A deformation field prediction method based on image quality evaluation and adaptive pre-training

Abstract

The application relates to a deformation field prediction method based on image quality evaluation and adaptive pre-training. Wavelet packet decomposition is performed on a reference image and a deformation image of a collected component to obtain coefficients of each subband of each layer of the images, and Pearson correlation coefficients of each subband are calculated based on all the coefficients corresponding to the subband. The subbands are screened based on L2 norms of coefficient vectors of the subbands and the Pearson correlation coefficients, and the screened coefficient vectors of the subbands are reconstructed to obtain reconstructed reference images and reconstructed deformation images. Quality evaluation indexes of each reconstructed sample pair are calculated. Training branches to which the corresponding reconstructed sample pairs belong are determined based on the quality evaluation indexes, a U-Net model is preliminarily trained by using a preset supervised training strategy based on the reconstructed sample pairs of different training branches, the U-Net model is unsupervisedly trained, and displacement fields and strain fields are predicted based on the trained U-Net model.

Description

Technical Field

[0001] This application relates to the field of deformation field prediction technology, and in particular to a deformation field prediction method based on image quality evaluation and adaptive pre-training. Background Technology

[0002] Digital image correlation (DIC) technology has been widely applied in fields such as materials mechanics and structural deformation monitoring to achieve high-resolution measurements of displacement and deformation fields. However, in practical applications, when reference or deformation images exhibit extreme conditions such as large deformations or speckle tearing, the prediction results of traditional DIC models often suffer from poor accuracy and insufficient robustness. Furthermore, traditional DIC methods heavily rely on labeled data of real displacement or strain fields during model training, and these high-quality labels are often difficult or even impossible to obtain in real-world engineering scenarios. Additionally, existing DIC displacement prediction models generally employ purely supervised learning, resulting in lengthy training processes and difficulty in fully utilizing large amounts of unlabeled real-world measurement data, significantly limiting the model's engineering applicability and scalability. Summary of the Invention

[0003] Therefore, it is necessary to provide a deformation field prediction method based on image quality assessment and adaptive pre-training, including:

[0004] S1: Perform three-level wavelet packet decomposition on the reference image and deformation image of the acquired component to obtain the coefficients of each sub-band of each layer of the corresponding image; calculate the L2 norm of the coefficient vector of each sub-band, and calculate the Pearson correlation coefficient of each sub-band based on all the coefficients corresponding to each sub-band.

[0005] S2: Subbands are selected based on the L2 norm and Pearson correlation coefficient of the coefficient vectors of each subband, and inverse wavelet packet reconstruction is performed based on the coefficient vectors of the selected subbands to obtain the reconstructed reference image and the reconstructed deformed image.

[0006] S3: Calculate the quality evaluation index of any pair of reconstructed sample pairs based on contrast, signal-to-noise ratio, and gradient entropy. The reconstructed sample pairs include the reconstructed reference image and the reconstructed deformed image of the same part. Determine the training branch to which the corresponding reconstructed sample pair belongs based on the quality evaluation index. Use the corresponding preset supervised training strategy to initially train the U-Net model based on the reconstructed sample pairs of different training branches. Use the unsupervised training strategy to train the U-Net model a second time based on all the reconstructed sample pairs.

[0007] S4: Input the reference image and deformation image to be tested into the trained U-Net model, and output the predicted displacement field and strain field.

[0008] Beneficial effects: This method performs wavelet packet decomposition on the acquired reference image and deformation image of the component to obtain the coefficients of each sub-band of each layer of each image, and calculates the Pearson correlation coefficient of each sub-band based on all the coefficients corresponding to each sub-band; the sub-bands are filtered based on the L2 norm and Pearson correlation coefficient of the coefficient vector of each sub-band, and reconstructed based on the coefficient vector of the filtered sub-bands to obtain the reconstructed reference image and reconstructed deformation image; the quality evaluation index of each reconstructed sample pair is calculated; the training branch to which the corresponding reconstructed sample pair belongs is determined based on the quality evaluation index; the U-Net model is initially trained using the corresponding preset supervised training strategy based on the reconstructed sample pairs of different training branches, and the U-Net model is then trained unsupervised; the displacement field and strain field are predicted based on the trained U-Net model. This method significantly improves the engineering practicality, automatic adaptability and generalization ability of the prediction model in complex scenarios. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of the deformation field prediction method based on image quality evaluation and adaptive pre-training in the embodiments of this application. Detailed Implementation

[0011] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0012] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0013] like Figure 1 As shown, this embodiment provides a deformation field prediction method based on image quality assessment and adaptive pre-training, including:

[0014] S1: Perform three-level wavelet packet decomposition on the reference image and deformation image of the acquired component to obtain the coefficients of each sub-band of each layer of the corresponding image; calculate the L2 norm of the coefficient vector of each sub-band, and calculate the Pearson correlation coefficient of each sub-band based on all the coefficients corresponding to each sub-band.

[0015] Specifically, reference and deformation images are acquired through experimental platforms or actual working conditions. Both are two-dimensional grayscale functions of size H×W. They are denoted as follows: , ;

[0016] in, This represents the pixel grayscale value in the reference state (i.e., the reference image); This represents the pixel value at the same location after deformation (i.e., the deformed image). and These are the pixel position indices for the horizontal and vertical positions, respectively. , H and W represent the height and width of the image, respectively; i is the sample index, i=1,2,…,N. Both are usually speckle or random texture maps, used as tensor inputs for subsequent processing and training steps.

[0017] Furthermore, regarding the reference image With deformed images Perform three-level wavelet packet decomposition to calculate the coefficient vector of each sub-band at each level of the reference map and deformation map. , Specifically, for the reference image... With deformed images Perform a three-level wavelet packet decomposition. Expand the coefficients of the reference image at the k-th sub-band of the l-th level into a coefficient vector:

[0018] ;

[0019] Expand the coefficients of the deformed image in the same sub-band of the same layer into a coefficient vector:

[0020] ;

[0021] in, , Representing reference images With deformed images In the Layer The coefficient vector of each sub-band and Representing reference images With deformed images In the Layer The coefficients of each wavelet packet within a sub-band, where Q is the number of coefficients within the corresponding sub-band.

[0022] Calculate the L2 norm of all subband coefficients:

[0023] ;

[0024] ;

[0025] in, and The reference image and the deformed image are respectively in the 1st... Layer The total number of coefficients in each sub-band; The first reference image Layer The coefficient vector of each sub-band The first image representing the deformed image Layer The coefficient vector of each sub-band The first reference image Layer The first of the sub-bands One coefficient, The first image representing the deformed image Layer The first of the sub-bands One coefficient; and Corresponding to the reference image and the deformed image respectively Layer L2 norm of each subband; This indicates the calculation of the L2 norm.

[0026] Furthermore, the formula for calculating the Pearson correlation coefficient of the subband is:

[0027] ;

[0028] in, Indicates the first Layer Pearson correlation coefficient of individual bands Indicates the total number of coefficients. The first reference image Layer The first of the sub-bands One coefficient, express The mean, The first image representing the deformed image Layer The first of the sub-bands One coefficient, express The mean, Indicates the number of decomposition layers.

[0029] S2: Subbands are selected based on the L2 norm and Pearson correlation coefficient of the coefficient vectors of each subband, and inverse wavelet packet reconstruction is performed based on the coefficient vectors of the selected subbands to obtain the reconstructed reference image and the reconstructed deformed image.

[0030] Specifically, the steps include:

[0031] Set energy threshold Correlation threshold ;

[0032] when , and At that time, the reference image and the deformed image are preserved. Layer The coefficient vectors of each sub-band are used; otherwise, the coefficient vectors of the corresponding sub-bands of the reference image and the deformed image are set to zero. The first reference image Layer The coefficient vector of each sub-band The first image representing the deformed image Layer The coefficient vector of each sub-band Indicates the first Layer Pearson correlation coefficient of individual bands;

[0033] Inverse wavelet packet reconstruction is performed based on the coefficient vectors of the sub-bands retained in the reference image to obtain the reconstructed reference image. The reconstruction formula is as follows:

[0034] ;

[0035] in, Represents the reconstructed reference image. This indicates the inverse wavelet packet transform operation. The first reference image Layer The coefficient vector of each sub-band;

[0036] Inverse wavelet packet reconstruction is performed based on the coefficient vectors of the sub-bands preserved in the deformed image to obtain the reconstructed deformed image. The reconstruction formula is as follows:

[0037] ;

[0038] in, This represents a reconstructed deformed image. This indicates the inverse wavelet packet transform operation. The first image representing the deformed image Layer The coefficient vector of each sub-band.

[0039] S3: Calculate the quality evaluation index of any pair of reconstructed sample pairs based on contrast, signal-to-noise ratio, and gradient entropy. The reconstructed sample pairs include the reconstructed reference image and the reconstructed deformed image of the same part. Determine the training branch to which the corresponding reconstructed sample pair belongs based on the quality evaluation index. Use the corresponding preset supervised training strategy to initially train the U-Net model based on the reconstructed sample pairs of different training branches. Use the unsupervised training strategy to train the U-Net model a second time based on all the reconstructed sample pairs.

[0040] In this embodiment, the formula for calculating the quality evaluation index of the reconstructed sample pair is:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] in, Indicates the first The quality evaluation index for reconstructed sample pairs, Weighting coefficients representing image contrast. Indicates the first The image contrast of the reconstructed reference image or the reconstructed deformed image in the reconstructed sample pair. The weighting coefficients representing the signal-to-noise ratio. Indicates the first The signal-to-noise ratio of the reconstructed reference image or the reconstructed deformed image in the reconstructed sample pair. The weighting coefficients represent the image gradient entropy. Indicates the first The image gradient entropy of the reconstructed reference image or the reconstructed deformed image in the reconstructed sample pair. Indicates the first The standard deviation of grayscale values ​​for the reconstructed reference image or reconstructed deformed image in the reconstructed sample pair. Indicates the first The pixel mean of the reconstructed reference image or reconstructed deformed image in the reconstructed sample pair. Indicates the first The variance of noise estimation for the reconstructed reference image or reconstructed deformed image in the reconstructed sample pair. Indicates the grayscale gradient magnitude. The grayscale gradient magnitude is represented by The probability distribution, This represents the index used when performing discrete statistics on the grayscale gradient magnitude. , , Based on experience or optimization settings.

[0046] In this embodiment, the U-Net model is a standard U-Net structure. The main body of the network consists of four backbone layers for upsampling and downsampling, with the number of channels gradually expanding to 32-64-128-256. The LeakyReLU activation function is used, and the final output is pixel-level displacement and strain fields.

[0047] Furthermore, the step of determining the training branch to which the corresponding reconstructed sample pair belongs based on the quality evaluation index includes:

[0048] The quantile of all quality evaluation index values ​​is taken as the image quality evaluation threshold;

[0049] The quality evaluation index is compared with the image quality evaluation threshold. When the quality evaluation index is greater than or equal to the image quality evaluation threshold, the corresponding reconstructed sample pair is classified into the standard training branch; otherwise, the corresponding reconstructed sample pair is classified into the augmentation training branch.

[0050] Based on the reconstructed sample pairs of the standard training branch and the augmented training branch, the U-Net model is initially trained using the corresponding pre-set supervised training strategy.

[0051] Furthermore, the U-Net model is initially trained using corresponding preset supervised training strategies on the reconstructed sample pairs based on different training branches, including:

[0052] For the standard training branch, the first dynamic pre-training round number is calculated based on the quality evaluation index of each reconstructed sample pair in the standard training branch. The mean squared error loss and correlation loss are calculated based on any pair of reconstructed sample pairs in the standard training branch. The mean squared error loss and correlation loss are weighted and summed to obtain the first supervised loss of the corresponding reconstructed sample pair. The average of the first supervised losses of all reconstructed sample pairs is taken to obtain the supervised loss of the standard training branch. The network parameters of the U-Net model are backpropagated according to the supervised loss of the standard training branch until the first dynamic pre-training round number is reached, thus completing the initial training of the U-Net model in the standard training branch.

[0053] In this embodiment, the formula for calculating the first dynamic pre-training round number in the standard training branch is:

[0054] ;

[0055] ;

[0056] in, Indicates the first dynamic pre-training round number. Indicates the standard reference round number. This represents the first index adjustment factor. This indicates the first minimum round number of guaranteed items. Indicates rounding up. This represents the representative quality index of the reconstructed sample pairs in the standard training branch. Indicates the image quality evaluation threshold. This indicates the number of reconstructed sample pairs in the standard training branch. This represents the set of reconstructed samples in the standard training branch. Indicates the first A quality evaluation index for reconstructed sample pairs.

[0057] The standard formula for calculating the supervised loss of the training branch is:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] in, This represents the supervised loss of the standard training branch. Indicates the first The first supervised loss for reconstructing sample pairs, Indicates the first weight. Indicates the first The mean squared error loss for reconstructed sample pairs, Indicates the second weight. Indicates the first The correlation loss for reconstructed sample pairs, Indicates the first For the total number of pixels in the reconstructed sample pair, Indicates the first The horizontal displacement predicted by the U-Net model at each pixel. Indicates the first The actual horizontal displacement at each pixel. Indicates the first Vertical displacement predicted by the U-Net model at each pixel Indicates the first The actual vertical displacement at each pixel. This represents the relevance value predicted by the U-Net model. This represents the true correlation value. This represents the correlation coefficient function.

[0063] For the augmented training branch, the second dynamic pre-training round number is calculated based on the quality evaluation index of each reconstructed sample pair in the augmented training branch. The perceptual loss and physical consistency loss are calculated based on any pair of reconstructed sample pairs in the augmented training branch. The second supervised loss of the corresponding reconstructed sample pair is calculated based on the perceptual loss and physical consistency loss. The average of the second supervised losses of all reconstructed sample pairs is taken to obtain the supervised loss of the augmented training branch. The network parameters of the U-Net model are backpropagated according to the supervised loss of the augmented training branch until the second dynamic pre-training round number is reached, thus completing the initial training of the U-Net model of the augmented training branch.

[0064] In the augmented training branch, the formula for calculating the second dynamic pre-training round is:

[0065] ;

[0066] ;

[0067] in, Indicates the second dynamic pre-training round number. Indicates the number of enhancement reference rounds, This represents the second index adjustment factor. This indicates the second minimum round number guarantee item. Indicates rounding up. This represents the representative quality index of the reconstructed sample pairs in the augmented training branch. Indicates the image quality evaluation threshold. This indicates the number of reconstructed sample pairs in the augmentation training branch. This represents the set of reconstructed samples in the augmentation training branch. Indicates the first A quality evaluation index for reconstructed sample pairs.

[0068] The formula for calculating the supervision loss of the enhanced training branch is:

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] in, This indicates the enhanced training branch supervision loss. Indicates the first The second supervised loss for reconstructing sample pairs, Indicates the third weight. Indicates the first branch of reinforcement training The second mean square error loss for the reconstructed sample pairs, Indicates the fourth weight. Indicates the first branch of reinforcement training For the second correlation loss of the reconstructed sample pairs, Indicates the fifth weight. Indicates the first The perceptual loss for reconstructed sample pairs, Indicates the sixth weight. Indicates the first The physical consistency loss of the reconstructed sample pairs, This indicates the number of network layers in the U-Net model. Indicates the first The perceptual feature vector output by the layer network The U-Net model is based on the first The image representation predicted from the reconstructed deformed image in the reconstructed sample pair. Indicates the first The image representation of the reconstructed reference image in the reconstructed sample pair. Indicates the first For the total number of pixels in the reconstructed sample pair, Indicates the number of strain component categories. Indicates the first The predicted pixel at the nth pixel Strain-like components, Indicates difference, Indicates the first The horizontal displacement predicted by the U-Net model at each pixel.

[0074] In this embodiment, in the reinforcement training branch, the following auxiliary branches can be activated based on the backbone network:

[0075] Edge perception discrimination branch: A shallow CNN is connected in parallel, the input is the last layer feature of the main branch, and the output is a pixel-level edge confidence map. The number of channels is 1, and the sigmoid function is normalized to [0,1].

[0076] Attention mechanism: Channel attention modules are inserted during the network encoding stage to adaptively amplify anomalous features, often implemented using SE-block. The network output includes the main prediction / auxiliary branch results.

[0077] Furthermore, the second training of the U-Net model based on all reconstructed sample pairs using an unsupervised training strategy includes:

[0078] Based on all reconstructed sample pairs, the pixel reconstruction loss, third correlation loss, second physical consistency loss, and second perceptual loss are calculated, and then weighted summation of these losses yields the unsupervised joint training loss. The expression for the unsupervised joint training loss is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] in, This indicates the losses from unsupervised joint training. , , , Let represent the weight hyperparameters of the pixel reconstruction loss, the third correlation loss, the second physical consistency loss, and the second perceptual loss, respectively. Indicates pixel reconstruction loss, This represents the third correlation loss. This represents the second physical consistency loss. This indicates the loss of second perception; It is the original reference image of the i-th reconstructed sample pair; It is the original deformed image of the i-th reconstructed sample pair. and These are the reconstructed reference image and the reconstructed deformed image obtained from the network reconstruction output, respectively. The sum of absolute values ​​for each pixel; For the set of subbands participating in the constraints, The number of sub-bands within the set. For reference image and deformed images In the Layer The Pearson correlation coefficient is calculated from the coefficient vector of each sub-band. To reconstruct the reference image and reconstructing deformed images The calculated Pearson correlation coefficient; The gradient of the image, where I refers to the image. For gradient entropy; This represents a perceptual feature extraction network; This represents the square of the L2 norm.

[0086] The network parameters of the U-Net model are backpropagated based on the unsupervised joint training loss until the preset number of unsupervised training rounds is reached (this number of rounds can be flexibly adjusted according to the overall training convergence and experimental experience), thus completing the secondary training of the U-Net model.

[0087] S4: Input the reference image and deformation image to be tested into the trained U-Net model, and output the predicted displacement field and strain field.

[0088] Specifically, through the aforementioned optimization training, the model ultimately forms an end-to-end dual-branch output, which can directly output the corresponding predicted displacement field and deformation field based on the input reference map and deformation map:

[0089] ;

[0090] ;

[0091] In the above formula, For the first For the image at the pixel The predicted output tensor; This is the U-Net model after training and optimization; This represents the reference image to be tested. Represents the deformation image to be measured; For the first For the image at the pixel The predicted horizontal displacement component; For the first For the image at the pixel The predicted vertical displacement component; For the first For the image at the pixel The predicted first type of strain component, For the first For the image at the pixel The predicted type 2 strain component For the first For the image at the pixel The predicted first Strain-like components. The predicted output size for each pair of samples is consistent with the original input, requiring no additional differencing or post-processing, and can be directly used for large-scale automatic monitoring and analysis in engineering scenarios.

[0092] In this embodiment, experiments were conducted using aerospace-grade aluminum alloy sheet as the component to verify the feasibility of the method:

[0093] 1. Acquisition of Reference and Deformation Images: Taking the experiment of aerospace aluminum alloy thin plates under large deformation and crack propagation conditions as an example, an industrial camera with a resolution of 2048×2048 pixels and a pixel size of 5.5μm was used to simultaneously acquire images of the reference state and the deformation state on a physical loading testing machine at a sampling frequency of 5Hz. Each sample set includes one reference image and one deformation image, totaling 8000 sets. All images are high-contrast pure speckle distributions, covering extreme conditions such as crack propagation, tearing, regional blurring, and edge anomalies. All data were uniformly input as H×W grayscale tensors.

[0094] 2. Image Hierarchical Reconstruction via Wavelet Packet Decomposition: A three-level Daubechies 4 wavelet packet decomposition is applied to each set of reference and deformed images to obtain multi-scale sub-band coefficients. Subsequently, the L2 norm and Pearson correlation coefficient of each sub-band are calculated. Based on set energy thresholds (e.g., 60% of the mean of each layer) and correlation thresholds (e.g., 0.4), high-quality feature sub-bands are selected and retained. The selected coefficient set is then reconstructed using inverse wavelet packet decomposition to obtain the cleaned reference and deformed images, effectively achieving noise filtering and feature enhancement, providing higher-quality input for subsequent modeling and training.

[0095] 3. Calculate the image quality evaluation index for each pair of samples. For all processed image pairs, the contrast C, signal-to-noise ratio (SNR), and gradient entropy H are extracted respectively, by setting weight parameters ( =0.35, =0.3, =0.35) Calculate the comprehensive quality evaluation index , The distribution range is [0.41, 0.98]. After being uniformly normalized, it is used for subsequent branch training allocation and dynamic resource scheduling.

[0096] 4. Training branch decision for each pair of samples: based on the quality index of each sample group. With preset threshold ( Perform automatic branch decision: when At that time, the samples were assigned to the standard training branch, using the standard U-Net structure, with loss function weights α=0.8 (mean squared error) and β=0.2 (correlation loss), and network encoding-decoding channels of 32, 64, 128, and 256 respectively. The first dynamic pre-training rounds were evaluated using a representative quality index. Dynamic allocation; when When this happens, it is categorized into the abnormal enhancement training branch. The backbone network structure remains the same, with an additional parallel edge-aware discrimination CNN branch and an SE-block attention mechanism inserted. The weights of the loss term are adjusted. =0.5、 =0.1、 =0.25 (perceptual loss) =0.15 (physical consistency loss), training epochs are based on representative quality index Adaptive scaling. This strategy enables differentiated management of training resources and learning difficulty, significantly improving the model's stability and generalization ability under various extreme conditions.

[0097] 5. Supervised Pre-training and Unsupervised Joint Training: All samples undergo supervised pre-training according to their respective configurations based on branch decisions, using the Adam optimizer with an initial learning rate of 1e-4 and a batch size of 8. The branch training loss uses the mean loss of all samples within each branch as the backpropagation objective, and training is completed using adaptive branch iterations. After supervised pre-training converges, the model uniformly enters the unsupervised joint training phase, applying a multi-objective joint loss (reconstruction loss, correlation loss, physical consistency loss, and perceptual loss, etc., with weights...) to all labeled and unlabeled samples. =0.4、 =0.2、 =0.25、 =0.15), with a total of 50 rounds of training.

[0098] 6. End-to-end dual-branch output: After final optimization, the model can automatically output displacement and strain field prediction results with the same resolution as the input image, end-to-end. Specifically, each pair of samples can directly obtain horizontal displacement, vertical displacement, and multiple strain field components. No additional post-processing is required, allowing direct application to large-scale automatic monitoring and analysis in real-world engineering scenarios. Experimental verification shows that under high noise and strong abnormal conditions, the mean square error of the predicted displacement field is reduced from 0.17 pixels to 0.11 pixels, and the strain field error is reduced from 0.28% to 0.14%. The overall training and inference speed is increased to 1.9 times that of the original method, fully demonstrating the significant engineering advantages and practical application value of this innovative technology system.

[0099] This embodiment addresses the bottlenecks of existing digital image correlation (DIC) techniques, such as insufficient robustness under extreme conditions and strong dependence on high-quality labeled samples. It innovatively constructs an end-to-end displacement and strain field prediction method with image quality assessment as its core and adaptive capabilities. This method significantly improves the engineering practicality, automatic adaptability, and generalization ability of the prediction model in complex scenarios by proposing multi-scale image quality assessment, a novel image quality assessment index, and adaptive branch training decisions based on the image quality assessment index. The innovations are specifically reflected in the following three aspects:

[0100] 1. An image quality-driven end-to-end deformation / strain field prediction framework is proposed: This invention proposes an integrated multi-module end-to-end prediction system, covering image acquisition, wavelet packet decomposition denoising, image quality assessment, adaptive decision-making for training branches, dynamic loss optimization, and a joint supervised-unsupervised overall training process. This overall framework streamlines the entire process from raw image input to displacement and strain field prediction outputs, significantly improving automation and engineering adaptability, while providing a strong technical foundation for large-scale structural health monitoring.

[0101] 2. A method for measuring image quality is proposed: This invention innovatively designs a comprehensive image quality evaluation index that integrates three factors: contrast, signal-to-noise ratio, and gradient entropy. This index can quantify the quality of different image samples, sensitively distinguishing between high-quality and low-quality samples, and providing a scientific basis for subsequent training branches and dynamic allocation of computing resources.

[0102] 3. Establishing an adaptive branch training and resource allocation mechanism based on image quality evaluation index: This invention addresses the significant differences in input sample quality by automatically assigning each pair of samples to either a standard training branch or an anomaly enhancement training branch based on their comprehensive quality evaluation index and a preset threshold. The number of training rounds for each branch is dynamically determined according to the quality index distribution of its sample set, achieving differentiated training round scheduling for samples. This ensures efficient and accurate convergence for high-quality samples and robust fitting for low-quality and anomaly samples. After completing adaptive branch supervised training, all branch samples are uniformly incorporated into the unsupervised joint training stage, further enhancing the model's generalization ability and engineering adaptability under diverse and unlabeled conditions.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A deformation field prediction method based on image quality assessment and adaptive pre-training, characterized in that, include: S1: Perform three-level wavelet packet decomposition on the reference image and deformation image of the acquired component to obtain the coefficients of each sub-band of each layer of the corresponding image; Calculate the L2 norm of the coefficient vector of each sub-band, and calculate the Pearson correlation coefficient of each sub-band based on all the coefficients corresponding to each sub-band. S2: Subbands are selected based on the L2 norm and Pearson correlation coefficient of the coefficient vectors of each subband, and inverse wavelet packet reconstruction is performed based on the coefficient vectors of the selected subbands to obtain the reconstructed reference image and the reconstructed deformed image. S3: Calculate the quality evaluation index of any pair of reconstructed sample pairs based on contrast, signal-to-noise ratio, and gradient entropy. The reconstructed sample pairs include the reconstructed reference image and the reconstructed deformed image of the same part. The training branch to which the corresponding reconstructed sample pair belongs is determined based on the quality evaluation index, including: The quantile of all quality evaluation index values ​​is taken as the image quality evaluation threshold; The quality evaluation index is compared with the image quality evaluation threshold. When the quality evaluation index is greater than or equal to the image quality evaluation threshold, the corresponding reconstructed sample pair is classified into the standard training branch; otherwise, the corresponding reconstructed sample pair is classified into the augmentation training branch. Based on the reconstructed sample pairs of the standard training branch and the augmented training branch, the U-Net model is initially trained using the corresponding pre-set supervised training strategy; The U-Net model is initially trained using reconstructed sample pairs based on different training branches and corresponding pre-defined supervised training strategies, including: For the standard training branch, the first dynamic pre-training round number is calculated based on the quality evaluation index of each reconstructed sample pair in the standard training branch. The mean squared error loss and correlation loss are calculated based on any pair of reconstructed sample pairs in the standard training branch. The mean squared error loss and correlation loss are weighted and summed to obtain the first supervised loss of the corresponding reconstructed sample pair. The average of the first supervised losses of all reconstructed sample pairs is taken to obtain the supervised loss of the standard training branch. The network parameters of the U-Net model are backpropagated according to the supervised loss of the standard training branch until the first dynamic pre-training round number is reached, thus completing the initial training of the U-Net model in the standard training branch. For the augmented training branch, the second dynamic pre-training round number is calculated based on the quality evaluation index of each reconstructed sample pair in the augmented training branch. The perceptual loss and physical consistency loss are calculated based on any pair of reconstructed sample pairs in the augmented training branch. The second supervised loss of the corresponding reconstructed sample pair is calculated based on the perceptual loss and physical consistency loss. The average of the second supervised losses of all reconstructed sample pairs is taken to obtain the supervised loss of the augmented training branch. The network parameters of the U-Net model are backpropagated according to the supervised loss of the augmented training branch until the second dynamic pre-training round number is reached, thus completing the initial training of the U-Net model of the augmented training branch. The U-Net model is then retrained using an unsupervised training strategy based on all the reconstructed sample pairs. S4: Input the reference image and deformation image to be tested into the trained U-Net model, and output the predicted displacement field and strain field.

2. The deformation field prediction method based on image quality assessment and adaptive pre-training according to claim 1, characterized in that, The formula for calculating the Pearson correlation coefficient of the subband is: ； in, Indicates the first Layer Pearson correlation coefficient of individual bands Indicates the total number of coefficients. The first reference image Layer The first of the sub-bands One coefficient, express The mean, The first image representing the deformed image Layer The first of the sub-bands One coefficient, express The mean.

3. The deformation field prediction method based on image quality evaluation and adaptive pre-training according to claim 1, characterized in that, S2 include: Set energy threshold Correlation threshold ; when , and At that time, the reference image and the deformed image are preserved. Layer The coefficient vectors of each sub-band are used; otherwise, the coefficient vectors of the corresponding sub-bands of the reference image and the deformed image are set to zero. The first reference image Layer The coefficient vector of each sub-band The first image representing the deformed image Layer The coefficient vector of each sub-band Indicates the first Layer Pearson correlation coefficient of individual bands; Inverse wavelet packet reconstruction is performed based on the coefficient vector of the sub-band retained in the reference image to obtain the reconstructed reference image; The reconstructed deformed image is obtained by performing inverse wavelet packet reconstruction based on the coefficient vector of the sub-band retained in the deformed image.

4. The deformation field prediction method based on image quality assessment and adaptive pre-training according to claim 1, characterized in that, In S3, the formula for calculating the quality evaluation index of the reconstructed sample pairs is: ； ； ； ； in, Indicates the first The quality evaluation index for reconstructed sample pairs, Weighting coefficients representing image contrast. Indicates the first The image contrast of the reconstructed reference image or the reconstructed deformed image in the reconstructed sample pair. The weighting coefficients representing the signal-to-noise ratio. Indicates the first The signal-to-noise ratio of the reconstructed reference image or the reconstructed deformed image in the reconstructed sample pair. The weighting coefficients represent the image gradient entropy. Indicates the first The image gradient entropy of the reconstructed reference image or the reconstructed deformed image in the reconstructed sample pair. Indicates the first The standard deviation of grayscale values ​​for the reconstructed reference image or reconstructed deformed image in the reconstructed sample pair. Indicates the first The pixel mean of the reconstructed reference image or reconstructed deformed image in the reconstructed sample pair. Indicates the first The variance of noise estimation for the reconstructed reference image or reconstructed deformed image in the reconstructed sample pair. Indicates the grayscale gradient magnitude. The grayscale gradient magnitude is represented by The probability distribution, This represents the index used when performing discrete statistics on grayscale gradient magnitudes.

5. The deformation field prediction method based on image quality assessment and adaptive pre-training according to claim 1, characterized in that, In the standard training branch, the formula for calculating the first dynamic pre-training round is: ； ； in, Indicates the first dynamic pre-training round number. Indicates the standard reference round number. This represents the first index adjustment factor. This indicates the first minimum round number of guaranteed items. Indicates rounding up. This represents the representative quality index of the reconstructed sample pairs in the standard training branch. Indicates the image quality evaluation threshold. This indicates the number of reconstructed sample pairs in the standard training branch. This represents the set of reconstructed samples in the standard training branch. Indicates the first A quality evaluation index for reconstructed sample pairs; The standard formula for calculating the supervised loss of the training branch is: ； ； ； ； in, This represents the supervised loss of the standard training branch. Indicates the first The first supervised loss for reconstructing sample pairs, Indicates the first weight. Indicates the first The mean squared error loss for reconstructed sample pairs, Indicates the second weight. Indicates the first The correlation loss for reconstructed sample pairs, Indicates the first For the total number of pixels in the reconstructed sample pair, Indicates the first The horizontal displacement predicted by the U-Net model at each pixel. Indicates the first The actual horizontal displacement at each pixel. Indicates the first Vertical displacement predicted by the U-Net model at each pixel Indicates the first The actual vertical displacement at each pixel. This represents the relevance value predicted by the U-Net model. This represents the true correlation value. This represents the correlation coefficient function.

6. The deformation field prediction method based on image quality assessment and adaptive pre-training according to claim 1, characterized in that, In the augmented training branch, the formula for calculating the second dynamic pre-training round is: ； ； in, Indicates the second dynamic pre-training round number. Indicates the number of enhancement reference rounds, This represents the second index adjustment factor. This indicates the second minimum round number guarantee item. Indicates rounding up. This represents the representative quality index of the reconstructed sample pairs in the augmented training branch. Indicates the image quality evaluation threshold. This indicates the number of reconstructed sample pairs in the augmentation training branch. This represents the set of reconstructed samples in the augmentation training branch. Indicates the first A quality evaluation index for reconstructed sample pairs; The formula for calculating the supervision loss of the enhanced training branch is: ； ； ； ； in, This indicates the enhanced training branch's supervised loss. Indicates the first The second supervised loss for reconstructing sample pairs, Indicates the third weight. Indicates the first branch of reinforcement training The second mean square error loss for the reconstructed sample pairs, Indicates the fourth weight. Indicates the first branch of reinforcement training For the second correlation loss of the reconstructed sample pairs, Indicates the fifth weight. Indicates the first The perceived loss for reconstructed sample pairs, Indicates the sixth weight. Indicates the first The physical consistency loss of the reconstructed sample pairs, This indicates the number of network layers in the U-Net model. Indicates the first The perceptual feature vector output by the layer network. The U-Net model is based on the first The image representation predicted from the reconstructed deformed image in the reconstructed sample pair. Indicates the first The image representation of the reconstructed reference image in the reconstructed sample pair. Indicates the first For the total number of pixels in the reconstructed sample pair, Indicates the number of strain component categories. Indicates the first The predicted pixel at the nth pixel Strain-like components, Indicates difference, Indicates the first The horizontal displacement predicted by the U-Net model at each pixel.

7. The deformation field prediction method based on image quality assessment and adaptive pre-training according to claim 1, characterized in that, In S3, the second training of the U-Net model based on all reconstructed sample pairs using an unsupervised training strategy includes: Calculate pixel reconstruction loss, third correlation loss, second physical consistency loss, and second perceptual loss based on all reconstructed sample pairs, and then sum the pixel reconstruction loss, third correlation loss, second physical consistency loss, and second perceptual loss in a weighted manner to obtain the unsupervised joint training loss; The network parameters of the U-Net model are backpropagated based on the unsupervised joint training loss until the preset number of unsupervised training rounds are reached, thus completing the secondary training of the U-Net model.

8. The deformation field prediction method based on image quality assessment and adaptive pre-training according to claim 1, characterized in that, The component includes aerospace-grade aluminum alloy sheet.

Images