are merged. By setting different scale parameters (such as σ = 1 and σ = M), both local details and global features of the image can be captured simultaneously, significantly improving feature representation capability and displacement field reconstruction accuracy.

[0177] Intelligent neural network model units such as Figure 3 As shown, a multi-branch parallel network structure is employed. The model contains multiple processing branches, each corresponding to a Fourier feature at a specific scale, and consists of multiple fully connected neural networks. The outputs of each branch are fused and then used in the final output layer to simultaneously predict the displacement field (u, v) and the strain field (εxx, εyy, εxy). Figure 3As shown, each branch network contains an input layer, multiple hidden layers, and an output layer. The hidden layers employ activation functions to enhance nonlinear modeling capabilities. The final fusion layer combines the outputs of each branch and maps them to five output channels through a fully connected network, corresponding to horizontal displacement, vertical displacement, and three strain components, respectively. The model employs initialization methods such as Xavier to improve training stability and convergence speed.

[0178] Physical information constraint unit such as Figure 4 As shown, it includes two parts: image consistency constraints and physical compatibility constraints. Image consistency constraints are implemented by comparing the mean square error between the reconstructed deformed image and the actual deformed image, ensuring the accuracy of the displacement field prediction. Figure 4 As shown on the left, this constraint first reconstructs the deformed image based on the predicted displacement field (u, v), and then calculates the difference between the reconstructed image and the actual deformed image as part of the loss function. Physical compatibility constraints are as follows... Figure 4 As shown on the right, this is achieved by comparing the strain calculated from the displacement gradient with the strain directly predicted by the model, ensuring that the displacement field and strain field satisfy the fundamental relationship of continuum mechanics. Specifically, the displacement gradient is calculated based on methods such as interpolation or finite difference to obtain a strain estimate based on displacement. This estimate is then compared with the strain directly predicted by the neural network, and the difference between the two is incorporated into the loss function, thereby guiding the model to produce physically reasonable predictions.

[0179] The results output unit is responsible for visualizing and exporting the analysis results. This unit includes a displacement field visualization module, a strain field visualization module, a vector field visualization module, an equivalent strain calculation module, and a data export module. The displacement field visualization module displays the distribution of horizontal displacement (u) and vertical displacement (v); the strain field visualization module displays the distribution of normal strain (εxx, εyy) and shear strain (εxy); the vector field visualization module uses arrows to indicate the direction and magnitude of displacement; the equivalent strain calculation module calculates the equivalent strain index by integrating each strain component; and the data export module supports saving the analysis results in the form of images, numerical data, etc.

[0180] Example 15:

[0181] The method of using the AI-driven surface image deformation detection system includes the following specific steps:

[0182] Step 1: Load the reference image and the deformed image, and perform preprocessing.

[0183] First, read the grayscale images of the reference state and the deformed image, check if the two images are the same size, and then normalize the image grayscale values ​​to an appropriate range to create a normalized pixel coordinate grid system. For example... Figure 1 As shown, this step establishes the foundational data structure for subsequent analysis.

[0184] Step 2: Generate multi-scale Fourier features

[0185] like Figure 2 As shown, based on preset scale parameters (σ value) and the number of features at each scale, pixel coordinates are mapped to a high-dimensional feature space through random frequency matrix, coordinate projection, and trigonometric function transformation. For the standard implementation, two scales, σ = 1 and σ = M, can be selected, with 2M features generated at each scale, thereby capturing deformation features at different scales.

[0186] Step 3: Construct a neural network model

[0187] like Figure 3 As shown, a multi-branch parallel neural network model is constructed, with each branch processing Fourier features at a specific scale. For the standard implementation, a 32-node hidden layer and a 5-layer network depth can be used, employing a hyperbolic tangent activation function to ensure sufficient nonlinear expressive power. The outputs of all branches are merged in a fusion layer and then simultaneously predicted for both displacement and strain fields through the final output layer.

[0188] Step 4: Define the composite loss function for physical information constraints

[0189] like Figure 4 As shown, a comprehensive loss function is constructed by fusing image consistency constraints and physical compatibility constraints. Image consistency constraints ensure that the displacement field can accurately reconstruct the deformed image; physical compatibility constraints ensure that the displacement field and strain field satisfy the fundamental relationships of continuum mechanics. The relative importance of the two constraints is adjusted by weighting coefficients (λ_image, λ_physics).

[0190] Step 5: Iteratively train the model

[0191] The model is iteratively trained using an optimization algorithm at a preset learning rate. An adaptive learning rate adjustment strategy can be employed during training; the learning rate is reduced when the loss function stops decreasing, thereby improving optimization efficiency.

[0192] Step 6: Predict the displacement and strain fields

[0193] The trained model is used to simultaneously predict the displacement field (u, v) and strain field (εxx, εyy, εxy), and the equivalent strain is calculated based on the strain components.

[0194] Step 7: Visualize and save the results

[0195] The analysis results are displayed and saved in the form of images, numerical data, etc., including displacement field distribution maps, strain field distribution maps, displacement vector fields, strain curves, etc., to facilitate further analysis and research.

[0196] like Figure 5As shown, the artificial intelligence-based digital image correlation technology of the present invention can achieve accurate reconstruction of displacement fields and high-precision simulation of deformed images. Figure 5 (a) is a reference image. Figure 5 (b) is the actual deformed image. Figure 5 (c) is the deformed image reconstructed based on the predicted displacement field. Figure 5 (d) shows the difference map. It can be observed that the reconstructed image is highly consistent with the actual deformation image, and the error in the difference map is extremely small, indicating the accuracy of the displacement field prediction. The displacement and strain fields obtained by the method of this invention not only meet the requirements of image matching but also follow the basic principles of continuum mechanics, providing reliable data support for material deformation analysis.

[0197] The digital image correlation technology based on artificial intelligence of this invention achieves high-precision and high-reliability simultaneous measurement of displacement and strain fields by integrating Fourier feature encoding, neural network modeling, and physical information constraints. Compared with traditional sub-region-based correlation methods, this invention eliminates the need for manual setting of parameters such as sub-region size and automatically adapts to different deformation modes. Compared with existing intelligent DIC methods, this invention ensures the physical rationality of the measurement results through physical information constraints, thereby improving the reliability and accuracy of the measurement.

[0198] This invention is applicable to deformation measurement and analysis in fields such as materials science, mechanical engineering, and civil engineering. It is particularly suitable for handling complex situations such as large deformation and discontinuous deformation, and provides reliable technical support for applications such as material mechanical property evaluation and structural safety monitoring.

Claims

1. An artificial intelligence-driven digital image deformation detection system, characterized in that: It includes a reference image and deformed image processing unit, a Fourier feature encoding unit, an intelligent neural network model unit, a physical information constraint unit, and a result output unit; The reference image and deformed image processing unit reads the reference image and deformed image data of the surface of the object under test and performs preprocessing to generate a pixel coordinate grid of the reference image and deformed image. The Fourier feature coding unit maps the image coordinates in the pixel coordinate grid to a high-dimensional feature space to obtain the feature data of the reference image and the deformed image. The physical information constraint unit generates physical information constraints between the displacement field and the strain field; The intelligent neural network model unit constructs a multi-scale displacement field and strain field prediction model that satisfies physical information constraints. The intelligent neural network model unit inputs the spatial coordinates of the reference image and the deformed image into the multi-scale displacement field and strain field prediction model to obtain the displacement field and strain field analysis results of the object under test. The result output unit is used to visualize and export displacement field and strain field analysis results; The physical information constraints include image consistency constraints and physical compatibility constraints; The image consistency constraint predicts the displacement field through a neural network, then uses the displacement field to perform pixel-level remapping interpolation on the original reference image to generate a reconstructed deformed image. Then, this reconstructed image is compared pixel by pixel with the actual deformed image captured, and the difference between the two gray values ​​is calculated. This difference measure is used as a constraint term of the loss function. The physical compatibility constraint is measured by the relationship between displacement gradient and strain prediction value; The physical compatibility constraints are as follows: εxx = ∂u / ∂x(1) εyy = ∂v / ∂y(2) εxy = 0.5 * (∂u / ∂y + ∂v / ∂x) (3) Where u and v are the horizontal and vertical displacements, respectively; εxx and εyy are the normal strain components; εxy is the shear strain component; and (x, y) represents the coordinates.

2. The artificial intelligence-driven digital image deformation detection system according to claim 1, characterized in that: The reference image and deformed image processing unit includes an image reading module, an image normalization module, and a coordinate grid generation module; The image reading module is used to read reference image and deformed image data of the object's surface; the reference image of the object's surface is the image of the object before deformation; The image normalization module standardizes the grayscale values ​​of the reference image and the deformed image data. The coordinate grid generation module generates a regularized pixel coordinate grid based on the standardized grayscale values.

3. The artificial intelligence-driven digital image deformation detection system according to claim 1, characterized in that: The Fourier feature coding unit maps image coordinates in the pixel coordinate grid to a high-dimensional feature space through a multi-scale Fourier feature coding method. The steps include: random frequency matrix generation, coordinate projection transformation, feature scaling, and trigonometric function mapping.

4. The artificial intelligence-driven digital image deformation detection system according to claim 1, characterized in that: The intelligent neural network model unit is a multi-branch parallel structure, including multiple parallel processing branches and a fusion output layer; Each processing branch corresponds to a Fourier feature at a specific scale, consisting of a multi-layer fully connected network and a nonlinear activation function; the fusion output layer combines the outputs of each branch and outputs the displacement field and strain field predictions simultaneously.

5. The artificial intelligence-driven digital image deformation detection system according to claim 1, characterized in that: The result output unit includes a displacement field visualization module, a strain field visualization module, a vector field visualization module, an equivalent strain calculation module, and a data export module. The displacement field visualization module is used to display the horizontal and vertical displacement distribution; The strain field visualization module is used to display the distribution of normal strain and shear strain; The vector field visualization module is used to display the direction and magnitude of the displacement; The equivalent strain calculation module is used to calculate the comprehensive strain index equivalent_strain, that is: equivalent_strain = sqrt(2 / 3 (εxx 2 + εyy 2 + 2εxy 2 )) (4) Where εxx and εyy are normal strain components, εxy is shear strain component; sqrt is the square root function; The data export module is used to save the analysis results.

6. The artificial intelligence-driven digital image deformation detection system according to claim 1, characterized in that: The objects to be tested include metal specimens, concrete structures, and rock samples.

7. A method of using an artificial intelligence-driven surface image deformation detection system, applied to an artificial intelligence-driven digital image deformation detection system as described in any one of claims 1-6, characterized in that, Includes the following steps: Step 1: Load the reference image and the deformed image, perform preprocessing, and create a normalized coordinate grid; Step 2: Generate multi-scale Fourier features based on preset scale parameters, and map the coordinate information to a high-dimensional feature space; Step 3: Construct a multi-branch neural network model and set the number of layers and nodes; Step 4: Define a composite loss function that includes image consistency constraints and physical compatibility constraints; Step 5: Iterate through the training of the model and update the network parameters until the loss function converges; Step 6: Use the trained model to predict the displacement field and strain field, and calculate the equivalent strain; Step 7: Visualize the analysis results and save the displacement field and strain field data.

8. The method of use according to claim 7, characterized in that, The Fourier feature generation in step 2 includes: generating a random frequency matrix for each scale, projecting the normalized coordinates onto the frequency matrix, and applying sine and cosine functions for nonlinear mapping to obtain a high-dimensional feature representation.

9. The method of use according to claim 7, characterized in that, The physical compatibility constraint in step 4 is achieved by the following formula: εxx = u_x(5) εyy = v_y(6) εxy = 0.5 * (u_y + v_x) (7) Where u_x and u_y are the partial derivatives of the horizontal displacement u with respect to x and y, respectively; v_x and v_y are the partial derivatives of the vertical displacement v with respect to x and y, respectively; εxx and εyy are the normal strain components calculated by displacement gradient; and εxy is the shear strain component calculated by displacement gradient. The equivalent strain in step 6 is shown below: equivalent_strain = sqrt(2 / 3 (εxx 2 + εyy 2 + 2εxy 2 ))(8) Where εxx and εyy are normal strain components, and εxy is shear strain component.

Citation Information

Patent Citations

  • Prostate MR image multi-task registration method based on deep convolutional neural network

    CN111260705A

  • Fan blade damage rapid detection method based on inverse finite element reconstruction image recognition

    CN115575104A