High-dynamic surface single-frame measurement method based on multi-view digital speckle correlation method

Abstract

A high-dynamic surface single-frame measurement method based on a multi-view digital speckle correlation method. The measurement method comprises: on the basis of a four-view camera and a lithographic speckle projection module, constructing a speckle projection collection system, and on the basis of the speckle projection collection system, collecting a single-frame image of a workpiece under test; on the basis of a multi-scale robust multi-view matching algorithm, processing the single-frame image of the workpiece under test, so as to obtain an intermediate depth map and pixel-wise spatial oblique planes; on the basis of a digital speckle correlation method, performing iterative calculation on intermediate parameters of the pixel-wise spatial oblique planes, so as to obtain final parameters of the spatial oblique planes; and on the basis of the final parameters of the spatial oblique planes and a left-right consistency cross-check, fusing final depth maps, so as to obtain final point cloud data. The high-dynamic surface single-frame measurement method based on a multi-view digital speckle correlation method can use a single-frame image to rapidly re-construct a workpiece having a high-dynamic surface in a field of view, so as to obtain point cloud with high integrity and high precision, and the measurement precision thereof can reach 0.03-0.07 mm.

Description

A High Dynamic Surface Single-Frame Measurement Method Based on Multi-View Digital Speckle Correlation Technical Field

[0001] This invention belongs to the field of multi-view digital speckle technology, and particularly relates to a high dynamic surface single-frame measurement method based on multi-view digital speckle correlation. Background Technology

[0002] In the manufacturing of sheet metal parts for aerospace, automotive, and other fields, accurate morphological measurement is essential. Despite the many advantages of structured light technology, accurately reconstructing the 3D point cloud of high dynamic range (HDR) surfaces remains a challenging problem. HDR surfaces exhibit both specular and diffuse reflection. Specular reflection occurs only in areas of significantly higher surface reflectivity. Due to insufficient camera measurement range, structured light patterns reflected from high-reflectivity locations are overly bright, easily leading to intensity saturation, while those from low-reflectivity locations are very dark, resulting in a low signal-to-noise ratio. Intensity saturation and low signal-to-noise ratio severely impact the accuracy of 3D reconstruction.

[0003] Many researchers have proposed multiple exposure techniques, which can obtain high-quality structured light patterns by varying the exposure time. However, the effectiveness of these techniques is highly dependent on the accuracy of identifying saturated pixels in the image. To address this issue, many researchers have proposed adaptive fringe projection techniques, which typically obtain high-quality images by adjusting the maximum input grayscale. However, this method is based on the assumption that the measured surface is smooth, therefore it cannot be used to measure geometrically complex objects with high dynamic range surfaces.

[0004] As mentioned above, specular reflections are only distributed in a specific area of ​​the camera's viewpoint. Therefore, some researchers increase the number of viewpoints to 2-3, which can change the camera's viewpoint to avoid capturing specular reflection areas in the image, for example, by measuring the object from at least three different directions or using a translation stage to adjust the camera's viewpoint. However, this technique cannot guarantee that all saturated areas in one camera view will be completely unsaturated in another camera viewpoint, and using a translation stage significantly increases measurement time. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a high-dynamic surface single-frame measurement method based on multi-view digital speckle correlation, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a high-dynamic surface single-frame measurement method based on multi-view digital speckle correlation, comprising:

[0007] A speckle projection acquisition system is constructed based on a quad-camera and a photolithographic speckle projection module, and a single-frame image of the workpiece under test is acquired based on the speckle projection acquisition system.

[0008] The single-frame image of the workpiece under test is processed based on a multi-scale robust multi-view matching algorithm to obtain an intermediate depth map and a pixel-by-pixel spatial oblique plane.

[0009] The intermediate parameters of the spatial oblique plane are iteratively calculated based on the digital speckle correlation method to obtain the final parameters of the spatial oblique plane.

[0010] The final depth map is fused based on the final parameters of the spatial oblique plane and the left-right consistency superposition test to obtain the final point cloud data.

[0011] Preferably, the process of constructing a speckle projection acquisition system based on a multi-view camera includes:

[0012] The four cameras around the perimeter are labeled, and a speckle projection module is installed at the symmetrical center of the cameras.

[0013] The labeled quad-camera is then subjected to single-target and dual-target calibration. The intrinsic and extrinsic parameters of the quad-camera are then optimized using the bundle adjustment method to generate the speckle projection acquisition system.

[0014] Preferably, the process after acquiring a single frame image of the workpiece based on the speckle projection acquisition system further includes:

[0015] Distortion correction processing is performed on a single frame image of the workpiece under test to obtain a distortion-corrected image;

[0016] The distortion-corrected image is processed using a median filtering algorithm to obtain a filtered and denoised image.

[0017] The filtered image is processed using a Gaussian filtering algorithm to obtain a preprocessed single-frame image.

[0018] Preferably, the process of obtaining the final depth map and the pixel-by-pixel spatial oblique plane includes:

[0019] The single-frame image of the workpiece under test is downsampled twice, and the spatial oblique plane is initialized within the assumed range of the depth to be measured and the imaging angle.

[0020] One image from a single frame of the workpiece under test is set as a reference image (each image is used as a reference image and the following algorithm is applied). The remaining images are source images. The binocular matching cost between the local matching box of the reference pixel and the corresponding matching box of each source image is calculated. The multi-view matching cost is obtained by weighting the binocular matching cost.

[0021] Based on the multi-view matching cost, the spatial plane assumption is iterated, and then the spatial plane is refined to obtain the final depth map corresponding to the reference image after two downsamplings and the pixel-by-pixel set of spatial oblique planes.

[0022] The reference image's corresponding depth map and pixel-by-pixel spatial oblique plane set are subjected to repeated view selection, propagation, and refinement to generate the coarsest-scale depth map.

[0023] The coarsest-scale depth map is refined using a multi-scale consistency matching cost. Then, the refined coarsest-scale depth map is upsampled and followed by the depth map refinement module twice to obtain the intermediate depth map and the pixel-wise spatial oblique plane.

[0024] Preferably, the expression for the multi-view matching cost is:

[0025] in This represents the weight of the reference pixel for each source image, m. i,j The stereo matching cost is calculated between the reference box at the reference pixel and the corresponding box in the source image.

[0026] Preferably, the process of obtaining the final parameters of the spatial inclined plane includes:

[0027] Based on the obtained pixel-by-pixel intermediate spatial oblique plane, the iterative initial box of the digital speckle correlation method is obtained, and the parameters of the iterative initial box are calculated to obtain the pixel-by-pixel iterative initial value.

[0028] The pixel-by-pixel initial value is iteratively calculated using the digital speckle correlation method, and the final parameters of the spatial oblique plane are determined based on the hyperparameter threshold.

[0029] Preferably, the process of iteratively calculating the pixel-by-pixel initial value using the digital speckle correlation method further includes: applying the parameter increment to the reference sub-region using the ZNSSD correlation function before each iteration;

[0030] The expression for the ZNSSD correlation function in Δp form is:

[0031] in, It is a Hessian matrix. in Here is the gray-level gradient matrix of the reference sub-region in the x, y directions. Let be the Jacobian matrix of shape functions.

[0032] Preferably, the process of obtaining the final point cloud data further includes:

[0033] The two-dimensional coordinates in the final depth map are reprojected back to the world coordinate system and then projected onto the remaining depth maps to obtain the projected subpixel coordinates.

[0034] The subpixel coordinates after projection are reprojected into the world coordinate system, and the spatial coordinates after reprojection into the world coordinate system are projected back into the reference depth map to obtain the subpixel coordinates.

[0035] The Euclidean distance between the original coordinates and the sub-pixel coordinates yields the final consistency distance;

[0036] The final point cloud data is obtained based on the final consistency distance and the threshold of the angle between the spatial oblique plane.

[0037] Compared with the prior art, the present invention has the following advantages and technical effects:

[0038] This invention provides non-contact measurement with high precision, high speed, and a large field of view. It can quickly reconstruct workpieces with highly dynamic surfaces within the field of view using a single frame, obtaining point clouds with high integrity and high precision. Its measurement accuracy can reach 0.03-0.07mm. Attached Figure Description

[0039] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0040] Figure 1 is a schematic diagram of a four-eye speckle projection acquisition system according to an embodiment of the present invention;

[0041] Figure 2 is a schematic diagram of two different random initialization assumptions in an embodiment of the present invention, wherein (a) is a schematic diagram of the leading-edge parallel plane with integer value disparity, and (b) is a schematic diagram of the inclined plane with continuous normal vector and depth used in the present invention.

[0042] Figure 3 is a diagram of the red and black checkerboard iterative mode according to an embodiment of the present invention;

[0043] Figure 4 is a schematic diagram of the reconstructed three-dimensional point cloud of the standard part according to an embodiment of the present invention. Detailed Implementation

[0044] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0046] Example 1

[0047] This embodiment provides a high-dynamic surface single-frame measurement method based on multi-view digital speckle correlation, including:

[0048] A speckle projection acquisition system is constructed based on a quad-camera and a photolithographic speckle projection module, and a single-frame image of the workpiece under test is acquired based on the speckle projection acquisition system.

[0049] The single-frame image of the workpiece under test is processed based on a multi-scale robust multi-view matching algorithm to obtain an intermediate depth map and a pixel-by-pixel spatial oblique plane.

[0050] The intermediate parameters of the spatial oblique plane are iteratively calculated based on the digital speckle correlation method to obtain the final parameters of the spatial oblique plane.

[0051] The final depth map is fused based on the final parameters of the spatial oblique plane and the left-right consistency superposition test to obtain the final point cloud data.

[0052] Specifically:

[0053] The exposure time of the multi-view camera was adjusted to a reasonable value, and a carefully designed four-view speckle projection acquisition system was used to acquire single-frame images of the workpiece under test. The pixel-by-pixel spatial coarse intermediate oblique plane corresponding to each camera on the workpiece was obtained through edge-guided multi-scale robust multi-view matching. The pixel-by-pixel spatial oblique plane parameters were mapped to the source image using a homography matrix to initialize the initial iteration parameters of the digital speckle correlation method. The digital speckle correlation method was used for 1-2 iterations, and the final parameters of the spatial oblique plane were determined based on hyperparameter thresholds. Finally, depth map fusion was performed through left-right consistency superposition checks to obtain the final point cloud for accuracy evaluation and analysis.

[0054] Further optimization of the scheme, the process of constructing a speckle projection acquisition system based on a multi-view camera includes:

[0055] The four cameras around the perimeter are labeled, and a speckle projection module is installed at the symmetrical center of the cameras.

[0056] The labeled quad-camera is then subjected to single-target and dual-target calibration. The intrinsic and extrinsic parameters of the quad-camera are then optimized using the bundle adjustment method to generate the speckle projection acquisition system.

[0057] Further optimization of the scheme includes the following process after acquiring a single-frame image of the workpiece under test based on the speckle projection acquisition system:

[0058] Distortion correction processing is performed on a single frame image of the workpiece under test to obtain a distortion-corrected image;

[0059] The distortion-corrected image is processed using a median filtering algorithm to obtain a filtered and denoised image.

[0060] The filtered image is processed using a Gaussian filtering algorithm to obtain a preprocessed single-frame image.

[0061] Specifically:

[0062] The first step is to label the surrounding four cameras, with the speckle projection module at the center, as shown in Figure 1. First, perform single-target calibration on each camera individually; then, perform two-target calibration on cameras 2, 3, and 4 respectively with camera 1; finally, use bundle adjustment to optimize the intrinsic and extrinsic parameters of the multi-camera system.

[0063] The second step is to adjust the exposure time of the multi-view camera to a reasonable value, adjust the focal length of the speckle projection module to make the image clear, and use the carefully designed four-view speckle projection acquisition system to acquire a single frame image of the workpiece under test.

[0064] Thirdly, in the image preprocessing stage, to improve image quality and reduce interference from subsequent processing, this embodiment employs multiple steps to eliminate noise in the images. First, distortion correction is performed on the four images to eliminate image distortion caused by camera lens or other factors, ensuring the geometric accuracy of the images. Subsequently, to remove salt-and-pepper noise from the images, this embodiment applies a median filtering algorithm for smoothing. The working principle of median filtering is that for each pixel in the image, its value is replaced with the median of all pixel values ​​in its neighborhood. Finally, to further remove noise from the images, this embodiment uses Gaussian filtering. Gaussian filtering works by assigning a weighted average of the pixel values ​​in its neighborhood to each pixel in the image, where the weights are determined by a Gaussian function. This filtering method can smooth the image, remove most of the noise, and retain the overall edge information of the image, providing clearer image data for subsequent image analysis or recognition tasks.

[0065] The process of obtaining the final depth map and the pixel-by-pixel spatial oblique plane includes:

[0066] The single-frame image of the workpiece under test is downsampled twice, and the spatial oblique plane is initialized within the assumed range of the depth to be measured and the imaging angle.

[0067] One image from a single frame of the workpiece under test is set as a reference image (each image is used as a reference image and the following algorithm is applied). The remaining images are source images. The binocular matching cost between the local matching box of the reference pixel and the corresponding matching box of each source image is calculated. The multi-view matching cost is obtained by weighting the binocular matching cost.

[0068] Based on the multi-view matching cost, the spatial plane assumption is iterated, and then the spatial plane is refined to obtain the final depth map corresponding to the reference image after two downsamplings and the pixel-by-pixel set of spatial oblique planes.

[0069] The reference image's corresponding depth map and pixel-by-pixel spatial oblique plane set are subjected to repeated view selection, propagation, and refinement to generate the coarsest-scale depth map.

[0070] The coarsest-scale depth map is refined using a multi-scale consistency matching cost. Then, the refined coarsest-scale depth map is upsampled and followed by the depth map refinement module twice to obtain the intermediate depth map and the pixel-wise spatial oblique plane.

[0071] Specifically:

[0072] The first step is to downsample the initial image twice and initialize the spatial oblique plane within the assumed range of the depth to be measured and the imaging angle.

[0073] Reference image I after downsampling twice ref Each pixel is randomly initialized with a spatial slant plane called the hypothetical plane (which includes continuous normal vectors and depth). This spatial slant plane is different from the front parallel plane at integer parallax values ​​and is closer to the fact that the workpiece being measured is not parallel to the camera imaging plane, as shown in Figure 2.

[0074] The second step is to calculate the multi-view matching cost (treating each image as a separate reference image to calculate the matching cost).

[0075] We define one image as the reference image, and the rest as source images. Consider a pixel in the reference image as reference pixel X. ref Since it corresponds to a randomly initialized spatial plane, combined with multi-view calibration parameters, the center and shape of the matching boxes in other images corresponding to the local matching box of the reference pixel can be calculated using the principle of plane-induced homography. The binocular matching cost is calculated between the local matching box of the reference pixel and the corresponding matching boxes in each source image using the proposed adaptive matching cost. Then, an adaptive weight for each source view is calculated using a multi-hypothesis joint view selection strategy, with the binocular matching cost as the clue. Finally, the multi-view matching cost is calculated using a weighted average. The multi-view matching cost calculation formula is shown below:

[0076] in This represents the weight of the reference pixel for each source image, m. i,j The stereo matching cost calculated between the reference box at the reference pixel and the corresponding box in the source image can be expressed as follows:

[0077] Where ZNCC(*,*) represents the zero-mean normalized cross-correlation matching cost.

[0078] The third step is to iterate the spatial plane assumption based on the multi-view matching cost.

[0079] First, the entire image is divided into a red-black checkerboard pattern, as shown in Figure 3(a). Red pixels can be updated simultaneously with black pixels, and vice versa. The positions of the eight red-filled pixels on the left side of Figure 3(a) are expanded into four V-shaped regions and four strip regions, containing 7 and 11 pixels respectively, as shown in Figure 3(b). Based on this, the optimal multi-view aggregation matching cost (Equation 1) is selected from the pixels in these eight regions to represent the matching degree of the center point. Finally, the checkerboard propagation model is used to update the estimated value of the center pixel using the minimum cost of multi-view aggregation.

[0080] The fourth step is to refine the spatial plane.

[0081] Due to the random initialization of each hypothesis, this embodiment applies a thinning step after each red-black iteration to approximate the true spatial plane corresponding to each pixel in a larger solution space. This embodiment obtains new depths by perturbing the current depth and randomly generating new normals. Subsequently, this embodiment obtains new hypotheses by combining these new depths and normals with the current depth and normals. The hypothesis with the lowest aggregation cost is considered the final estimate for the pixel. After multiple repetitions of view selection, propagation, and thinning, a final depth map corresponding to the reference image after two downsampling iterations can be obtained.

[0082] Step 5: Multi-scale geometric consistency iteration.

[0083] Since surface texture is a relative metric, multi-scale information can be used to improve matching accuracy and completeness. While two downsampling operations have already been performed to enhance surface texture, the reconstruction of thin-structured regions requires finer-scale considerations. After applying multiple iterations of view selection, propagation, and refinement, the coarsest-scale depth map is obtained (steps two through four). The depth map is then refined using a multi-scale consistency matching cost, as shown in the following equation:

[0084] Where Δ i,j λ represents the subpixel reprojection error term, and λ is a constant that balances photometric consistency and geometric consistency.

[0085] The depth map is then upsampled twice. After each upsampling, the depth map refinement module balances the accuracy between thin structures and large planes, and multi-scale consistency matching is run again to further refine the depth map. The above two upsampling and refinement of the depth map yield the final depth map and the pixel-by-pixel spatial oblique plane.

[0086] To improve the overall matching speed while ensuring the reconstruction accuracy of thin structures, a 5×5 matching frame is used in the multi-view patch matching process to obtain a rough intermediate spatial oblique plane in the near-plane region with low curvature and an accurate intermediate spatial oblique plane in the thin structure region with high curvature.

[0087] The process of obtaining the final parameters of the spatial inclined plane includes:

[0088] Based on the obtained pixel-by-pixel intermediate spatial oblique plane, the iterative initial box of the digital speckle correlation method is obtained, and the parameters of the iterative initial box are calculated to obtain the pixel-by-pixel iterative initial value.

[0089] The pixel-by-pixel initial value is iteratively calculated using the digital speckle correlation method, and the final parameters of the spatial oblique plane are determined based on the hyperparameter threshold.

[0090] Furthermore, the process of iteratively calculating the pixel-by-pixel initial value using the digital speckle correlation method also includes: applying the parameter increment to the reference sub-region using the ZNSSD correlation function before each iteration;

[0091] The first step is to obtain the initial bounding box for iteration based on the pixel-by-pixel spatial oblique plane obtained by the multi-view patch matching algorithm.

[0092] Through theoretical derivation and experiments, it can be concluded that there is a one-to-one correspondence between the spatial oblique plane and the first-order shape function of the digital speckle correlation method. Therefore, the pixel-by-pixel intermediate spatial oblique plane obtained by the multi-view patch matching algorithm can be used to obtain the initial value of the pixel-by-pixel iteration of the digital speckle correlation method, so as to avoid the problems of too few seed points and obstructed propagation caused by the seed point selection and propagation process of the traditional digital speckle correlation method. At the same time, GPU acceleration can be used to accelerate the iteration process of the digital speckle correlation method.

[0093] The second step is to iterate and calculate the parameters of the initial bounding box.

[0094] Since there is a one-to-one correspondence between the spatial oblique plane and the first-order shape function of the digital speckle correlation method, the coordinates of the four corner points of the matching box can be used to initialize the first-order shape function. The expression of the first-order shape function is as follows:

[0095] where p1=(u,u x ,u y ,v,v x ,v y ) T Let be the sub-pixel offset and gradient parameters corresponding to the first-order shape function. The first-order shape function can be initialized by simultaneously solving the following two equations:

[0096] By solving the two equations above simultaneously, all the necessary parameters of the first-order shape function can be obtained.

[0097] The third step is to iterate using the multi-view digital speckle correlation method.

[0098] A reference image is set, and the IC-GN algorithm is used to iterate the digital speckle correlation method on each source image. The basic idea is to assume that the reference sub-region can better match the target sub-region after deformation, and to apply parameter increments to the reference sub-region before each iteration. The similarity metric ZNSSD (Zero Normalized Sum of Squares) is used as the similarity metric between the reference and target sub-regions, as shown below:

[0099] Assuming that after deformation of the reference subregion, the Δp form of its ZNSSD correlation function can be expressed as:

[0100] in It is a Hessian matrix. in Here is the gray-level gradient matrix of the reference sub-region in the x, y directions. Let be the Jacobian matrix of the shape functions. After obtaining the mapping parameter increment from the above equation, apply it to the reference subregion to obtain the increment shape function matrix; then invert it and combine it with the current shape function matrix to obtain the updated shape function matrix: W(x,y;p)=W(x,y;p)·W -1 (x,y;Δp;);

[0101] This completes one iteration.

[0102] The fourth step is to set the iteration threshold.

[0103] Because a larger matching box leads to more accurate iterative matching using digital image correlation (DIR), but also results in a loss of matching accuracy at thin structures, after one or two iterations, the ZNCC matching cost threshold is set to 0.85 (a larger threshold indicates more accurate matching), and the Δp threshold is set to 0.05. The optimal source image matching cost that satisfies the Δp threshold is used as the final reconstructed matching box. The surrounding points and the center point are reconstructed by triangulation with the reference image and fitted to a spatial plane. The depth and normal vector of this spatial plane correspond to the updated depth and normal vector of the pixel-by-pixel spatial oblique plane. A 19×19 matching box is used during DIR matching to maximize the reconstruction accuracy of near-planes at areas with small curvature.

[0104] The fifth step is to fuse the four final depth maps.

[0105] In the depth map fusion step, left-right consistency checks and robust fusion strategies are performed. Specifically, this embodiment reprojects the 2D coordinates in the reference depth image back to the world coordinate system, and projects them onto four other depth maps using a homography matrix based on the spatial slant plane information at that pixel. The depth and normal at sub-pixels are obtained through interpolation. Subsequently, this embodiment reprojects them into the world coordinate system and projects them back into the reference depth map to obtain sub-pixel coordinates. The final consistency distance is the Euclidean distance between the original coordinates and the reprojected sub-pixel coordinates. Here, a consistency threshold is defined: a relative depth difference of 0.01, an angle between normals of 30°, and a reprojection error limit of 0.3 pixels. Any corresponding matches at pixels less than the threshold are counted. When the count is greater than 1, the hypothetical spatial slant plane is retained, and the spatial coordinates and normals corresponding to these hypothetical 3D points are averaged as the final result. The 3D spatial coordinates are aggregated to obtain the point cloud of the workpiece surface, as shown in Figure 4.

[0106] The process of obtaining the final point cloud data also includes:

[0107] The two-dimensional coordinates in the final depth map are reprojected back to the world coordinate system and then projected onto the remaining depth maps to obtain the projected subpixel coordinates.

[0108] The subpixel coordinates after projection are reprojected into the world coordinate system, and the spatial coordinates after reprojection into the world coordinate system are projected back into the reference depth map to obtain the subpixel coordinates.

[0109] The Euclidean distance between the original coordinates and the sub-pixel coordinates yields the final consistency distance;

[0110] The final point cloud data is obtained based on the final consistency distance and the threshold of the angle between the spatial oblique plane.

[0111] The measurement results show that the proposed method achieves high integrity reconstruction on highly dynamic surfaces. Multiple reconstruction measurements were performed on the finished standard sheet metal parts, and the results were compared with the standard digital model to evaluate the accuracy. It can be concluded that the proposed method can achieve a measurement accuracy of 0.03-0.07 mm, which meets industrial requirements.

[0112] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A high-dynamic surface single-frame measurement method based on multi-view digital speckle correlation, characterized in that, Includes the following steps: A speckle projection acquisition system is constructed based on a quad-camera and a photolithographic speckle projection module, and a single-frame image of the workpiece under test is acquired based on the speckle projection acquisition system. The single-frame image of the workpiece under test is processed based on a multi-scale robust multi-view matching algorithm to obtain an intermediate depth map and a pixel-by-pixel spatial oblique plane. The intermediate parameters of the spatial oblique plane are iteratively calculated based on the digital speckle correlation method to obtain the final parameters of the spatial oblique plane. The final depth map is fused based on the final parameters of the spatial oblique plane and the left-right consistency superposition test to obtain the final point cloud data.

2. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 1, characterized in that, The process of constructing a speckle projection acquisition system based on a multi-view camera includes: The four cameras around the perimeter are labeled, and a speckle projection module is installed at the symmetrical center of the cameras. The labeled quad-camera is then subjected to single-target and dual-target calibration. The intrinsic and extrinsic parameters of the quad-camera are then optimized using the bundle adjustment method to generate the speckle projection acquisition system.

3. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 1, characterized in that, The process following the acquisition of a single frame image of the workpiece by the speckle projection acquisition system also includes: Distortion correction processing is performed on a single frame image of the workpiece under test to obtain a distortion-corrected image; The distortion-corrected image is processed using a median filtering algorithm to obtain a filtered and denoised image. The filtered image is processed using a Gaussian filtering algorithm to obtain a preprocessed single-frame image.

4. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 1, characterized in that, The process of obtaining the final depth map and the pixel-by-pixel spatial oblique plane includes: The single-frame image of the workpiece under test is downsampled twice, and the spatial oblique plane is initialized within the assumed range of the depth to be measured and the imaging angle. One image from a single frame of the workpiece under test is set as a reference image, and the remaining images are source images. The binocular matching cost between the local matching box of the reference pixel and the corresponding matching box of each source image is calculated. The multi-view matching cost is obtained by weighting the binocular matching cost. Based on the multi-view matching cost, the spatial plane assumption is iterated, and then the spatial plane is refined to obtain the final depth map corresponding to the reference image after two downsamplings and the pixel-by-pixel set of spatial oblique planes. The reference image's corresponding depth map and pixel-by-pixel spatial oblique plane set are subjected to repeated view selection, propagation, and refinement to generate the coarsest-scale depth map. The coarsest-scale depth map is refined using multi-scale consistency matching cost, and then the refined coarsest-scale depth map is upsampled and followed by the depth map refinement module twice to obtain the intermediate depth map and the pixel-wise spatial oblique plane.

5. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 4, characterized in that, The expression for the multi-view matching cost is: in This represents the weight of the reference pixel for each source image, m. i,j The stereo matching cost is calculated between the reference box at the reference pixel and the corresponding box in the source image.

6. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 1, characterized in that, The process of obtaining the final parameters of the spatial inclined plane includes: Based on the obtained pixel-by-pixel intermediate spatial oblique plane, the iterative initial box of the digital speckle correlation method is obtained, and the parameters of the iterative initial box are calculated to obtain the pixel-by-pixel iterative initial value. The pixel-by-pixel initial value is iteratively calculated using the digital speckle correlation method, and the final parameters of the spatial oblique plane are determined based on the hyperparameter threshold.

7. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 6, characterized in that, The process of iteratively calculating the pixel-by-pixel initial value using the digital speckle correlation method further includes: applying the parameter increment to the reference sub-region using the ZNSSD correlation function before each iteration; The expression for the ZNSSD correlation function in Δp form is: in, It is a Hessian matrix. in Here is the gray-level gradient matrix of the reference sub-region in the x, y directions. Let be the Jacobian matrix of the shape functions.

8. The high dynamic surface single-frame measurement method based on multi-view digital speckle correlation according to claim 1, characterized in that, The process of obtaining the final point cloud data also includes: The two-dimensional coordinates in the final depth map are reprojected back to the world coordinate system and then projected onto the remaining depth maps to obtain the projected subpixel coordinates. The subpixel coordinates after projection are reprojected into the world coordinate system, and the spatial coordinates after reprojection into the world coordinate system are projected back into the reference depth map to obtain the subpixel coordinates. The Euclidean distance between the original coordinates and the sub-pixel coordinates yields the final consistency distance; The final point cloud data is obtained based on the final consistency distance and the threshold of the angle between the spatial oblique plane.

Images