Adaptive subset configuration method and device based on feature matching in dic measurement

Abstract

The application discloses a kind of in DIC measurement based on image feature matching's self-adapting subset configuration method and device, method includes: using SIFT algorithm to extract speckle image feature point, and using improved GMS algorithm to eliminate false matching point to obtain matching feature point pair;Based on matching feature point pair, construct Delaunay triangular grid, and divide region with triangle as basic unit;By calculating the minimum circumscribed square of each triangle, determine the adaptive subset size of each local area;With the preliminary displacement obtained by feature point matching as the basis, the whole field displacement is calculated by using IC-GN method.The device includes: processor and memory.The application uses digital image processing and feature matching technology, with the help of the distribution characteristics and matching characteristics of feature points in speckle images before and after deformation, realizes the adaptive configuration of subset size, so as to optimize the measurement accuracy and robustness without manual intervention.

Description

Technical Field

[0001] This invention relates to the field of digital image correlation (DIC) measurement technology, and in particular to an adaptive subset configuration method and apparatus based on feature matching in DIC measurement. Background Technology

[0002] Digital image correlation (DIC) is a non-contact optical measurement technique that compares the pixel displacements of two speckle images before and after deformation to obtain full-field deformation information. This method can accurately measure displacement and strain under various environmental conditions. Currently, DIC has become an indispensable tool in the field of object deformation measurement and is widely used in numerous engineering fields.

[0003] Currently, there are various implementations of DIC methods, which can be categorized into local subset methods and global DIC methods. Local subset methods select a fixed-size subset (e.g., 21×21 pixels) centered at each calculation point and match the grayscale distribution of the subset before and after deformation using a cross-correlation function. This type of algorithm is widely used, including the forward Gauss-Newton method (FA-GN) and the inverse Gauss-Newton method (IC-GN). Global Digital Image Correlation (GDIC) treats the entire image region as a whole and achieves matching by optimizing global displacement field parameters. It is suitable for large or non-uniform deformations, but has high computational complexity and is often combined with the finite element method for displacement field interpolation. In addition, there are emerging DIC measurement methods based on deep learning or neural networks, which have achieved good results in specific speckle images. However, the black-box nature of deep learning neural network methods leads to controversy regarding the accuracy of their predictions.

[0004] In subset-based Discrete Indication (DIC) measurement methods, the size of the subset is a crucial factor affecting measurement accuracy and robustness. To ensure correct registration between the reference and target images before and after deformation, the speckle patterns in the two speckle images need to be randomly distributed. The subset needs to be large enough to cover a sufficient number of speckles because a larger subset contains richer gray-level distribution features, uniquely identifying the region for correct registration, reducing the impact of random noise, and improving the stability of the measurement results. However, in bending or high strain gradient regions of complex deformation fields, an excessively large subset may contain multiple deformation modes. Since DIC measurement methods typically use first- or second-order shape functions, a large subset cannot accurately describe the deformation information, leading to systematic errors. Conversely, an excessively small subset may result in insufficient features within the subset region, failing to uniquely match the image regions before and after deformation, causing matching errors and reducing accuracy. Furthermore, small subsets are more sensitive to noise, especially in low-contrast or high-noise environments, leading to larger fluctuations in measurement results. Therefore, the choice of subset size requires a balance between feature uniqueness and deformation approximation accuracy. The core challenge lies in ensuring that the subset simultaneously meets the requirements of feature uniqueness and deformation approximation accuracy.

[0005] Currently, in existing DIC implementations, the size and shape of the subset are typically configured manually by the user, heavily reliant on the operator's skill level. For users lacking extensive experience, setting a subset that strikes a balance between measurement accuracy and resolution often requires trial and error, a time-consuming and laborious process. Furthermore, most current DIC software uses globally configured subsets to process all points of interest (POIs), which cannot guarantee the quality of measurement results for POIs located in regions with sparse speckle patterns or drastic deformation. Local defects and complex, unknown deformations in speckle patterns often make selecting a suitable subset for the test sample surface difficult.

[0006] In recent years, researchers have proposed several adaptive subset configuration methods. For example, Sun et al. introduced the concept of subset entropy to quantify the quality of subset image patterns in DIC analysis, setting lower and upper limits for subset size based on subset entropy. Ma et al. proposed a method using pixel-level ZNCC (zero-mean normalized cross-correlation) function values ​​as correlation criteria, pre-setting various subsets of different forms according to practical application scenarios. The proposed algorithm can automatically select the most suitable subset for matching under the guidance of the pixel-level ZNCC value distribution range. Yuan et al. proposed a DIC strategy based on point shape function-assisted subset selection (SFSS), which determines the optimal subset size for each sampling point by quantitatively evaluating the deformation deviation within a given subset size range. However, existing adaptive subset configuration methods still have limitations in complex deformation scenarios, making it difficult to simultaneously consider the requirements of local speckle pattern quality and deformation field smoothness. Summary of the Invention

[0007] This invention provides an adaptive subset configuration method and apparatus based on feature matching in DIC measurement. Utilizing digital image processing and feature matching techniques, and leveraging the distribution and matching characteristics of feature points in speckle images before and after deformation, this invention achieves adaptive subset size configuration, thereby optimizing measurement accuracy and robustness without manual intervention. Details are described below:

[0008] A first aspect is an adaptive subset configuration method based on image feature matching in DIC measurement, the method comprising:

[0009] The SIFT algorithm is used to extract feature points from speckle images, and an improved GMS algorithm is used to eliminate mismatched points to obtain matching feature point pairs.

[0010] Delaunay triangular meshes are constructed based on matching feature point pairs, and regions are divided using triangles as the basic units. The size of the adaptive subset of each local region is determined by calculating the minimum circumscribed square of each triangle.

[0011] Based on the preliminary displacement obtained by feature point matching, the IC-GN method is used to traverse and calculate the displacement of the entire field.

[0012] The step of using the improved GMS algorithm to eliminate mismatched points and obtain matching feature point pairs is as follows:

[0013] A statistical algorithm is used to calculate the optimal grid edge distance and assign edge feature points to adjacent grids, thus addressing the shortcomings in edge feature point processing.

[0014] The method of statistically analyzing the feature scores of eight adjacent symmetrical grids was changed to a five-grid feature score statistical method to eliminate erroneous matching points in the initial matching.

[0015] Using the aspect ratio of the input image as a constraint, the values ​​of P and Q are initialized with empirical values ​​E, so that the divided grid is close to a regular square.

[0016] The method for calculating the feature scores of the five-square grid is as follows:

[0017] The sum of the five grid feature scores is called the five-grid feature score S: , Let the feature score of the j-th grid within the five-square grid containing the i-th grid be calculated. Then, rotate the five-square grid clockwise and calculate the maximum feature score of the grid after three rotations. ,

[0018]

[0019] In the formula, To calculate the feature score of the j-th grid after rotating the i-th grid k times, we calculate the average number of coarse matching features among the grids in the current five-grid:

[0020]

[0021] In the formula, The number of coarse matching features within the j-th grid in the five-grid containing the i-th grid, when the grid feature score Greater than the grid feature score threshold At that time, the judgment A correct match is defined as a match where the condition is not met, and an incorrect match is defined as a grid feature score threshold. Calculation formula:

[0022]

[0023]

[0024] In the formula , , For threshold coefficient, Let F represent the i-th pair of matching points, where F represents an incorrect match and T represents a correct match.

[0025] The values ​​of P and Q are initialized using the empirical value E:

[0026]

[0027] In the formula, and Given the width and height of the input speckle image I, respectively, shift the grid pattern by half a cell width in both the x and y directions and the x and y directions, and then repeat the algorithm 3 times.

[0028] The step of determining the adaptive subset size of each local region by calculating the minimum circumscribed square of each triangle is as follows:

[0029] The minimum circumscribed square of the triangle can be obtained by rotation, and the rotation angle can be determined by gradient descent or bisection, or the optimal solution can be obtained by traversing all angles.

[0030] Based on the optimal solution, the minimum circumscribed square side length of the subset is L obtained through feature point distribution statistics. Therefore, the subset size S is set to the smallest odd number not less than L, i.e.:

[0031]

[0032] in, Indicates rounding up. It is the preset maximum subset size.

[0033] A second aspect is an adaptive subset configuration apparatus based on image feature matching in DIC measurement, characterized in that the apparatus comprises: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the apparatus to perform the method described in any one of the first aspects.

[0034] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0035] The beneficial effects of the technical solution provided by this invention are:

[0036] 1. This method utilizes the distribution and matching characteristics of feature points in speckle images, and uses feature point density and distribution quality as indicators to measure the local quality of speckle patterns. It automatically determines the optimal subset size of each local region through Delaunay triangular mesh, solving the problem that the subset size needs to be manually configured in traditional DIC measurement.

[0037] 2. This method takes into account both the quality of the local speckle pattern used for registration and the smoothness of the deformation field. In dense speckle areas, a smaller subset is automatically used to capture local deformation details, while in sparse speckle areas, a larger subset is used to ensure the uniqueness of the matching and noise resistance, thus achieving an adaptive balance between measurement accuracy and robustness.

[0038] 3. This method uses an improved GMS (Grid Motion Statistics) algorithm to eliminate mismatched points. Compared with the traditional RANSAC (Random Sample Consensus) method, it significantly improves the matching rate while maintaining a large number of matching point pairs, providing more reliable initial values ​​for subsequent subset configuration and displacement calculation.

[0039] 4. This method uses feature point density and distribution quality as quantitative indicators of the local quality of speckle images, and adaptively correlates the subset size with the local information richness of the speckle pattern, effectively overcoming the adverse effects of uneven speckle image quality on the measurement results.

[0040] 5. Experimental results of this invention show that, under conditions of significant local quality differences in speckle patterns, this method achieves high and uniform correlation across the entire domain (average ZNCC of approximately 0.98), eliminating false high correlation and matching degradation caused by local image quality differences. Under non-ideal imaging conditions and scenarios with uneven speckle quality, it significantly improves the confidence and reliability of measurement results. Attached Figure Description

[0041] Figure 1 A schematic diagram of feature point matching and matching point pairs after eliminating mismatched points;

[0042] Figure 2 A schematic diagram of the constructed Delaunay triangular mesh and the removed edge triangles;

[0043] Figure 3 A schematic diagram to determine the circumscribed square;

[0044] Figure 4 This is a schematic diagram of the displacement field within the corresponding triangle;

[0045] Figure 5 The generated speckle image and the theoretical displacement value in the x-direction;

[0046] Figure 6 A comparison chart showing the theoretical displacement and the error of the comparison algorithm;

[0047] Wherein, (a) is the theoretical displacement; (b) is the SIFT+DIC error; (c) is the SURF+DIC error; and (d) is the algorithm of this invention.

[0048] Figure 7The weld location and the area to be measured are marked in the acquired speckle image;

[0049] Figure 8 This is a flowchart of an adaptive subset configuration method based on image feature matching in DIC measurement. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0051] Example 1

[0052] To overcome the shortcomings of existing technologies, this invention presents an adaptive subset configuration method based on image feature matching in DIC measurement, which consists of the following core steps:

[0053] 101: For example Figure 1 The SIFT (Scale Invariant Feature Transform) algorithm is used to extract feature points from speckle images, and an improved GMS algorithm is used to eliminate mismatched points to obtain high-precision matching feature point pairs.

[0054] The core inventive concept of this step is that the spatial distribution density of feature points directly reflects the richness of local information in the speckle pattern, while the matching accuracy is a prerequisite for the stability of subsequent geometric subdivision.

[0055] Existing DIC methods, such as SIFT and SURB, can extract a large number of feature points, but the matching results contain a large number of mismatched point pairs. Directly using point sets containing mismatches for triangulation will lead to incorrect triangle topological relationships, thus rendering subset configurations completely ineffective. Therefore, it is essential to eliminate mismatches with the highest possible accuracy while ensuring the number of matching point pairs is maintained.

[0056] Therefore, this step first uses the SIFT algorithm to extract feature points from the reference speckle image and the target speckle image. Its core advantage lies in constructing a scale space through the difference of Gaussian pyramid, which enables the extracted feature points to have scale invariance and rotation invariance, and can maintain stable extraction results under conditions such as image deformation and illumination changes.

[0057] Building upon this foundation, this step employs an improved GMS algorithm to eliminate mismatches. The basic principle of the GMS algorithm is that correctly matched feature points exhibit spatial consistency within their neighboring regions; that is, correctly matched points tend to correctly match their surrounding pairs of matched points as well. Incorrect matches, however, do not possess this statistical characteristic. Based on this statistical law, the GMS algorithm distinguishes between correct and incorrect matches by statistically analyzing the matching consistency within the neighborhood. Essentially, it transforms the feature matching problem into a statistical testing problem of neighborhood support.

[0058] 102: Construct Delaunay triangular meshes based on matching feature point pairs, and divide regions using triangles as the basic units;

[0059] The core inventive concept of this step is to transform the discrete feature point set into a continuous region partition with geometric constraints through Delaunay triangulation, so that the spatial variation of the subset size is subject to the natural constraints of the triangle geometry, thereby avoiding abrupt changes in the subset size.

[0060] After obtaining highly reliable matching feature point pairs, the discretely distributed feature points need to be organized into computable spatial units. This step uses Delaunay triangulation to construct a triangular mesh, whose geometric properties are highly compatible with the requirements of this invention.

[0061] (1) Maximize the minimum angle property. Among all possible triangulations of a given set of points, the Delaunay triangulation maximizes the minimum angle of the triangle, avoiding the generation of elongated triangles. This property ensures that the side lengths of the circumscribed squares of adjacent triangles have a smooth transition in space, thereby avoiding abrupt changes in subset size due to mesh distortion.

[0062] (2) Uniqueness and low time complexity. When no four points in the point set are concyclic, the Delaunay triangulation is unique and the construction complexity is O(n log n), which is suitable for efficient processing of large-scale feature point sets.

[0063] (3) Empty circumcircle criterion. The circumcircle of any triangle does not contain any other points. This criterion ensures that the triangle is most sensitive to the local point set density, and the side length of the circumscribed square can accurately reflect the local feature distribution density of the region.

[0064] After constructing the triangular mesh, to eliminate the elongated triangles caused by the sparse feature points in the edge regions (whose estimation of subset size is unreliable), edge triangles whose perimeter-to-area ratio exceeds a preset threshold are discarded, and only triangles with higher internal quality are retained for subsequent calculations. Through the above processing, the spatial distribution characteristics of the matching feature points are completely mapped to a geometrically constrained triangular mesh, providing structured spatial units for step 103.

[0065] 103: By calculating the minimum bounding square of each triangle, the adaptive subset size of each local region is automatically determined;

[0066] The core inventive concept of this step is that the side length of the circumscribed square of the triangle is a geometric measure of the feature distribution density of the local area. This step establishes for the first time a quantitative mapping relationship of "feature point distribution density → geometric measure → subset size", so that the subset size can adaptively respond to changes in local speckle quality and deformation complexity.

[0067] The physical significance of this mapping relationship lies in the fact that, in this embodiment of the invention, instead of relying on human experience or fixed thresholds, the size of the subset is directly defined as the geometric response of the speckle local information entropy. Large subsets are automatically configured for sparse feature point regions to ensure matching robustness, while small subsets are automatically configured for dense feature point regions to ensure spatial resolution, achieving an adaptive optimal balance between measurement accuracy and robustness.

[0068] 104: Based on the preliminary displacement obtained by feature point matching, the IC-GN method is used to traverse and calculate the displacement of the entire field.

[0069] The core inventive concept of this step is that the triangular mesh constructed in step 102 is not only used for subset configuration, but also naturally provides personalized initial displacement values ​​for each calculation point, which improves the IC-GN iterative optimization from a globally unified initial value to an adaptive prediction initial value, thereby accelerating convergence and improving accuracy.

[0070] After completing the adaptive subset configuration in steps 101 to 103, this step proceeds to the precise calculation stage of the displacement field. Unlike the existing DIC method, which typically uses zero displacement or the displacement of the previous adjacent point as the initial value for all calculation points, this method utilizes the triangulation geometry to provide personalized initial displacement predictions for each calculation point.

[0071] In summary, the embodiments of the present invention achieve adaptive configuration of subset size, thereby improving the measurement accuracy and robustness of complex deformation fields without manual intervention.

[0072] Steps 101 to 104 above constitute a complete closed-loop control strategy of "perception-modeling-decision-execution":

[0073] Step 101 (Perception): Extract feature points from the speckle image, obtain highly reliable matching point pairs through the improved GMS algorithm, and perceive the speckle quality and deformation distribution in the local area of ​​the image.

[0074] Step 102 (Modeling): Discrete feature points are organized into a spatial structure with geometric constraints through Delaunay triangulation, and a computational unit under the continuity assumption is established.

[0075] Step 103 (Decision): Based on the geometric measure of the circumscribed square of the triangle, quantitatively determine the optimal subset size for each local region.

[0076] Step 104 (Execution): Drive IC-GN iteration with adaptive initial values ​​provided by triangulation to perform high-precision full-field displacement calculation.

[0077] Existing DIC methods treat subset size only as a user-specified input parameter, while this invention redefines subset size as the output response of local speckle quality. This fundamental difference allows the method to operate without manual intervention throughout the entire process, and the subset size automatically adapts to local changes in the speckle pattern, thus solving the measurement failure problem of traditional DIC methods' subset configuration strategies in scenarios with non-uniform speckle quality.

[0078] Example 2

[0079] The scheme in Example 1 will be further described below with specific calculation formulas and examples:

[0080] I. Feature point extraction and mismatch point elimination

[0081] The core inventive concept of this section lies in transforming the feature matching problem into a hypothesis testing problem based on neighborhood statistical consistency. The spatial distribution consistency of matching point pairs is used as the criterion for judging their correctness, thereby achieving highly reliable mismatch removal without relying on prior deformation models. In digital image correlation methods, image feature points and speckle quality are interdependent. High-quality speckle improves matching accuracy by enhancing the uniqueness and stability of feature points, while effective extraction and matching of feature points are the core basis for speckle quality assessment. Therefore, this step first extracts feature points from the speckle image and eliminates mismatched points to obtain reliable matching feature point pairs.

[0082] Step 201: Obtain the reference speckle image and the target speckle image;

[0083] The reference speckle image is the initial speckle image captured by the camera before the specimen to be measured is loaded and deformed, while the target speckle image is the speckle image captured at different times after the specimen to be measured is loaded and deformed. The specific acquisition steps are well known to those skilled in the art, and will not be described in detail in this embodiment of the invention.

[0084] Step 202: Extract feature points from the reference speckle image and the target speckle image using the SIFT algorithm;

[0085] By utilizing the Gaussian convolution kernel of the SIFT algorithm, a scale space is efficiently constructed, rapidly generating a multi-scale image space with scale invariance. This achieves fast coverage across the entire scale range, thereby extracting more scale-robust feature points. In a 480×480 pixel speckle image, more than 4000 feature points can be extracted.

[0086] Step 203: Eliminate mismatched points based on the improved margin grid motion statistics GMS algorithm to obtain matching feature point pairs;

[0087] The improved GMS algorithm improves upon the original GMS algorithm in the following three aspects:

[0088] (1) Statistical edge distance grid: The optimal grid edge distance is calculated using a statistical algorithm, and the edge feature points are assigned to adjacent grids to solve the shortcomings of edge feature point processing;

[0089] The GMS algorithm, through rigorous mathematical proof and experimental demonstration, shows that there are several pairs of features that match in the adjacent regions of a correct match, while the number of features that match in the adjacent regions of a wrong match is small or even non-existent. The GMS algorithm distinguishes between correct and wrong matches by counting the number of features that match in the adjacent pixel regions of the coarse matching point set.

[0090] Reference speckle image The neighborhood representation of the matching feature point set is as follows Target speckle image The neighborhood representation of the matching feature point set is as follows ,in , For the i-th pair of matching points ,statistics Neighborhood Feature point set in This is the neighborhood and feature count of the i-th matching point pair. ;statistics Neighborhood Feature point set in The matching feature point set is located in Neighborhood Number of features in For ease of description, This is called the feature neighborhood score, and the calculation process is as follows:

[0091]

[0092]

[0093] In the formula, for Does the feature matched by the k-th feature point lie in the... In the state, if If the feature is within the specified range, it is evaluated as 1; otherwise, it is evaluated as 0. Then, the feature neighborhood scores for correctly matched features are calculated. Feature neighborhood scores of mismatches Statistical analysis was conducted based on... and Setting a score threshold based on proportional relationships To determine Whether it is a correct match. Finally, iterate through all matching feature point sets to eliminate incorrect matches, obtaining a refined matching feature point set. .

[0094] (2) Improved five-grid feature score statistics method: By changing the feature score statistics method of adjacent symmetrical 8 grids to the five-grid feature score statistics method, the erroneous matching points in the initial matching are eliminated more quickly, and the matching accuracy is improved;

[0095] The GMS algorithm divides the matching image into a non-overlapping grid (G=P*Q). For feature points located at the grid edges, the GMS algorithm calculates the feature score for each grid and also calculates the feature scores of the eight adjacent symmetrical grids. Because the grid is symmetrical, calculating the feature scores of eight adjacent grids for a single grid would increase unnecessary computation. Therefore, to accelerate the removal of incorrect matching pairs and further increase distinguishability, this invention proposes an improved five-grid feature score calculation method. Specifically, considering the symmetry of the grid distribution, only the feature scores of the four adjacent and symmetrical grids are calculated. The sum of these five grid feature scores is called the five-grid feature score S, defined by the following formula: In the formula The feature score is the j-th grid within the five-grid containing the i-th grid.

[0096] To avoid the impact of image rotation on statistics, the five-grid is rotated clockwise. The rotation does not change the grid's position, but adjacent grids move clockwise. After the fourth rotation, the grid feature distribution will be consistent with the initial situation. Therefore, we only need to calculate the maximum grid feature score after three rotations. Compared to the original GMS method, which requires seven rotations in a 3x3 grid, this method reduces the computational load by more than half.

[0097]

[0098] In the formula, Let j be the feature score of the j-th grid after rotating the five-square grid containing the i-th grid k times. Then, calculate the mean of the number of coarse matching features of the grids within the current five-square grid:

[0099]

[0100] In the formula, This represents the number of coarse-matching features within the j-th grid in the five-grid containing the i-th grid. When the grid feature score... Greater than the grid feature score threshold At that time, the judgment A correct match is indicated by a non-correct match, otherwise an incorrect match. Grid feature score threshold. Calculation formula:

[0101]

[0102]

[0103] In the formula , , This is the threshold coefficient. The coefficient is generally set relatively high to address situations where there are few correct matches and many incorrect matches within the grid. The coefficient is the weight of the mean of the number of features, and is generally set to be close to 1. This setting is for cases where the input to the logarithmic function with base e is small, and is generally set to a smaller value.

[0104] (3) Multi-scale and multi-rotation strategy: improve the robustness of the algorithm in rotation and scale change scenarios.

[0105] In the GMS algorithm, the values ​​P and Q for the number of grids are empirical values. For images with inconsistent aspect ratios, rectangular grids are generated, resulting in uneven distribution of coarse matching features within each grid. This embodiment of the invention uses the aspect ratio of the input image as a constraint and initializes the P and Q values ​​with empirical values ​​E, making the divided grids approximately regular squares.

[0106]

[0107] In the formula, and These represent the width and height of the input speckle image I, respectively. In reality, many features are located on the grid edges. To accommodate this, this embodiment of the invention shifts the grid pattern by half a cell width in both the x and y directions, and repeats the algorithm three times.

[0108] The improved algorithm increases recall by approximately 10%, improves real-time performance by 30%, and reduces runtime by approximately 30% compared to the traditional ORB+RANSAC method. For image regions with few feature points, bicubic interpolation can be used for feature enhancement. In a 480×480 pixel speckle image, the improved GMS algorithm can achieve over 3700 matching feature point pairs, with a matching accuracy exceeding 91%.

[0109] II. Determining Adaptive Subsets

[0110] This step automatically determines the optimal subset size of each local region based on the spatial distribution of matching feature point pairs through geometric partitioning.

[0111] Step 204: As Figure 2 As shown, a Delaunay triangular mesh is constructed based on matching feature point pairs to eliminate narrow triangles at the edges;

[0112] Delaunay triangulation is a classic method for constructing triangular meshes from a planar point set. Its core objective is to generate optimal triangulation. Among all possible triangulations, Delaunay triangulation produces triangles with the largest minimum angle, avoiding elongated triangles. It also possesses uniqueness when no four points in the point set are concyclic, exhibiting low time complexity and suitability for large-scale point sets. Since elongated triangles are prone to appearing at edges, making them unsuitable for subsequent calculations, they are eliminated, retaining only high-quality triangles from the internal regions.

[0113] Step 205: As Figure 3 As shown, calculate the minimum bounding square of each triangle, and take the integer part of the side length of the minimum bounding square as the size of the adaptive subset at the center point of the triangle;

[0114] For each retained triangle, its minimum circumscribed square is obtained through either a rotation method or an analytical method. The rotation method can use gradient descent or bisection to determine the rotation angle, or it can obtain the optimal solution by traversing all angles. The circumscribed square around the center point of each triangle, defined by three feature points, has a side length that reflects the distribution density of feature points in that local region: the circumscribed square is smaller in regions with dense feature points and larger in regions with sparse feature points. Therefore, the odd value of the circumscribed square determined by the triangle, rounded up, is used as the size of the subset at the center point.

[0115] Suppose that the side length of the smallest circumscribed square of the subset is L, obtained through statistical analysis of the feature point distribution. Then, the size S of the subset is set to the smallest odd number not less than L, i.e.:

[0116]

[0117] in, This indicates rounding up, and Smax is the preset maximum subset size. Typically, Smax is set to 51 pixels, corresponding to a subset size range of 11×11 pixels to 51×51 pixels.

[0118] III. Determination of Total Displacement

[0119] After completing the adaptive subset configuration, the precise calculation of the displacement field begins.

[0120] Step 206: As Figure 4 As shown, the initial displacement obtained by feature point matching is used as the initial value;

[0121] The average displacement of the matching feature point pairs at the three vertices of each triangle is used as the initial displacement estimate for that triangular region, thus providing a reliable initial value for subsequent IC-GN iterative optimization. To reduce random uncertainty, the average of the three maximum and minimum displacements is taken as the maximum and minimum values, respectively. Based on this displacement, the corresponding subset size is determined between 11×11 and 51×51 pixels.

[0122] Step 207: Calculate the full-field displacement using the IC-GN method to generate a uniformly distributed displacement field.

[0123] Among them, the IC-GN method is widely used in DIC for accurate calculation of full-field displacement, and has the advantages of high efficiency, stability and fast convergence speed.

[0124] The iterative calculation formula for the IC-GN method is as follows:

[0125]

[0126] Where k is the current iteration number, p is the shape function parameter vector, and the embodiment of this invention uses a first-order shape function, with corresponding p= , It is the Hessian matrix. It is the gradient of the error function. This refers to the parameter update amount. After obtaining the sub-pixel precision displacement of each calculation point through iterative optimization, all points to be calculated on the reference image are traversed to finally obtain the displacement field and strain field of the entire field.

[0127] In the correlation matching calculation process, the zero-mean normalized squared distance cross-correlation function (ZNSSD) is used as the correlation criterion. This correlation function is insensitive to linear changes in the gray values ​​of the target subset (such as uneven illumination or intensity changes), and has good resistance to illumination changes. At the same time, subtracting the average value from the image gray values ​​in the ZNSSD correlation function can better highlight the difference between the main peak and the secondary peak, making it easier to obtain the peak value in the correlation calculation. Furthermore, the value range is narrowed after normalization, making the statistics more convenient.

[0128] A first-order shape function is used to describe the correspondence between the internal point positions of the subset before and after deformation. This first-order shape function, based on rigid body translation, considers simple deformations of the image subset (such as rigid body translation, rotation, stretching, compression, and dilation), and has six degrees of freedom. It can describe the deformation state of the subset relatively accurately, and can meet the requirements in terms of computational accuracy and speed.

[0129] Example 3

[0130] The experimental verification process of the method of the present invention will be described in detail below with specific experimental data and comparative analysis.

[0131] I. Displacement Measurement Accuracy Evaluation Experiment

[0132] To verify the displacement measurement accuracy of the method of the present invention, see [link to relevant documentation]. Figure 5 This experiment created a 1024×256 pixel sinusoidal displacement simulation speckle map. In this simulation image, the frequency decreases linearly from the left side (0.05) to the right side (0), the displacement calculation component v is a sinusoidal displacement with an amplitude of 1 pixel, and the horizontal displacement component u is set to 0. Bicubic interpolation was used during image generation to ensure image quality.

[0133] This experiment systematically compared the method of this invention with methods using five different fixed subset sizes (11×11, 21×21, 31×31, 41×41, and 51×51 pixels). The experimental results were analyzed from two dimensions: correlation and displacement field.

[0134] Regarding correlation, ZNCC was used as the evaluation metric. The ZNCC distribution under different subset sizes shows that as the fixed subset increases, the ZNCC correlation on both sides of the image gradually decreases, typically indicating decorrelation when the correlation coefficient is less than 0.8. Specifically, while small subsets (11×11, 21×21 pixels) exhibit high correlation within the entire image region, this is a spurious high correlation due to insufficient features. As the subset size increases to 41×41 and 51×51 pixels, rapidly changing areas on both sides of the image exhibit numerous low-correlation regions due to violation of the assumption of uniform displacement within the window. In contrast, the correlation within the measurement region obtained by the method of this invention did not show a significant decrease, with the average ZNCC remaining above 0.98 across the entire region, achieving highly uniform correlation and effectively solving the problem of large-scale matching failure.

[0135] Regarding displacement field distribution, the comparative results show that the traditional fixed subset strategy suffers from an inherent dilemma: high noise at small scales and low accuracy at large scales. While small subsets (11×11, 21×21 pixels) produce fine displacement fringes, noise peaks are significant, especially in the central region of the image where displacement changes are gradual, leading to drastic fluctuations in measurement error and considerable uncertainty. Large subsets (41×41, 51×51 pixels), while smoothing the noise, result in loss of detail and blurred transitions in high-frequency regions on both sides of the image, significantly increasing the deviation from theoretical values ​​and drastically reducing the fitting area of ​​the measured values. The method of this invention achieves low-noise smoothing consistency while maintaining high-resolution fringe detail. Furthermore, the measured values ​​in the image edge region remain within 0.5 pixels, and the error fluctuation in the central region is also very small. Random errors are effectively controlled, and the stability, accuracy, and reliability of the measurement system all reach a high level, verifying the stability and robustness of the method in measuring complex deformation regions.

[0136] II. Performance Comparison Experiment of Feature Matching and Initial Value Estimation

[0137] To verify the robustness of the method of the present invention in feature matching and mismatch point filtering, the method of the present invention was systematically compared with the classical SIFT method and SURF method based on RANSAC. (See [link to relevant documentation]). Figure 6 .

[0138] Experimental setup: A region of interest was selected in a 512×512 pixel simulated reference speckle image, with a subset size of 31×31 pixels. The reference image was rotated 10° clockwise to create a deformed image, simulating rotational deformation scenarios that may occur in actual measurements.

[0139] This experiment used five quantitative indicators—number of matched pairs, matching rate, average measurement error, average correlation coefficient, and average running time—to objectively evaluate the matching results. Each experiment was repeated five times, and the average results were taken. The comparison results are shown in Table 1.

[0140] Table 1 Comparison results of different feature matching methods

[0141]

[0142] The following analytical conclusions can be drawn from the comparative data in Table 1: (1) In terms of the number of matching point pairs, the SIFT algorithm extracts the most matching point pairs (4738 pairs), but the matching rate is the lowest (80.56%) and the correlation is the worst (0.86). This is because the RANSAC method has randomness when filtering out false matches, and it is difficult to guarantee the stability of the matching quality. The SURF method has fewer matching point pairs (3924 pairs), and the matching rate (90.25%) and correlation (0.89) are improved. The method of this invention maintains a matching point pair number comparable to the SIFT algorithm (4592 pairs), while improving the matching rate to 96.78% and the correlation coefficient to 0.95, indicating that the improved GMS algorithm is more effective and reliable in eliminating false matches. (2) Regarding measurement error, the average measurement errors of SIFT+DIC and SURF+DIC methods are 0.2186° and 0.1859°, respectively, while the average measurement error of the method of the present invention is only 0.0515°, which is a reduction of 76.4% and 72.3%, respectively, proving that the method of the present invention has a significant advantage in measurement accuracy. (3) Regarding running time, the method of the present invention (12.93s) is slightly longer than SIFT+DIC (11.78s) and SURF+DIC (10.41s). This is because the improved GMS algorithm adopts a multi-scale and multi-rotation strategy to improve matching quality, but it is still within an acceptable range, and the significant improvement in measurement accuracy and reliability is enough to compensate for the slight increase in computation time.

[0143] From the perspective of spatial distribution error of the displacement field: synthesizing the arithmetic square roots of the displacements in the X and Y directions into a comprehensive displacement field can more intuitively reflect the overall distribution characteristics of the displacement field. The deviations between the displacement fields obtained by the SIFT+DIC and SURF+DIC methods and the theoretical displacements are both within ±0.2 pixels, while the measurement deviation of the method of this invention is within ±0.05 pixels, with an error range only one-quarter of that of the comparative methods. This result fully verifies the significant advantages of the method of this invention in improving the accuracy and matching reliability of displacement field measurement.

[0144] III. Experimental Verification in Non-uniform Tensile Fields

[0145] To further verify the effectiveness of the method of this invention in practical engineering applications, it was applied to a uniaxial tensile test after laser welding of aluminum alloys. During laser welding, high temperatures cause changes in the microstructure of the alloy material, thereby deteriorating its properties, especially the mechanical properties of the heat-affected zone near the weld. This experiment analyzes the deformation information of the welded area during the tensile process by measuring the full-field displacement deformation in the area near the weld after laser welding of aluminum alloys.

[0146] Experimental conditions: The welding material used was an aluminum alloy sheet with magnesium and silicon as the main alloying elements. Tensile tests were conducted under conditions where there was no obvious necking of the sample after laser welding. 120 speckle images were acquired within the first 60 seconds, with an image size of 1024×768 pixels. The weld location and the area to be measured were marked in the acquired speckle images, with the weld location indicated by the dashed box and the area to be calculated within the rectangular box. (See [reference]). Figure 7 .

[0147] Experimental procedure and parameter settings: The 10th image at the 5th second was selected as the reference image, and the tensile images at the 15th second (30th frame), 25th second (50th frame), 35th second (70th frame), 45th second (90th frame), and 55th second (110th frame) were used as the target images to be measured. A total of five sets of deformation information of the area to be calculated during the tensile process of aluminum alloy were analyzed.

[0148] The method of this invention was used to perform DIC-related calculations to obtain the corresponding displacement field. The experimental results show that: (1) The displacement field gradually increases with time, and the gradient change near the weld is significantly higher than that of the unwelded aluminum alloy substrate area. The displacement field in the X direction increases from about 0.3 pixels at 15 seconds to about 3 pixels at 55 seconds, and the displacement field in the Y direction increases from about 0.5 pixels at 15 seconds to about 4 pixels at 55 seconds. The displacement change is significant and directly reflects the gradual stretching and elongation process of the aluminum alloy material. (2) The displacement gradient at the weld is significantly higher than that near the substrate, indicating that the deformation mainly occurs in the area near the weld, which is consistent with the mechanical properties of local weakening of the material caused by laser welding. (3) During the stretching process of the aluminum alloy welded specimen, there is a clear relationship between the displacement field distribution and the weld distance: the X-direction displacement is approximately symmetrically distributed with respect to the weld line, and the Y-direction displacement is approximately anti-symmetrically distributed with respect to the midpoint of the weld.

[0149] The results of the above-mentioned tensile test on laser welding of aluminum alloys verify the effectiveness and engineering application value of the method of the present invention in the measurement of actual non-uniform and complex deformation fields. The adaptive subset configuration strategy can automatically adjust the subset size according to the non-uniformity of speckle feature distribution near the weld. In areas with large deformation gradients at the weld, a smaller subset size is used to capture local deformation details, while a larger subset size is used in the substrate area far from the weld to ensure measurement stability and noise resistance, thereby achieving high-precision displacement measurement across the entire field.

[0150] Example 4

[0151] An adaptive subset configuration apparatus based on image feature matching in DIC measurement, the apparatus comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the apparatus to perform the following method steps in Embodiment 1:

[0152] The SIFT algorithm is used to extract feature points from speckle images, and an improved GMS algorithm is used to eliminate mismatched points to obtain matching feature point pairs.

[0153] Delaunay triangular meshes are constructed based on matching feature point pairs, and regions are divided using triangles as the basic units. The size of the adaptive subset of each local region is determined by calculating the minimum circumscribed square of each triangle.

[0154] Based on the preliminary displacement obtained by feature point matching, the IC-GN method is used to traverse and calculate the displacement of the entire field.

[0155] Among them, the improved GMS algorithm is used to eliminate mismatched points to obtain matching feature point pairs as follows:

[0156] A statistical algorithm is used to calculate the optimal grid edge distance and assign edge feature points to adjacent grids, thus addressing the shortcomings in edge feature point processing.

[0157] The method of statistically analyzing the feature scores of eight adjacent symmetrical grids was changed to a five-grid feature score statistical method to eliminate erroneous matching points in the initial matching.

[0158] Using the aspect ratio of the input image as a constraint, the values ​​of P and Q are initialized with empirical values ​​E, so that the divided grid is close to a regular square.

[0159] The statistical method for the five-square grid feature scores is as follows:

[0160] The sum of the five grid feature scores is called the five-grid feature score S: , Let the feature score of the j-th grid within the five-square grid containing the i-th grid be calculated. Then, rotate the five-square grid clockwise and calculate the maximum feature score of the grid after three rotations. ,

[0161]

[0162] In the formula, To calculate the feature score of the j-th grid after rotating the i-th grid k times, we calculate the average number of coarse matching features among the grids in the current five-grid:

[0163]

[0164] In the formula, The number of coarse matching features within the j-th grid in the five-grid containing the i-th grid, when the grid feature score Greater than the grid feature score threshold At that time, the judgment A correct match is defined as a match where the condition is not met, and an incorrect match is defined as a grid feature score threshold. Calculation formula:

[0165]

[0166]

[0167] In the formula , , For threshold coefficient, Let F represent the i-th pair of matching points, where F represents an incorrect match and T represents a correct match.

[0168] The values ​​of P and Q are initialized using the empirical value E:

[0169]

[0170] In the formula, and Given the width and height of the input speckle image I, respectively, shift the grid pattern by half a cell width in both the x and y directions and the x and y directions, and then repeat the algorithm 3 times.

[0171] Specifically, the size of the adaptive subset of each local region is determined by calculating the minimum circumscribed square of each triangle:

[0172] The minimum circumscribed square of the triangle can be obtained by rotation, and the rotation angle can be determined by gradient descent or bisection, or the optimal solution can be obtained by traversing all angles.

[0173] Based on the optimal solution, the minimum circumscribed square side length of the subset is L obtained through feature point distribution statistics. Therefore, the subset size S is set to the smallest odd number not less than L, i.e.:

[0174]

[0175] in, Indicates rounding up. It is the preset maximum subset size.

[0176] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0177] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0178] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0179] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0180] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0181] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0182] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0183] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0184] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0185] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive subset configuration method based on image feature matching in DIC measurement, characterized in that, The method includes: The SIFT algorithm is used to extract feature points from speckle images, and an improved GMS algorithm is used to eliminate mismatched points to obtain matching feature point pairs. Delaunay triangular meshes are constructed based on matching feature point pairs, and regions are divided using triangles as the basic units. The size of the adaptive subset of each local region is determined by calculating the minimum circumscribed square of each triangle. Based on the preliminary displacement obtained by feature point matching, the IC-GN method is used to traverse and calculate the displacement of the entire field.

2. The adaptive subset configuration method based on image feature matching in DIC measurement according to claim 1, characterized in that, The improved GMS algorithm is used to eliminate mismatched points to obtain matching feature point pairs as follows: A statistical algorithm is used to calculate the optimal grid edge distance and assign edge feature points to adjacent grids, thus addressing the shortcomings in edge feature point processing. The method of statistically analyzing the feature scores of eight adjacent symmetrical grids was changed to a five-grid feature score statistical method to eliminate erroneous matching points in the initial matching. Using the aspect ratio of the input image as a constraint, the values ​​of P and Q are initialized with empirical values ​​E, so that the divided grid is close to a regular square.

3. The adaptive subset configuration method based on image feature matching in DIC measurement according to claim 2, characterized in that, The method for calculating the feature scores of the five-square grid is as follows: The sum of the five grid feature scores is called the five-grid feature score S: , The feature score of the j-th grid within the five-grid containing the i-th grid. Rotate the five-square grid clockwise and calculate the maximum grid feature score after three rotations. , In the formula, To calculate the feature score of the j-th grid after rotating the i-th grid k times, we calculate the average number of coarse matching features among the grids in the current five-grid: In the formula, The number of coarse matching features within the j-th grid in the five-grid containing the i-th grid, when the grid feature score Greater than the grid feature score threshold At that time, the judgment A correct match is defined as a match where the condition is not met, and an incorrect match is defined as a grid feature score threshold. Calculation formula: In the formula , , For threshold coefficient, Let F represent the i-th pair of matching points, where F represents an incorrect match and T represents a correct match.

4. The adaptive subset configuration method based on image feature matching in DIC measurement according to claim 2, characterized in that, Initialize P and Q values ​​using the empirical value E: In the formula, and Given the width and height of the input speckle image I, respectively, shift the grid pattern by half a cell width in both the x and y directions and the x and y directions, and then repeat the algorithm 3 times.

5. The adaptive subset configuration method based on image feature matching in DIC measurement according to claim 2, characterized in that, The adaptive subset size of each local region is determined by calculating the minimum circumscribed square of each triangle: The minimum circumscribed square of the triangle can be obtained by rotation, and the rotation angle can be determined by gradient descent or bisection, or the optimal solution can be obtained by traversing all angles. Based on the optimal solution, the minimum circumscribed square side length of the subset is L obtained through feature point distribution statistics. Therefore, the subset size S is set to the smallest odd number not less than L, i.e.: in, Indicates rounding up. It is the preset maximum subset size.

6. An adaptive subset configuration device based on image feature matching in DIC measurement, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-5.

Images