A Deep Learning-Based Method and System for Measuring Large Deformations

Abstract

This invention discloses a deep learning-based method and system for measuring large deformations. It extracts features and deformation information from a reference image and a deformed image to estimate initial values. Then, it uses a recursive iterative approach to refine the initial displacement values, constructing a measurement model. Finally, the obtained measurement model is trained and optimized to obtain the final model. Inputting the image to be measured into the trained measurement model yields the deformation field between images. Compared to existing deep learning digital image correlation measurement methods, this method can measure both small and large deformations. By estimating initial values ​​on low-scale features and then using the estimated initial values ​​for iterative refinement on high-scale features, it ensures the model can adapt to both small and large deformation measurements. Compared to traditional methods, it overcomes the limitations of slow computation speed and the need for manual parameter setting in digital image correlation methods. This method no longer relies on sub-region size settings and has a significant advantage in measuring high-frequency changing deformation fields.

Description

Technical Field

[0001] This invention belongs to the field of digital image-related technology, and in particular relates to a method and system for measuring large deformations based on deep learning. Background Technology

[0002] With the vigorous development of society and the continuous progress of science and technology, measurement technology is also developing rapidly. Modern measurement methods are widely used in the research of various materials, component structures and product performance. They play a huge feedback and correction role in production and research in basic manufacturing, automotive industry, aerospace and defense, and have a pivotal position in industrial development. Displacement and strain, the two measured quantities, can usually directly reflect the mechanical properties of materials and structures, thus providing theoretical support for material selection. The measurement methods are generally divided into two types: contact measurement and non-contact measurement. Contact measurement mainly includes displacement gauges, strain gauges, etc. This measurement method has high accuracy and good stability, but because it requires contact with the measured object, it will affect the accuracy of the measurement results to a certain extent. Non-contact measurement mainly includes photoelasticity, fringe projection, moiré interferometry and digital image correlation methods. Among them, digital image correlation method stands out from many methods with its special advantages such as simple measurement equipment and low environmental requirements, and is currently the most active measurement method in the field of optical mechanics research.

[0003] Digital image correlation (DIC) is a non-contact optical metrology technique for measuring motion and deformation across the entire field. This technique requires only a camera as the image acquisition device to measure deformation. Over the past thirty years, DIC has been increasingly applied in many fields involving non-contact measurement, such as experimental mechanics, cellular mechanics, and structural health monitoring. The core of traditional DIC methods is to use images of the object before and after deformation as the reference image and the deformed image, respectively. Several equidistant sub-regions are selected on the reference image, and the center of each sub-region is tracked by matching the sub-region with the highest correlation on the deformed image. These sub-region centers are also called seed points. The surface deformation of the object can be measured by tracking the position of these seed points. However, to ensure high computational efficiency, a relatively small number of seed points are usually set to describe the surface deformation. After calculating the displacement of the seed points, the displacement of the entire field is obtained through interpolation, which can lead to significant interpolation errors. Therefore, the number of seed points must be balanced between computational efficiency and interpolation accuracy. Clearly, designing a DIC algorithm with high computational efficiency and a sufficient number of seed points is of great significance.

[0004] The theoretical research on the DIC method mainly focuses on integer-pixel search and subpixel registration. Integer-pixel search methods include Fourier transform, SIFT, and genetic algorithms. Regarding subpixel registration, since the introduction of the Newton-Raphson partial differential iteration method in 1989, it has been a major breakthrough in the development of the DIC method. However, when the deformed image is updated, the Newton-Raphson method needs to recalculate the Hessian matrix, which greatly limits its computational efficiency. However, the overall accuracy of subpixel registration is very high. Subsequently, the inverse compositional Gauss-Newton iteration method, also known as the IC-GN (Inverse Compositional Gauss-Newton algorithm), was proposed for image registration. It avoids the repeated calculation of the Hessian matrix and has become the most efficient subpixel matching method. To date, the IC-GN algorithm has become the mainstream method for subpixel registration in current DIC calculations.

[0005] The computational efficiency of the mainstream methods mentioned above is difficult to meet the requirements of real-time measurement, especially for images with high resolution. It cannot satisfy the requirement of obtaining the displacement of every pixel in the entire field. Moreover, this method requires manual determination of the sub-region size and iteration parameters based on the type of deformation. This relies on human experience, and requires a longer iteration time for some complex deformations. It may even fail for some high-gradient deformations. Recently, some scholars have applied deep learning to DIC measurement to obtain full-field displacement measurement. This method can greatly improve computational efficiency. However, their network is just a simple modification of U-net. The network can only perform small displacement measurements, has a large number of parameters, and has poor generalization ability.

[0006] In recent years, the rapid development of high-speed and high-resolution digital cameras has led to a surge in images and a surge in the number of pixels that need to be processed in each image, making the demand for high-precision, high-speed DIC increasingly urgent. In addition, for the special requirements of real-time motion tracking in some applications, the computational efficiency of DIC methods has become increasingly important in recent years. It is of great significance to study a DIC method that is fast, robust, and can overcome the shortcomings of traditional methods. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for measuring large deformations based on deep learning, so as to overcome the problems that existing deep learning digital image correlation methods can only measure small displacements, have a large number of parameters, low computational efficiency, and poor generalization ability.

[0008] A deep learning-based method for measuring large deformations includes the following steps:

[0009] S1. Use a Boolean model to render a reference image, apply a deformation field to the reference image to obtain a deformation image, and construct a dataset using the reference image, the deformation image and the corresponding deformation field.

[0010] S2, Feature Extraction Stage: Extracting feature and deformation information from the images in the S1 dataset;

[0011] Initial value estimation stage: The initial displacement value is estimated based on the feature information and deformation information of the image;

[0012] Iterative refinement stage: The initial displacement value is iteratively refined to obtain the final displacement, thus completing the construction of the measurement model;

[0013] S3, train and optimize the measurement model to obtain the final measurement model;

[0014] S4. Acquire images and perform grayscale processing on the acquired images. For a series of frame images, the first frame image is used as the reference image, and each subsequent frame image is used as the deformation image. Inputting the reference image and the deformation image into the trained model will yield the deformation field between the two frames.

[0015] Furthermore, in S1, the displacement is randomly generated, with the specified pixel coordinate system as the reference. The corresponding pixel of the reference frame is retrieved on the deformed frame, and it is determined whether the updated index coordinates are out of bounds. If an out-of-bounds error occurs, the displacement is regenerated.

[0016] Furthermore, in the feature extraction stage, feature information at different scales of the image is extracted through a feature encoder. The feature information of the reference image is: The characteristic information of the deformation diagram is Context encoders extract deformation information at different scales from images.

[0017] Furthermore, in the initial value estimation stage, the obtained displacement features are used to estimate the deformation characteristics. Feature distortion processing is performed to obtain By utilizing the feature information of the reference image and the feature information of the distorted image, a corresponding cost volume is constructed, thereby obtaining the motion information expression:

[0018]

[0019] Where Deformation is the amount of deformation, D k-1 It is the deformation field calculated from the previous layer, C k Represents the relevant cost body, These are the features of the distorted image. F describes the difference between the distorted feature and the reference feature. m ,F c,F φ ,F δ F1 and F2 represent linear mappings performed using convolutional layers;

[0020] The obtained deformation and image deformation information The update module is sent to the device, and the update is completed via GRU.

[0021] Furthermore, the GRU formula is:

[0022] z t =σ(Conv 3×3 ([h t-1 ,x t ],W z ))

[0023] r t =σ(Conv 3×3 ([h t-1 ,x t ],W r ))

[0024]

[0025]

[0026] in, and For image deformation information The two parts split on the C channel, Conv 3×3 This represents a 3×3 convolution, [] denotes concatenation, σ and tanh denote activation functions, and W... z W r W h For different weight parameters;

[0027] h is finally obtained through the GRU formula. t , making D k =Conv(h t W k Mask = Conv(h) t W m )

[0028] Where Conv represents a convolutional layer, W k With W m It is the weight parameter, D k The variable represents the deformation field, and Mask represents the mask information used to upsample the deformation field.

[0029] Furthermore, the iterative refinement stage is the same as the initial value estimation stage, used to estimate the residual displacement Δd. The displacement estimate obtained in each iteration is: dk+1 =Δd+d k From this, a series of displacement estimates can be obtained.

[0030] Furthermore, the series of displacement estimates are all at 1 / 4 resolution of the original image, and are upsampled by a factor of 4 using masking information to output a series of displacements d. i .

[0031] Furthermore, the measurement model was implemented using the PyTorch framework and trained on an RTX3090. AdamW was chosen as the optimization method, and the loss was defined as:

[0032]

[0033] Where γ = 0.8, N represents the number of training rounds, and d gt Given the true displacement, d i This represents a series of displacements output by the model.

[0034] Furthermore, the training optimization results are evaluated using the mean endpoint error. During the training process, if the current result achieves the best mean endpoint error (EPE) on the test set, then the training parameters for that set are determined, and the final measurement model is determined.

[0035] Furthermore, the dataset generation module is used to generate a series of reference diagrams and deformation diagrams, as well as the corresponding deformation fields;

[0036] The model building module is used to extract information from the image generation module, complete the initial displacement value estimation, and perform iterative updates and refinement.

[0037] The model training and optimization module trains and optimizes the model obtained from the model building module to determine the optimal training parameters and thus the final model.

[0038] The measurement module is used to acquire images and perform grayscale processing. The images are then input into a trained measurement model to obtain the deformation field between the images.

[0039] Compared with the prior art, the present invention has the following beneficial technical effects:

[0040] This invention discloses a deep learning-based method for measuring large deformations. It extracts features and deformation information from a reference image and the deformed image to estimate initial values, then uses recursive iteration to refine the initial displacement values, thus constructing a measurement model. Finally, the obtained measurement model is trained and optimized to obtain the final model. Compared to existing deep learning digital image methods, this method can measure both small and large deformations. By estimating initial values ​​on low-scale features and then using the estimated initial values ​​for iterative refinement on high-scale features, the model can adapt to both small and large deformation measurements. During the iterative refinement stage, model parameters are shared in each iteration, reducing model parameters and lowering training costs without affecting the model's generalization ability. This application also addresses the slow computation speed and manual parameter setting issues of traditional methods. Compared to traditional methods, this method no longer relies on the setting of sub-region size, giving it an absolute advantage in measuring high-frequency changing displacement fields.

[0041] Preferably, during the stage of creating the reference image and the deformed image, for large deformations, when the displacement is large, pixels will be moved in and moved out. Using a specified pixel coordinate system as a reference, the corresponding pixels of the deformed frame are retrieved on the reference frame, and it is determined whether the updated index coordinates are out of bounds. If they are out of bounds, the displacement is regenerated.

[0042] Preferably, in the initial value estimation stage, feature distortion is performed on the feature information of the deformation map. Feature distortion reduces the feature space distance between the feature information of the reference map and the deformation map, so as to facilitate the processing of large displacement prediction. Attached Figure Description

[0043] Figure 1 This is a flowchart of a deep learning-based large deformation measurement method in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of deformation map generation using a deep learning-based large deformation measurement method in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram illustrating the model construction of a large deformation measurement method based on deep learning in an embodiment of the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0047] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0048] This invention discloses a deep learning-based method for measuring large deformations, comprising the following steps:

[0049] S1. Create a speckle dataset, which involves rendering a reference image using a Boolean model, applying a displacement (deformation field) to the reference image to obtain a deformed image, and constructing the dataset from the reference image, the deformed image, and the corresponding deformation field.

[0050] Unlike traditional methods, deep learning methods require labeled data with real displacements. Generally, obtaining such real labels is difficult because the displacements corresponding to pixels in real-world scenes are not easy to determine. The dataset should consist of paired blob images of reference and deformation, as well as their associated deformation fields. It must represent real images and deformations to ensure strong generalization ability.

[0051] Specifically, 200 reference frames of 256×256 pixels were first rendered using a Boolean model. To generate a rich set of displacement fields covering as wide a range of displacement field types as possible, grid regions with different intervals were first generated. Then, the displacements (deformation fields) of these nodes were randomly generated. These displacements (deformation fields) were treated as the actual displacements and then linear interpolation was used to transform them into 256×256 pixels. The applied random displacements were between [-30, 30] pixels. In addition, for large deformations, when the displacement amount is large, pixels will shift in and out, making it impossible to find the correspondence between the deformed frame and the reference frame pixels. This will increase the number of incorrect matches. However, since the displacement amount is between [-30, 30], only the boundary 30 pixels will have the problem of shifting in and out. The specific solution is as follows:

[0052] First, a pixel coordinate system (pixel index starting from 0) is specified as the reference. Rightward movement is positive and leftward movement is negative along the X-direction, while upward movement is negative and downward movement is positive along the Y-direction. Since the corresponding pixels in the reference frame are retrieved on the deformed frame, it is necessary to ensure that all pixels in the deformed frame originate from the reference frame. Therefore, it is only necessary to check whether the updated index coordinates are out of bounds. Whenever an out-of-bounds situation occurs, the displacement is regenerated. To reduce generation time, for the X-direction, the displacement of the left boundary is generated between [0, 30], and the displacement of the right boundary is generated between [-30, 0]. The same principle applies to the Y-direction. After generating the displacement, it is applied to the reference image to obtain the corresponding deformed image.

[0053] S2: Model building;

[0054] The structure of the model is mainly as follows: Figure 3 As shown, the model construction introduces the idea of ​​recursive residual refinement, which is mainly divided into three stages: ① feature extraction, ② initial value estimation, and ③ iterative update refinement. These three stages form an end-to-end trainable architecture.

[0055] The input image requires selecting a region of interest (ROI) for calculation. After selection, the image is converted to grayscale and normalized before being input into the network. The first stage completes feature extraction, mainly extracting image information and deformation information. The second stage estimates the initial displacement value based on these two parts of information. The third part iteratively refines the initial displacement value estimated in the previous stage. The number of iterations is set externally by the program. The program can output the displacement in both the X and Y directions simultaneously in one run.

[0056] Feature extraction stage

[0057] The feature extraction stage mainly consists of a feature encoder and a context encoder. The feature encoder and context encoder have identical structures, both employing a pyramid structure, specifically implemented using residual blocks. The reference image is defined as I1, and the deformed image as I2. The feature encoder is used to extract feature information at different scales of the image. Context encoders are used to extract deformation information at different scales of an image. Here, k represents different scales, each of which acts as a feature descriptor. Together, these two provide preparation for subsequent displacement estimation, and the feature network and context network only need to be executed once.

[0058] Initial value estimation stage

[0059] The initial value estimation module mainly includes a motion information fusion and update module, which incorporates features of different scales obtained from the reference image and the deformed image in the previous stage. The displacement pairs that have already been obtained will be used in subsequent operations. By performing feature distortion Reduced by feature distortion and The feature space distance between them is used to facilitate large displacement prediction. When the initial displacement is 0, distortion is not required; the distorted features and Using both, the corresponding cost volume is constructed, and the specific expression for motion information fusion is as follows:

[0060]

[0061] Where Deformation is the amount of deformation, D k-1 It is the deformation field calculated from the previous layer, C k Represents the relevant cost body, These are the features of the distorted image. F describes the difference between the distorted feature and the reference feature. m ,F c ,F φ ,F δ F1 and F2 represent linear mappings performed using convolutional layers;

[0062] The obtained motion feature deformation will be compared with the context features The data is sent to the update module and updated by the GRU. GRU (Gated Recurrent Unit) is a variant of recurrent neural network (RNN) that is widely used in sequence data modeling. GRU uses a gating mechanism to regulate the flow of information to help the network better capture long-term dependencies.

[0063] The formula for GRU can be expressed as:

[0064] Update Gate:

[0065] z t =σ(Conv 3×3 ([h t-1 ,x t ],W z ))

[0066] Reset Gate:

[0067] r t =σ(Conv 3×3 ([h t-1 ,x t ],W r ))

[0068] Candidate Hidden State:

[0069]

[0070] Final Hidden State:

[0071]

[0072] in, The C channel is divided into two parts: and make Conv 3×3 This represents a 3×3 convolution, where σ is the Sigmoid function; ⊙ represents element-wise multiplication (Hadamard product); [] represents concatenation; tanh represents the activation function; and W... z W r W h For different weight parameters;

[0073] GRU controls the influence of the previous hidden state and the current input information on the next time step through the update gate, while the reset gate determines which information from the previous hidden state needs to be forgotten. In other words, GRU is used to decide which information to discard and which new information to update. The updated feature volume predicts displacement and mask information through convolution, where the mask information is mainly used for upsampling. The displacement estimated by the initial value is 1 / 4 of the original image resolution.

[0074] h is finally obtained through the GRU formula. t , making D k =Conv(h t W k Mask = Conv(h) t W m Where Conv represents a convolutional layer, W k With W m It is the weight parameter, D k The variable represents the deformation field, and Mask represents the mask information used to upsample the deformation field.

[0075] Iterative refinement stage

[0076] This module primarily receives the initial value d0 from the previous stage. d0 represents the displacement of the deformation field obtained in the initial value estimation stage after upsampling with mask information, and the deformation information. With feature information Its module structure is shown in the figure. Similar to the second stage, it mainly consists of a motion information fusion and update module. This module executes n times to estimate the residual displacement Δd. The displacement estimate obtained each time is: d k+1 =Δd+d kA series of displacement estimates {d1,…,d} are obtained. n Since the displacement estimate obtained each time is only 1 / 4 of the resolution of the original image, it is upsampled by 4 times using mask information to output a series of displacements for subsequent training.

[0077] S3: Train and optimize the measurement model to obtain the final measurement model;

[0078] The model was implemented using the PyTorch framework and trained on an RTX 3090. AdamW was chosen as the optimization method, and the weight decay was set to 0.5 × 10⁻⁶. -4 The batch size is set to 32, the learning rate starts at 0.002, the loss function is L1 loss, and the predicted sequence is {d1,…,d}. N Given the true displacement d gt Loss is defined as

[0079]

[0080] In the experiment, γ was set to 0.8, N represents the number of training rounds, and the number of training rounds was set to 300. i This represents a series of displacements output by the model, where the actual displacement d is... gt This refers to the displacement (deformation field) applied to the reference image in S1. During training, the above training strategy is used as a benchmark, and fine-tuning can be performed according to different devices. The average endpoint error (EPE) is used for evaluation. During training, if the current result reaches the best average endpoint error (EPE) on the test set, this set of training parameters is saved, and then the final measurement model is determined.

[0081] Assume there is an image of size W×H, where W represents the width of the image and H represents the height of the image.

[0082] For each pixel position (x, y), the true displacement vector and estimate the displacement vector (u) i,j ,v i,j Average endpoint error:

[0083]

[0084] In this formula, we sum the values ​​at each pixel position in the image and divide the final result by the total number of pixels W×H in the image to obtain the average endpoint error (EPE). This calculation method can measure the accuracy of displacement estimation on a two-dimensional image. The error is calculated for each pixel position and the average value is obtained. A smaller average endpoint error (EPE) value indicates that the displacement estimation result is closer to the true displacement, indicating that the algorithm is more accurate.

[0085] S4: Image Deformation Measurement

[0086] The experimental setup used in the measurement experiment was relatively simple. The imaging software used was Pylon Viewer, the camera was an ACA2440-75UC, and the lens was an ML-M0822UR. One end of the sample was fixed, and the other end was pulled at a constant speed of 10 mm / s until it broke. A fixed acquisition frequency was set, the acquired images were processed into grayscale, and then the first frame was used as the reference frame and subsequent frames were used as deformation frames. These were input into the trained measurement model, and the deformation field (displacement in the X and Y directions) could be calculated through the model to complete the displacement measurement.

[0087] A deep learning-based large deformation measurement system includes a dataset generation module, a model building module, a model training and optimization module, and a measurement module. The image generation module generates a reference image, a deformation image, and corresponding displacements (deformation fields), and handles pixel shifting in and out when the displacement is large. The model building module extracts feature and deformation information from the reference and deformation images, estimates the initial displacement value, and iteratively refines it. The model training and optimization module trains and optimizes the model obtained from the model building module, defines a loss function, evaluates it using the average endpoint error, finds the best set of training parameters, and obtains the final measurement model. The measurement module acquires images and performs grayscale processing on the acquired images. The reference and deformation images to be measured are input into the trained measurement model to obtain the deformation field between the images, thus completing the measurement.

[0088] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A large deformation measurement method based on deep learning, characterized in that, Includes the following steps: S1. Use a Boolean model to render a reference image, apply a deformation field to the reference image to obtain a deformation image, and construct a dataset using the reference image, the deformation image and the corresponding deformation field. S2, Feature Extraction Stage: Extracting feature and deformation information from the images in the S1 dataset; Initial value estimation stage: The initial displacement value is estimated based on the feature information and deformation information of the image; Iterative refinement stage: The initial displacement value is iteratively refined to obtain the final displacement, thus completing the construction of the measurement model; S3, train and optimize the measurement model to obtain the final measurement model; S4. Acquire images and perform grayscale processing on the acquired images. For a series of frame images, the first frame image is used as the reference image, and each subsequent frame image is used as the deformation image. Inputting the reference image and the deformation image into the trained model will yield the deformation field between the two frames. In the initial value estimation stage, the obtained displacement is used to obtain the feature information of the deformed image The feature distortion processing is performed to obtain The reference image feature information and the distorted deformed image feature information are used to construct a corresponding cost volume, and a motion information expression is obtained. wherein, is the amount of deformation, is the deformation field computed in the previous layer, represents the correlation cost volume, , is the warped feature of the deformation map, describes the difference between the warped feature and the reference feature, , , , , , denotes a linear mapping with a convolutional layer; The result With image deformation information The update module is sent to the device, and the update is completed via GRU.

2. The method for measuring large deformations based on deep learning according to claim 1, characterized in that, In S1, the displacement is randomly generated. The specified pixel coordinate system is used as the reference. The corresponding pixel of the reference frame is retrieved on the deformed frame. It is determined whether the updated index coordinates are out of bounds. If they are out of bounds, the displacement is regenerated.

3. The method for measuring large deformations based on deep learning according to claim 1, characterized in that, In the feature extraction stage, feature information at different scales of the image is extracted through a feature encoder. The feature information of the reference image is The feature information of the deformation diagram is The context encoder extracts deformation information at different scales of the image. .

4. The method for measuring large deformations based on deep learning according to claim 1, characterized in that, The GRU formula is: in, , = , and For image deformation information The two parts are divided on the C channel. express convolution, Indicates splicing, and This represents the activation function. , , For different weight parameters; The final result obtained through the GRU formula is... ,make , in, Indicates a convolutional layer. and These are weight parameters. The variable represents the deformation field, and Mask represents the mask information used to upsample the deformation field.

5. The method for measuring large deformations based on deep learning according to claim 1, characterized in that, The iterative refinement stage is the same as the initial value estimation stage, used to estimate the residual displacement. The displacement estimate obtained in each iteration is: From this, a series of displacement estimates can be obtained.

6. The method for measuring large deformations based on deep learning according to claim 5, characterized in that, The series of displacement estimates are all at 1 / 4 resolution of the original image. They are upsampled by a factor of 4 using masking information to output a series of displacements. .

7. The method for measuring large deformations based on deep learning according to claim 1, characterized in that, The measurement model was implemented using the PyTorch framework and trained on an RTX3090. AdamW was chosen as the optimization method, and the loss was defined as: in, N represents the training rounds. Indicates a given actual displacement. This represents a series of displacements output by the model.

8. The method for measuring large deformations based on deep learning according to claim 1, characterized in that, The training optimization results are evaluated using the mean endpoint error. During training, if the current result achieves the best mean endpoint error (EPE) on the test set, then the training parameters for that set are determined, and the final measurement model is determined.

9. A large deformation measurement system based on deep learning, characterized in that, include: The dataset generation module is used to generate a series of reference images and deformation images, as well as the corresponding deformation fields; The model building module is used to extract information from the image generation module, complete the initial displacement value estimation, and perform iterative updates and refinement. The model training and optimization module trains and optimizes the model obtained from the model building module to determine the optimal training parameters and thus the final model. The measurement module is used to acquire images and perform grayscale processing, and input the images into the trained measurement model to obtain the deformation field between the images. In the initial value estimation stage, the obtained displacement features are used to estimate the deformation characteristics. Feature distortion processing is performed to obtain By utilizing the feature information of the reference image and the feature information of the distorted image, a corresponding cost volume is constructed, thereby obtaining the motion information expression: in, It is the amount of deformation. It is the deformation field calculated from the previous layer. Represents the relevant cost body, , These are the features of the distorted image. The differences between the distorted features and the reference features are described. , , , , , This indicates that a linear mapping is performed using a convolutional layer; The result With image deformation information The update module is sent to the device, and the update is completed via GRU.

Images