Multi-level deep learning neural network adaptive global deformation vision measurement method and system
By performing global gridding on images using a multi-layered deep learning neural network, and combining the zero-mean normalized minimum distance squared standard and the Adam optimization algorithm, the problems of displacement field discontinuity and insufficient measurement accuracy in visual deformation measurement are solved, and high-precision adaptive global deformation measurement is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2023-07-21
- Publication Date
- 2026-05-12
AI Technical Summary
Existing visual deformation measurement methods have shortcomings in terms of displacement field continuity and measurement accuracy. In particular, local methods cannot guarantee the displacement continuity between subsets, while global methods based on finite element methods are difficult to adapt to non-uniformly deformed objects due to the manual setting of mesh node positions.
An adaptive global deformation visual measurement method using a multi-level deep learning neural network is adopted. The image is processed into a global grid by the multi-level deep learning neural network. The zero-mean normalized minimum distance squared standard is used as the loss function. Combined with the Adam optimization algorithm, the neural network parameters and node displacements are optimized to generate shape functions to achieve high-precision global deformation measurement.
It improves the accuracy and robustness of visual deformation measurement, can adapt to different materials and deformation characteristics, avoids errors caused by changes in exposure intensity, and achieves high-precision adaptive global deformation measurement.
Smart Images
Figure CN116972764B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and measurement technology, and more specifically, to a multi-level deep learning neural network adaptive global deformable visual measurement method and system. Background Technology
[0002] Visual deformation measurement is an image-based optical measurement technique used to measure the displacement and strain field of an object's surface. The core of visual deformation measurement is the image matching algorithm, which can be mainly divided into two types: subset-based local methods and finite element-based global methods.
[0003] Subset-based local methods are the earliest and most commonly used visual deformability measurement methods. They divide the Region of Interest (ROI) in an image into several subsets, and then perform separate image matching (also known as correlation operation) on each subset to determine the displacement of the center of each subset. The advantages of this method are its simplicity, intuitiveness, and flexibility, but the continuity of displacement between subsets cannot be guaranteed.
[0004] Patent document CN115345781A (application number: CN202210956950.6) discloses a multi-viewpoint video stitching method based on deep learning. The steps are as follows: First, using the Airsim simulator, images and depth data are acquired at a set virtual common viewpoint to generate a dataset for the video stitching task. Preprocessing is performed on the images, such as cylindrical projection. Then, a convolutional neural network is used to design an artifact elimination module and a smooth transition module. The former considers the feature correlation of overlapping areas and aligns the overlapping areas through viewpoint regression to eliminate artifacts after fusion. The latter, based on the image's feature information, propagates the obtained deformation rules of overlapping areas to non-overlapping areas to guide smooth transitions between regions and improve visual appeal. Finally, the original viewpoint image is distorted according to the predicted displacement field and weighted linear fusion is performed to obtain the stitching result.
[0005] The finite element method (FEM) is a visual deformation measurement method that has emerged in recent years. It discretizes the region of interest (ROI) in an image into several finite element units (FEMs), and then simultaneously searches and matches all units to determine the displacement of all nodes. The advantages of this method are that it can explicitly guarantee displacement continuity, reduce image registration errors, and improve measurement accuracy. However, it also has some disadvantages, such as the fact that the positions of the mesh nodes are manually set and cannot be changed, making it difficult to guarantee measurement accuracy and stability when measuring non-uniformly deformed objects. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a multi-level deep learning neural network adaptive global deformation visual measurement method and system.
[0007] The multi-layer deep learning neural network adaptive global deformable visual measurement method provided by the present invention includes:
[0008] Step S1: Align the optical axis of the camera and lens perpendicularly with the sample surface and record information about the sample under different conditions;
[0009] Step S2: Generate shape functions by globally meshing the region of interest in the image based on a multi-level deep learning neural network, thereby realizing global deformation measurement;
[0010] Step S3: Using the zero-mean normalized minimum squared distance standard as the loss function, the Adam algorithm is used to optimize the parameters and node displacements of the neural network. The weights and biases are functions of the node positions. Mesh adaptation is achieved by optimizing the neural network parameters.
[0011] Step S4: Export the global displacement field and strain field based on the optimization results to achieve high-precision adaptive global deformation measurement.
[0012] Preferably, in step S2, a global mesh generation shape function is performed on the region of interest in the image based on a multi-level deep learning neural network, and the displacement of all nodes is determined by simultaneously matching all cells, thereby ensuring the continuity of the displacement field;
[0013] The form of the function is:
[0014]
[0015] Among them, u e (x) represents the displacement field; n is the number of grid cell nodes; x represents the displacement field coordinates; x i Let u be the coordinates of grid node i; i For grid node displacement; For (x, x i The network structure of shape functions;
[0016] The shape function form of a multi-layered deep learning neural network is:
[0017]
[0018] Among them, W i b represents the weights of the intermediate layers of the neural network. i This is due to deviations in the intermediate layers of the neural network;
[0019] The sharing of nodes and boundaries between adjacent elements ensures that the displacement field between elements meets the consistency requirements. At the element boundaries and nodes, the displacements of adjacent elements are the same, thus guaranteeing the global continuity of the displacement field.
[0020] The weights W of the intermediate layers of a neural networki and deviation b i It is the grid node coordinate x i Functions:
[0021] [W i ,b i ]=H(x i )
[0022] By training the neural network, its parameters are continuously optimized [W] i ,b i The weights and biases are the grid node coordinates x. i The function, optimize [W] i ,b i The coordinates of the grid nodes are also optimized.
[0023] Preferably, in step S3, the zero-mean normalized minimum distance squared standard (ZNSSD) is selected as the loss function to avoid errors caused by changes in exposure intensity at different times. The expression is:
[0024]
[0025] In the formula, I represents the total number of pixels in the unit; f i and g i These represent the grayscale values of the i-th pixel in the reference image and the deformed image, respectively. and These represent the average gray values of all pixels in the reference image and the target image, respectively. and ZNSSD represents the standard deviation of all pixel gray values in the reference image and the target image, respectively; ZNSSD is 0 when the two images are perfectly matched.
[0026] For all finite element elements, their ZNSSD is used as the loss function, expressed as:
[0027]
[0028] Where m represents the number of finite element elements; C ZNSSDj Let ZNSSD represent the j-th unit.
[0029] Preferably, the multi-layered deep learning neural network consists of three basic modules: a linear function module, a multiplication module, and an inversion module. These three modules generate different types of shape functions to adapt to different materials and deformation characteristics. For a bilinear element shape function, the formula is:
[0030]
[0031]
[0032]
[0033]
[0034] in, This represents the x-coordinate of the i-th node in the e-th four-node unit; This represents the y-coordinate of the i-th node in the e-th four-node unit.
[0035] Preferably, the order is reduced by regular tensor decomposition, thereby improving the solution efficiency. The specific process is as follows:
[0036]
[0037] Where x, y, z are the coordinates of the displacement field with respect to the coordinate axes; Q is the order of the tensor decomposition; X q ,Y q Z q The function is decomposed according to the coordinate axes;
[0038] For the x-axis direction:
[0039]
[0040] Where nx represents the number of nodes along the x-axis, β q The decomposition coefficients;
[0041] The shape function form based on a multi-layered deep learning neural network is:
[0042]
[0043] The multi-layered deep learning neural network adaptive global deformable visual measurement system provided by the present invention includes:
[0044] Module M1: Aligns the optical axis of the camera and lens perpendicularly to the sample surface to record information about the sample under different conditions;
[0045] Module M2: Based on a multi-level deep learning neural network, shape functions are generated by global meshing of regions of interest in an image, thereby enabling global deformation measurement;
[0046] Module M3: Using the zero-mean normalized minimum squared distance standard as the loss function, Adam is used to optimize the parameters and node displacements of the neural network. The weights and biases are functions of the node positions. Mesh adaptation is achieved by optimizing the neural network parameters.
[0047] Module M4: Exports the global displacement field and strain field based on the optimization results, enabling high-precision adaptive global deformation measurement.
[0048] Preferably, in module M2, a multi-level deep learning neural network is used to generate shape functions for the region of interest in the image through global meshing. All node displacements are determined by simultaneously matching all elements, thereby ensuring the continuity of the displacement field.
[0049] The form of the function is:
[0050]
[0051] Among them, u e (x) represents the displacement field; n is the number of grid cell nodes; x represents the displacement field coordinates; x i Let u be the coordinates of grid node i; i For grid node displacement; For (x, x i The network structure of shape functions;
[0052] The shape function form of a multi-layered deep learning neural network is:
[0053]
[0054] Among them, W i b represents the weights of the intermediate layers of the neural network. i This is due to deviations in the intermediate layers of the neural network;
[0055] The sharing of nodes and boundaries between adjacent elements ensures that the displacement field between elements meets the consistency requirements. At the element boundaries and nodes, the displacements of adjacent elements are the same, thus guaranteeing the global continuity of the displacement field.
[0056] The weights W of the intermediate layers of a neural network i and deviation b i It is the grid node coordinate x i Functions:
[0057] [W i ,b i ]=H(x i )
[0058] By training the neural network, its parameters are continuously optimized [W] i ,b i The weights and biases are the grid node coordinates x. i The function, optimize [W] i ,b i The coordinates of the grid nodes are also optimized.
[0059] Preferably, in module M3, the zero-mean normalized minimum distance squared standard (ZNSSD) is selected as the loss function to avoid errors caused by changes in exposure intensity at different times. The expression is:
[0060]
[0061] In the formula, I represents the total number of pixels in the unit; f i and g i These represent the grayscale values of the i-th pixel in the reference image and the deformed image, respectively. and These represent the average gray values of all pixels in the reference image and the target image, respectively. and ZNSSD represents the standard deviation of all pixel gray values in the reference image and the target image, respectively; ZNSSD is 0 when the two images are perfectly matched.
[0062] For all finite element elements, their ZNSSD is used as the loss function, expressed as:
[0063]
[0064] Where m represents the number of finite element elements; Let ZNSSD represent the j-th unit.
[0065] Preferably, the multi-layered deep learning neural network consists of three basic modules: a linear function module, a multiplication module, and an inversion module. These three modules generate different types of shape functions to adapt to different materials and deformation characteristics. For a bilinear element shape function, the formula is:
[0066]
[0067]
[0068]
[0069]
[0070]
[0071]
[0072]
[0073]
[0074] in, This represents the x-coordinate of the i-th node in the e-th four-node unit; This represents the y-coordinate of the i-th node in the e-th four-node unit.
[0075] Preferably, the order is reduced by regular tensor decomposition, thereby improving the solution efficiency. The specific process is as follows:
[0076]
[0077] Where x, y, z are the coordinates of the displacement field with respect to the coordinate axes; Q is the order of the tensor decomposition; X q ,Y q Z q The function is decomposed according to the coordinate axes;
[0078] For the x-axis direction:
[0079]
[0080] Where nx represents the number of nodes along the x-axis, β q The decomposition coefficients;
[0081] The shape function form based on a multi-layered deep learning neural network is:
[0082]
[0083] Compared with the prior art, the present invention has the following beneficial effects:
[0084] (1) This invention proposes a multi-level deep learning neural network adaptive global deformation visual measurement method and system. By using HiDeNN to perform global gridding processing on the image, global deformation measurement can be achieved, which can overcome the problem of displacement field incompatibility in traditional methods and thus improve the accuracy of visual deformation measurement.
[0085] (2) This invention uses a hierarchical deep learning neural network for meshing. This neural network can achieve mesh adaptation, thereby improving the local and global accuracy of the interpolation function. At the same time, it can adapt to different materials and deformation characteristics, as well as different problems and needs.
[0086] (3) The present invention uses the zero-mean normalized cross-correlation standard as the loss function. This standard can avoid the error caused by the change of exposure intensity at different times and help improve the robustness of the algorithm. Attached Figure Description
[0087] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0088] Figure 1 This is a flowchart of the present invention;
[0089] Figure 2 The specific network structure for bilinear elemental shape functions;
[0090] Figure 3 A comparison diagram of traditional global vision measurement methods and this patent;
[0091] Figure 4 e calculated for this invention xx Strain field;
[0092] Figure 5 e calculated for this invention xy Strain field;
[0093] Figure 6 e calculated for this invention yy Strain field. Detailed Implementation
[0094] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0095] Example 1:
[0096] like Figure 1 This invention provides a multi-level deep learning neural network adaptive global deformable visual measurement method, comprising the following steps:
[0097] Step S1: In order to accurately record the information of the sample under different conditions, it is necessary to ensure that the optical axis of the camera and lens is perpendicularly aligned with the sample surface. This can avoid errors introduced by angular deviation.
[0098] Step S2: Generate shape functions by globally meshing the Region of Interest (ROI) in the image based on a multi-level deep-learning neural network (HiDeNN), thereby realizing global deformation measurement;
[0099] Step S2 specifically includes the following steps:
[0100] Step S2.1: Divide the region of interest (ROI) in the image into several four-node finite element elements. Each element consists of four nodes connected together. The initial position of the nodes can be represented by coordinate vectors.
[0101] Step S2.2: Select an appropriate shape function type, such as Lagrange polynomial, NURBS, isogeometric, regenerating nucleoparticle method, etc. The selection and combination should be based on different problems and requirements. Taking bilinear element shape functions as an example, the formula is:
[0102]
[0103]
[0104]
[0105]
[0106]
[0107]
[0108]
[0109]
[0110] Let x represent the x-coordinate of the i-th node in the e-th four-node element. This represents the y-coordinate of the i-th node in the e-th four-node unit.
[0111] Step S2.3: Construct a hierarchical deep learning neural network (HiDeNN). This neural network consists of three basic modules: a linear function module, a multiplication module, and an inversion module. These three modules can generate different types of interpolation functions. Hierarchical deep learning neural network structures such as Figure 2 As shown, the network structures for other shape functions follow the same principle. Represents the linear rectified unit (ReLu).
[0112] Step S2.4: Treat the node positions and node displacements as variables to be solved in the optimization problem, thereby achieving image matching and mesh deformation.
[0113] Step S3: Using the Zero-Normalized Sum of Squared Differences Criterion (ZNSSD) as the loss function, adaptive moment estimation (Adam) is used to optimize the parameters and node displacements of the neural network. The weights and biases are functions of the node positions. Optimization of the neural network parameters enables mesh adaptation, resulting in measurement accuracy higher than traditional finite element-based global vision measurement methods. Figure 3 As shown;
[0114] Specifically, step S3 includes the following steps:
[0115] Step S3.1: For each finite element, calculate its gray value in the reference image and the target image, and then calculate the zero-mean normalized minimum squared distance standard (ZNSSD) between the two. This function is insensitive to overall changes in light intensity and has good robustness.
[0116]
[0117] In the formula, I represents the total number of pixels in the unit, and f i and g i These represent the grayscale values of the i-th pixel in the reference image and the deformed image, respectively. and These represent the average gray values of all pixels in the reference image and the target image, respectively. and ZNSSD represents the standard deviation of all pixels in the reference image and the target image, respectively; ZNSSD is 0 when the two images are perfectly matched.
[0118] Step S3.2: For all finite element elements, use their ZNSSD as the loss function, i.e.:
[0119]
[0120] Where m represents the number of finite element elements, Let ZNSSD represent the j-th unit.
[0121] Step S3.3: Use the Adaptive Moment Estimation (Adam) algorithm to optimize the parameters and node displacements of the neural network. This algorithm is an adaptive optimization algorithm that can adjust the learning rate according to the first and second moments of the gradient, thereby improving the convergence speed and stability.
[0122] Step S3.4: Repeat steps S3.1 to S3.3 until the iteration stopping criterion is met or the maximum number of iterations is reached, thereby obtaining the optimized neural network parameters and node displacements.
[0123] Step S4: Based on the optimization results, the global displacement field and strain field can be exported to achieve high-precision adaptive global deformation measurement.
[0124] This invention provides a multi-level deep learning neural network adaptive global deformable vision measurement system, comprising:
[0125] Module M1: Aligns the optical axis of the camera and lens perpendicularly to the sample surface to record information about the sample under different conditions;
[0126] Module M2: Based on a multi-level deep learning neural network, shape functions are generated by global meshing of regions of interest in an image, thereby enabling global deformation measurement;
[0127] Module M3: Using the zero-mean normalized minimum squared distance standard as the loss function, Adam is used to optimize the parameters and node displacements of the neural network. The weights and biases are functions of the node positions. Mesh adaptation is achieved by optimizing the neural network parameters.
[0128] Module M4: Exports the global displacement field and strain field based on the optimization results, enabling high-precision adaptive global deformation measurement.
[0129] In module M2, a multi-level deep learning neural network is used to generate shape functions for the region of interest in the image through global meshing. All node displacements are determined by matching all cells simultaneously, thereby ensuring the continuity of the displacement field.
[0130] The form of the function is:
[0131]
[0132] Among them, u e (x) represents the displacement field; n is the number of grid cell nodes; x represents the displacement field coordinates; x i Let u be the coordinates of grid node i; i For grid node displacement; For (x, x i The network structure of shape functions;
[0133] The shape function form of a multi-layered deep learning neural network is:
[0134]
[0135] Among them, W i b represents the weights of the intermediate layers of the neural network. i This is due to deviations in the intermediate layers of the neural network;
[0136] The sharing of nodes and boundaries between adjacent elements ensures that the displacement field between elements meets the consistency requirements. At the element boundaries and nodes, the displacements of adjacent elements are the same, thus guaranteeing the global continuity of the displacement field.
[0137] The weights W of the intermediate layers of a neural network i and deviation b i It is the grid node coordinate x i Functions:
[0138] [W i ,b i ]=H(x i )
[0139] By training the neural network, its parameters are continuously optimized [W] i ,b i The weights and biases are the grid node coordinates x. i The function, optimize [W] i ,b i The coordinates of the grid nodes are also optimized.
[0140] In module M3, the zero-mean normalized minimum distance squared standard (ZNSSD) is selected as the loss function to avoid errors caused by changes in exposure intensity at different times. The expression is:
[0141]
[0142] In the formula, I represents the total number of pixels in the unit; f i and g i These represent the grayscale values of the i-th pixel in the reference image and the deformed image, respectively. and These represent the average gray values of all pixels in the reference image and the target image, respectively. and ZNSSD represents the standard deviation of all pixel gray values in the reference image and the target image, respectively; ZNSSD is 0 when the two images are perfectly matched.
[0143] For all finite element elements, their ZNSSD is used as the loss function, expressed as:
[0144]
[0145] Where m represents the number of finite element elements; Let ZNSSD represent the j-th unit.
[0146] The multi-layered deep learning neural network consists of three basic modules: a linear function module, a multiplication module, and an inversion module. These three modules generate different types of shape functions to adapt to different materials and deformation properties. For a bilinear element shape function, the formula is:
[0147]
[0148]
[0149]
[0150]
[0151]
[0152]
[0153]
[0154]
[0155] in, This represents the x-coordinate of the i-th node in the e-th four-node unit; This represents the y-coordinate of the i-th node in the e-th four-node unit.
[0156] The order is reduced by using regular tensor decomposition, thereby improving the solution efficiency. The specific process is as follows:
[0157]
[0158] Where x, y, z are the coordinates of the displacement field with respect to the coordinate axes; Q is the order of the tensor decomposition; X q ,Y q Z q The function is decomposed according to the coordinate axes;
[0159] For the x-axis direction:
[0160]
[0161] Where nx represents the number of nodes along the x-axis, β q The decomposition coefficients;
[0162] The shape function form based on a multi-layered deep learning neural network is:
[0163]
[0164] Example 2:
[0165] To verify the effectiveness of this invention, the 2-D DIC Challenge 1.0 dataset provided by the International Digital Image Correlation Society (iDICs) was used for testing. This dataset is intentionally designed to have different exposure intensities between each image to test the robustness of the algorithm. A Cartesian coordinate system was established with the bottom left corner of the image as the origin, the top edge as the x-axis, and the left edge as the y-axis. The displacement field was directly solved using Adam, and the strain field was derived using the following formula:
[0166]
[0167] u represents the displacement along the x-axis, v represents the displacement along the y-axis, and e xy This represents the change in the angle between two small line segments in mutually perpendicular directions after deformation; e yyThis represents the ratio of the length increment of a small line segment along the y-axis due to deformation to its original length; it is positive for elongation. Since this is a dataset of rigid body translational images, the closer the calculated strains are to 0, the higher the measurement accuracy. The measurement results of this invention are as follows: Figures 4 to 6 As shown, the error magnitude reaches 10. -4 This demonstrates the high precision and robustness of the present invention.
[0168] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.
[0169] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A multi-level deep learning neural network adaptive global deformable visual measurement method, characterized in that, include: Step S1: Align the optical axis of the camera and lens perpendicularly with the sample surface and record information about the sample under different conditions; Step S2: Generate shape functions by globally meshing the region of interest in the image based on a multi-level deep learning neural network, thereby realizing global deformation measurement; Step S3: Using the zero-mean normalized minimum squared distance standard as the loss function, the Adam algorithm is used to optimize the parameters and node displacements of the neural network. The weights and biases are functions of the node positions. Mesh adaptation is achieved by optimizing the neural network parameters. Step S4: Export the global displacement field and strain field based on the optimization results to achieve high-precision adaptive global deformation measurement; In step S2, a global mesh generation shape function is performed on the region of interest in the image based on a multi-level deep learning neural network. The displacement of all nodes is determined by matching all cells at the same time, thereby ensuring the continuity of the displacement field. The form of the function is: in, For displacement field; This represents the number of nodes in the grid cell. These are the displacement field coordinates; Let i be the coordinates of grid node i; For grid node displacement; for The network structure of shape functions; The shape function form of a multi-layered deep learning neural network is: in, These are the weights of the intermediate layers of the neural network; This represents the bias in the intermediate layers of the neural network. The sharing of nodes and boundaries between adjacent elements ensures that the displacement field between elements meets the consistency requirements. At the element boundaries and nodes, the displacements of adjacent elements are the same, thus guaranteeing the global continuity of the displacement field. Weights of the intermediate layers of a neural network and deviation Grid node coordinates Functions: Parameters are continuously optimized by training the neural network. The weights and biases are the coordinates of the grid nodes. The function, optimization The coordinates of the time grid nodes are also optimized; The multi-layered deep learning neural network consists of three basic modules: a linear function module, a multiplication module, and an inversion module. These three modules generate different types of shape functions to adapt to different materials and deformation properties. For a bilinear element shape function, the formula is: in, Indicates the first The first four-node unit Nodes coordinate; Indicates the first The first four-node unit Nodes coordinate.
2. The multi-level deep learning neural network adaptive global deformable visual measurement method according to claim 1, characterized in that, In step S3, the zero-mean normalized minimum distance squared standard (ZNSSD) is selected as the loss function to avoid errors caused by changes in exposure intensity at different times. The expression is: In the formula, This represents the total number of pixels in the cell. and These represent the grayscale values of the i-th pixel in the reference image and the deformed image, respectively. and , representing the average grayscale value of all pixels in the reference image and the target image, respectively; and ZNSSD represents the standard deviation of all pixel gray values in the reference image and the target image, respectively; ZNSSD is 0 when the two images are perfectly matched. For all finite element elements, their ZNSSD is used as the loss function, expressed as: in, Indicates the number of finite element elements; Indicates the first ZNSSD for each unit.
3. The multi-level deep learning neural network adaptive global deformable visual measurement method according to claim 1, characterized in that, The order is reduced by using regular tensor decomposition, thereby improving the solution efficiency. The specific process is as follows: in, These are the coordinates of the corresponding coordinate axes of the displacement field; Let be the order of the tensor decomposition; The function is decomposed according to the coordinate axes; For the x-axis direction: in, This represents the number of nodes along the x-axis. The decomposition coefficients; The shape function form based on a multi-layered deep learning neural network is: 。 4. A multi-level deep learning neural network adaptive global deformable visual measurement system, characterized in that, include: Module M1: Aligns the optical axis of the camera and lens perpendicularly to the sample surface to record information about the sample under different conditions; Module M2: Based on a multi-level deep learning neural network, shape functions are generated by global meshing of regions of interest in an image, thereby enabling global deformation measurement; Module M3: Using the zero-mean normalized minimum squared distance standard as the loss function, Adam is used to optimize the parameters and node displacements of the neural network. The weights and biases are functions of the node positions. Mesh adaptation is achieved by optimizing the neural network parameters. Module M4: Exports the global displacement and strain fields based on the optimization results, enabling high-precision adaptive global deformation measurement; In module M2, a multi-level deep learning neural network is used to generate shape functions for the region of interest in the image through global meshing. All node displacements are determined by matching all cells simultaneously, thereby ensuring the continuity of the displacement field. The form of the function is: in, For displacement field; This represents the number of nodes in the grid cell. These are the displacement field coordinates; Let i be the coordinates of grid node i; For grid node displacement; for The network structure of shape functions; The shape function form of a multi-layered deep learning neural network is: in, These are the weights of the intermediate layers of the neural network; This represents the bias in the intermediate layers of the neural network. The sharing of nodes and boundaries between adjacent elements ensures that the displacement field between elements meets the consistency requirements. At the element boundaries and nodes, the displacements of adjacent elements are the same, thus guaranteeing the global continuity of the displacement field. Weights of the intermediate layers of a neural network and deviation Grid node coordinates Functions: Parameters are continuously optimized by training the neural network. The weights and biases are the coordinates of the grid nodes. The function, optimization The coordinates of the time grid nodes are also optimized; The multi-layered deep learning neural network consists of three basic modules: a linear function module, a multiplication module, and an inversion module. These three modules generate different types of shape functions to adapt to different materials and deformation properties. For a bilinear element shape function, the formula is: in, Indicates the first The first four-node unit Nodes coordinate; Indicates the first The first four-node unit Nodes coordinate.
5. The multi-level deep learning neural network adaptive global deformable visual measurement system according to claim 4, characterized in that, In module M3, the zero-mean normalized minimum distance squared standard (ZNSSD) is selected as the loss function to avoid errors caused by changes in exposure intensity at different times. The expression is: In the formula, This represents the total number of pixels in the cell. and These represent the grayscale values of the i-th pixel in the reference image and the deformed image, respectively. and , representing the average grayscale value of all pixels in the reference image and the target image, respectively; and ZNSSD represents the standard deviation of all pixel gray values in the reference image and the target image, respectively; ZNSSD is 0 when the two images are perfectly matched. For all finite element elements, their ZNSSD is used as the loss function, expressed as: in, Indicates the number of finite element elements; Indicates the first ZNSSD for each unit.
6. The multi-level deep learning neural network adaptive global deformable visual measurement system according to claim 4, characterized in that, The order is reduced by using regular tensor decomposition, thereby improving the solution efficiency. The specific process is as follows: in, These are the coordinates of the corresponding coordinate axes of the displacement field; Let be the order of the tensor decomposition; The function is decomposed according to the coordinate axes; For the x-axis direction: in, This represents the number of nodes along the x-axis. The decomposition coefficients; The shape function form based on a multi-layered deep learning neural network is: 。