A multi-spot denoising correction method and three-dimensional reconstruction method based on image conversion

By constructing the MSIDNet model and combining the eight-connected area segmentation and Gaussian fitting algorithm, the noise problems caused by long-term exposure and natural light interference of CMOS are solved, and high-precision denoising and three-dimensional reconstruction of multi-spot images are achieved, improving the accuracy of measurement Sex and reliability.

CN117952854BActive Publication Date: 2024-11-08GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410145291.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-02
Publication Date
2024-11-08
Estimated Expiration
2044-02-02

AI Technical Summary

Technical Problem

The prior art is difficult to effectively overcome the noise problems caused by long-term exposure of CMOS and natural light interference, affecting the precise positioning of the center of the multi-spot image, resulting in a decrease in accuracy and reliability when measuring cylindrical axial objects.

Method used

Using a multi-spot denoising correction method based on image conversion, by constructing an MSIDNet model, the image features are extracted and restored using an encoder and decoder, combined with eight-connected area segmentation and Gaussian fitting algorithm, noise is removed and three-dimensional reconstruction is performed.

Benefits of technology

The noise interference in multi-spot images was successfully removed, which improved the spot positioning accuracy and the accuracy of three-dimensional reconstruction, and significantly improved the reliability of measurement.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117952854B_ABST
    Figure CN117952854B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of image processing, and more specifically, it relates to a multi-spot denoising correction method and a three-dimensional reconstruction method based on image conversion. The main points of the technical solution are: obtaining a multi-spot image with noise; building an MSIDNet model; inputting the multi-spot image with noise into the MSIDNet model to obtain a denoised image; dividing the denoised image into multiple sub-spots using an eight-connected region ROI; extracting the central coordinates of the multiple sub-spots using a Gaussian fitting algorithm; matching corresponding sub-spots based on a moving window; and performing three-dimensional reconstruction based on an iterative method to obtain the three-dimensional morphology of the original object.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more specifically, to a multi-spot denoising correction method and a three-dimensional reconstruction method based on image conversion. Background Art

[0002] Optical elements are key components of optical systems and are widely used in imaging, refraction, transmission, reflection, polarization control and other fields. In these applications, the surface morphology of optical elements has a decisive influence on their performance and function. In previous studies, a multi-beam angle sensor (MBAS) for detecting plane errors and roundness errors was developed. The sensor has the characteristics of independence from spindle errors, miniaturization, high sensitivity, and suitability for factory-level applications. Using the circular scanning method, MBAS can stably and accurately measure the roundness error of cylindrical workpieces. In the process of morphology measurement, accurate positioning of multi-spot images is an important factor to ensure its high-precision measurement. However, with the long exposure of the camera and the interference of natural light, the noise of the multi-spot image increases, causing the spot to be distorted, affecting the precise positioning of the center of the multi-spot image.

[0003] At present, there are two main types of traditional methods for detecting the center of a spot image: the first type is based on the shape of the spot, such as the circle fitting method, the ellipse fitting method (EFM) and the Hough transform method (HFM). This type of method has high computational complexity, and its positioning accuracy is greatly affected by the structure of the spot. The second type is based on pixel values, such as the grayscale centroid method (GCM), the Otsu threshold centroid method (OTCM), and the Gaussian fitting method (GFM). Among them, GCM is simple to calculate, but its accuracy is average and it is easily affected by noise. Compared with GCM, OTCM has higher positioning accuracy when the contrast between the spot and the noise is obvious, but it cannot eliminate the influence of extreme noise types such as salt and pepper noise. GFM is suitable for spots with shapes close to Gaussian distribution and high image quality and low noise. It usually needs to be used in combination with other methods to obtain more accurate positioning. In recent years, deep learning methods have been widely used in image detection. Its powerful data fitting ability can capture nonlinear relationships in data and improve model prediction accuracy. In order to address the challenge of noise problem caused by long-term exposure of CMOS and interference of natural light in MBAS, which leads to distortion of light spot, a method to overcome the noise problem caused by long-term exposure of CMOS and interference of natural light, so as to improve the accuracy and reliability of MBAS in measuring cylindrical shaft objects, needs to be developed urgently. Summary of the invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a multi-spot denoising correction method and a three-dimensional reconstruction method based on image conversion, which overcomes the noise problem caused by long-term CMOS exposure and natural light interference and improves the accuracy of three-dimensional reconstruction.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions: A multi-spot denoising correction method based on image conversion, comprising the following steps:

[0006] Acquire multi-spot images with noise;

[0007] Build the MSIDNet model;

[0008] Inputting the noisy multi-spot image into the MSIDNet model to obtain a denoised image;

[0009] The MSIDNet model includes an encoder and a decoder;

[0010] The encoder includes a plurality of convolutional layers and pooling layers, wherein the convolutional layer is used to extract abstract features of the noisy multi-spot image, the convolutional layer is a 3*3 convolution, and the pooling layer is used to remove redundant information of the noisy multi-spot image, and fuse the plurality of abstract features based on a splicing operation of feature mapping;

[0011] The convolutional layer includes a batch normalization layer and a ReLU activation function;

[0012] The decoder comprises a plurality of convolutional layers and upsampling layers, wherein the upsampling layer performs an upsampling operation using a bilinear interpolation technique to restore a target image, and the upsampling layer performs a splicing operation before performing a second and a third upsampling operation to fuse high-level semantic features with low-level semantic features to transfer low-level and high-level feature information, and the convolutional layer is further used to identify noise in the noisy multi-spot image and output the denoised image;

[0013] A three-dimensional reconstruction method, characterized by comprising the following steps:

[0014] Acquire a denoised image, wherein the multi-spot image is obtained by the multi-spot denoising correction method based on image conversion;

[0015] The denoised image is divided into a plurality of sub-spots using eight connected regions R0I;

[0016] A Gaussian fitting algorithm is used to extract the center coordinates of the plurality of sub-spots;

[0017] Matching the corresponding sub-spot based on a moving window;

[0018] Perform three-dimensional reconstruction based on an iterative method to obtain the three-dimensional shape of the original object;

[0019] The step of dividing the denoised image into a plurality of sub-spots using eight connected regions ROI is specifically as follows:

[0020] The denoised image is converted into a grayscale image. Based on the connectivity characteristics of the spot image, the eight connected region labeling method is used to label each connected region of the grayscale image with the same label. The grayscale image is segmented according to the labeling parameters. Each connected region of the grayscale image has the same label, which is denoted as C. i , i = 1, 2, ..., N, N is the total number of connected regions, connected region C i It can be calculated by the MATLAB function bwlabel(), and the calculation formula is:

[0021] C i =bwlabel(I(x,y),i)

[0022] Among them, i is the index of the connected area, and I(x, y) is the grayscale image;

[0023] The regionprops() function of MATLAB is used to calculate the clipping matrix B of the single spot image after segmentation by the eight connected regions R0I. i , i = 1, 2, ..., N, the size of the matrix is ​​the area C of each spot connection after segmentation i The smallest rectangle of

[0024] The MATLAB function imcrop() is used to extract the mask image corresponding to the cropping matrix of the original denoised image. The specific operation is as follows:

[0025] mask i =imcrop(I2(x,y),B i )

[0026] Among them, mask i Represents the corresponding cropping matrix B in the multi-spot denoised image I2(x, y) i Each light spot connected area image, I2(x, y) is the denoised image after training with the MSIDNet model;

[0027] The steps are based on the iterative method to perform three-dimensional reconstruction, and the three-dimensional morphology of the original object is obtained as follows:

[0028] The coordinates of the corresponding sub-spot in the rotation plane are reconstructed through the slope difference. Taking the point on the rotation axis as the center of the circle, the coordinates of the first point are first determined. The calculation method is as follows:

[0029]

[0030] Among them, (x1, y1) is the coordinate of the first point, th is the rotation angle, and r is the distance from the starting point to the center of the circle;

[0031] The slope ki Calculated by the following formula:

[0032]

[0033] Among them, y i is the ordinate of the i-th point, x i is the horizontal coordinate of the i-th point, k i is the slope;

[0034] The point (x1, y1) and the point (x i ,y i ) is:

[0035] y i =tan(θ i-1 +th)x i (2)

[0036] Among them, (x1, y1) is the coordinate of the first point, th is the rotation angle, and r is the distance from the starting point to the center of the circle;

[0037] Taking π / 2+th / 2 as the initial angle of the first point, we can get the following formula:

[0038] θ i-1 =π / 2+i×th-th / 2 (3)

[0039] By combining formulas (1), (2), and (3), we can obtain the following relationship between the two points:

[0040]

[0041] Among them, th is the rotation angle, y i is the ordinate of the i-th point, x i is the horizontal coordinate of the i-th point, k i is the slope;

[0042] All points are calculated using the above formula, and the three-dimensional shape contour of the cylindrical shaft object can be reconstructed based on the calculated points.

[0043] In one embodiment, the step of obtaining a multi-spot image with noise is specifically as follows:

[0044] Use Zemax software to simulate and obtain multi-spot images;

[0045] Matlab software is used to add various noises to the multi-spot image to obtain the multi-spot image with noise.

[0046] In one embodiment, the step of inputting the noisy multi-spot image into the MSIDNet model to obtain the denoised image is specifically as follows:

[0047] The above-mentioned multi-spot denoising correction method and three-dimensional reconstruction method based on image conversion have the following beneficial effects:

[0048] The noisy multi-spot image is input into an encoder, the multiple convolution layers extract the abstract features of the noisy multi-spot image through convolution kernels, the pooling layer removes redundant information of the noisy multi-spot image, and the multiple abstract features are fused based on a splicing operation of feature mapping to obtain a feature map;

[0049] The feature map is input into the decoder, the upsampling layer performs upsampling operation by using bilinear interpolation technology, the upsampling layer performs splicing operation before performing the second and third upsampling operations, and the convolution layer outputs the denoised image after identifying the noise in the feature map.

[0050] In one embodiment, the step uses a Gaussian fitting algorithm to extract the fitting formula of the center coordinates of the plurality of sub-spots:

[0051]

[0052] Among them, f(x, y) is the grayscale distribution function of the sub-spot, and the sub-pixel spot center coordinates (x0, y0) can be obtained by least squares fitting.

[0053] In one embodiment, the step of matching the corresponding sub-spot based on the moving window is specifically as follows:

[0054] A threshold value T is selected to screen the corresponding sub-light spots. When the sub-light spot at the next angle is located in a window with the threshold value T as the radius window and the sub-light spot as the center, the two sub-light spots are identified as corresponding sub-light spots.

[0055] Starting from the initial angle, according to the offset of the sub-spot, the position of the window is moved, and the corresponding sub-spot is identified in a continuous cycle until a rotation is completed. Here, the value of T is as follows:

[0056]

[0057] Where P is the microlens array pitch, d s is the CMOS pixel size.

[0058] The above-mentioned multi-spot denoising correction method and three-dimensional reconstruction method based on image conversion have the following beneficial effects:

[0059] First, after the denoising and correction processing of MSIDNet, the noise interference in the multi-spot image is successfully removed, and the distorted sub-spots are effectively corrected;

[0060] Secondly, using the eight-connected domain segmentation technology, each sub-spot can be accurately segmented from the image, and the segmented sub-spot can use the Gaussian fitting method to calculate its sub-pixel spot center coordinates, thereby improving the spot positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the process of this embodiment;

[0062] Figure 2 Schematic diagram of the MSIDNet network structure in this embodiment;

[0063] Figure 3 2 is a schematic diagram of the identification results of sub-spots corresponding to two different cylindrical shaft objects under test in this embodiment;

[0064] Figure 4 In this embodiment, the point (x1, y1) and the point (x i ,y i ) diagram of the relationship between them; DETAILED DESCRIPTION

[0065] The present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0066] A three-dimensional reconstruction method, as shown in the figure, comprises the following steps:

[0067] Obtain a denoised image;

[0068] Specifically, Figure 1 As shown, the following steps are included:

[0069] Acquire multi-spot images with noise;

[0070] Specifically, a multi-spot image is obtained by simulation using Zemax software, and the optical component specifications of a multi-beam angle sensor (MBAS) are referred to when simulating using Zemax software to obtain an ideal multi-spot image;

[0071] Matlab software is used to add various noises to the multi-spot image to obtain a multi-spot image with noise.

[0072] Build the MSIDNet model;

[0073] like Figure 2 As shown, the MSIDNet model includes an encoder (D) and a decoder (T);

[0074] The encoder has multiple convolutional layers and pooling layers. The convolutional layer is used to extract the abstract features of the multi-spot image with noise. The convolutional layer is a 3*3 convolution. The pooling layer is used to remove the redundant information of the multi-spot image with noise. After the pooling layer eliminates the redundant information of the multi-spot image with substitute noise, the abstract features extracted by multiple convolutional layers are fused based on the splicing operation of feature mapping.

[0075] The convolutional layer includes a batch normalization layer and a ReLU activation function;

[0076] The decoder includes multiple convolutional layers and upsampling layers. The upsampling layer uses bilinear interpolation technology to perform upsampling operations to restore the target image. The upsampling layer performs splicing operations before the second and third upsampling operations to fuse high-level semantic features with low-level semantic features to transfer low-level and high-level feature information. The convolutional layer is also used to identify noise in the noisy multi-spot image and output a denoised image.

[0077] Input the noisy multi-spot image into the MSIDNet model to obtain the denoised image;

[0078] Specifically, a multi-spot image with noise is input into an encoder, multiple convolutional layers extract abstract features of the multi-spot image with noise through convolution kernels, redundant information of the multi-spot image with noise is removed in a pooling layer, and the abstract features extracted by multiple convolutional layers are fused based on a splicing operation of feature mapping to obtain a feature map;

[0079] The feature map is input into the decoder, and the upsampling layer uses bilinear interpolation technology to perform upsampling operations. The upsampling layer performs splicing operations before the second and third upsampling operations. After the convolution layer identifies the noise in the feature map, it outputs the denoised image.

[0080] The denoised image is divided into multiple sub-spots using the eight-connected region R0I;

[0081] Specifically, the denoised image is converted into a grayscale image. Based on the connectivity characteristics of the spot image, the eight-connected region labeling method is used to label each connected region of the grayscale image with the same label. The grayscale image is segmented according to the labeling parameters. Each connected region of the grayscale image has the same label, which is denoted as C. i , i = 1, 2, ..., N (where N is the total number of connected regions), connected region C i It can be calculated by the MATLAB function bwlabel(), and the calculation formula is:

[0082] C i =bwlabel(I(x,y),i)

[0083] Among them, i is the index of the connected area, and I(x, y) is the grayscale image;

[0084] The regionprops() function of MATLAB is used to calculate the clipping matrix Bi of a single spot image after segmentation by the eight connected regions R01, where i = 1, 2, ..., N. The size of the matrix is ​​the size of each spot connection region C after segmentation. i The smallest rectangle of

[0085] Use the MATLAB function imcrop() to extract the mask image corresponding to the cropping matrix of the original denoised image. The specific operations are as follows:

[0086] mask i =imcrop(I2(x,y),B i )

[0087] Among them, mask i Represents the corresponding cropping matrix B in the multi-spot denoised image I2(x, y) i I2(x, y) is the denoised image after training with the MSIDNet model.

[0088] The Gaussian fitting algorithm is used to extract the center coordinates of multiple sub-spots;

[0089] After the multi-spot image is denoised and corrected by the MSIDNet model and segmented by eight connections, each sub-spot in the multi-spot image can be approximately presented as a two-dimensional Gaussian model, in which the grayscale distribution function of each sub-spot can be approximately regarded as a two-dimensional Gaussian distribution and can be fitted by a two-dimensional Gaussian function. The fitting formula is as follows:

[0090]

[0091] Among them, f(x, y) is the grayscale distribution function of the sub-spot, and the sub-pixel spot center coordinates can be obtained by least squares fitting.

[0092] Match the corresponding sub-spots based on the moving window;

[0093] A threshold T is selected to screen the corresponding sub-spots. When the sub-spot of the next angle is located in a window with the threshold T as the radius and the sub-spot as the center, the two sub-spots are identified as corresponding sub-spots.

[0094] Starting from the initial angle, according to the offset of the sub-spot, the position of the window is moved, and the corresponding sub-spot is identified in a continuous cycle until a rotation is completed. Here, the value of T is as follows:

[0095]

[0096] Where P is the microlens array pitch, d s is the CMOS pixel size;

[0097] like Figure 3 As shown in the figure, the multi-spot images and the corresponding sub-spots are identified at various angles after two different cylindrical shaft objects rotated one circle at a rotation angle of 10°.

[0098] The white dots represent the centroid coordinates of the sub-spots of the multi-spot images at various angles, and the white lines represent the motion trajectories of the corresponding sub-spots.

[0099] Perform three-dimensional reconstruction based on an iterative method to obtain the three-dimensional shape of the original object;

[0100] Specifically, the coordinates of the corresponding sub-spot on the rotation plane (with the point on the rotation axis as the center) are reconstructed through the slope difference. First, the coordinates of the first point are determined. The calculation method is as follows:

[0101]

[0102] Among them, (x1, y1) is the coordinate of the first point, th is the rotation angle, and r is the distance from the starting point to the center of the circle;

[0103] Slope k i Calculated by the following formula:

[0104]

[0105] Among them, y i is the ordinate of the i-th point, x i is the horizontal coordinate of the i-th point, k i is the slope;

[0106] The point (x1, y1) and the point (x i ,y i ) is:

[0107] y i =tan(θ i-1 +th)x i (2)

[0108] Among them, (x1, y1) is the coordinate of the first point, th is the rotation angle, and r is the distance from the starting point to the center of the circle;

[0109] Taking π / 2+th / 2 as the initial angle of the first point, we can get the following formula:

[0110] θ i-1 =π / 2+i×th-th / 2 (3)

[0111] By combining formulas (1), (2), and (3), we can obtain the following relationship between the two points:

[0112]

[0113] Among them, th is the rotation angle, y i is the ordinate of the i-th point, x i is the horizontal coordinate of the i-th point, k i is the slope;

[0114] All points are calculated using the above formula, and the three-dimensional shape contour of the cylindrical shaft object can be reconstructed based on the calculated points.

[0115] Comparative Example

[0116] Multi-spot simulation positioning comparison test

[0117] CFT, GCM, GFM, OTCM and MSIDNet methods are used to extract the center of the sub-spot.

[0118] The center positioning error formula is used to calculate the center positioning error of the sub-spots obtained by the above methods. The specific formula is as follows:

[0119]

[0120] in, and is the actual spot center, x i and i It is the spot center calculated by several different algorithms.

[0121] After calculation, the average error of the GCM method is 0.8091 pixels. The OTCM method reaches 1.2710 pixels. The CFT and GFM methods are 0.1679 pixels and 0.3374 pixels respectively. The average error of the method of the present invention is only 0.0603 pixels. Experiments have shown that the method of the present invention can effectively remove noise in multi-spot images, thereby improving the positioning accuracy of the spots.

[0122] Algorithm stability analysis

[0123] Analyze different noise intensities and different noise types. Calculate the average error to evaluate the positioning accuracy. Analyze the four key performance indicators of the maximum error, minimum error, average error and standard deviation (std) of the above five algorithms, where the standard deviation is calculated by the formula. The formula is as follows:

[0124]

[0125] Where n is the number of samples, std is the standard deviation, E i is the positioning deviation of each sample, is the mean deviation of the sample.

[0126] The calculation shows that the maximum error of the method used in the present invention is 0.3034 pixels, the minimum error is 0.0085, and the average error is 0.0849, which are better than other algorithms. In addition, the std is only 0.0502. The comprehensive experimental results show that the MSIDNet algorithm exhibits high stability and excellent accurate positioning performance in different noise environments.

[0127] Analysis of 3D reconstruction accuracy

[0128] The contour after surface reconstruction of the MSIDNet positioning result is unfolded along the central axis, and its height represents the distance of the surface contour from the central axis. For the measured parts with radii of 10mm, 20mm and 22mm, the std value (standard deviation) and pv value (measured value) of the positioning method based on the Gaussian fitting method (GFM) and the method of the present invention are compared. The experimental results show that the method of the present invention can effectively reduce the std value and accurately reconstruct the surface morphology of cylindrical shaft parts.

[0129] Therefore, this method has high-precision center positioning and high stability, and realizes multi-spot image correction and three-dimensional shape measurement.

[0130] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the attached claims.

Claims

1. A three-dimensional reconstruction method, characterized in that: The following steps are involved: The multi-spot denoising correction based on image conversion obtains the denoised image; The denoised image is divided into a plurality of sub-spots by using eight connected regions ROI; A Gaussian fitting algorithm is used to extract the center coordinates of the plurality of sub-spots; Matching the corresponding sub-spot based on a moving window; Perform three-dimensional reconstruction based on an iterative method to obtain the three-dimensional shape of the original object; The step of obtaining a denoised image based on multi-spot denoising correction of image conversion is specifically as follows: Acquire multi-spot images with noise; Build the MSIDNet model; Inputting the noisy multi-spot image into the MSIDNet model to obtain the denoised image; The MSIDNet model includes an encoder and a decoder; The encoder includes a plurality of convolutional layers and pooling layers, wherein the convolutional layer is used to extract abstract features of the noisy multi-spot image, the convolutional layer is a 3*3 convolution, and the pooling layer is used to remove redundant information of the noisy multi-spot image, and fuse the plurality of abstract features based on a splicing operation of feature mapping; The convolutional layer includes a batch normalization layer and a ReLU activation function; The decoder comprises a plurality of convolutional layers and upsampling layers, wherein the upsampling layer performs an upsampling operation using a bilinear interpolation technique to restore a target image, and the upsampling layer performs a splicing operation before performing a second and a third upsampling operation to fuse high-level semantic features with low-level semantic features to transfer low-level and high-level feature information, and the convolutional layer is further used to identify noise in the noisy multi-spot image and output the denoised image; The step of dividing the denoised image into a plurality of sub-spots using eight connected regions ROI is specifically as follows: The denoised image is converted into a grayscale image. Based on the connectivity characteristics of the spot image, the eight connected region labeling method is used to label each connected region of the grayscale image with the same label. The grayscale image is segmented according to the labeling parameters. Each connected region of the grayscale image has the same label, which is denoted as C. i , I = 1, 2, ..., N, where N is the total number of connected regions, and the connected region C i It can be calculated by the MATLAB function bwlabel(), and the calculation formula is: C i =bwlabel(I(x,y),i) Among them, i is the index of the connected area, and I(x,y) is the grayscale image; The regionprops() function of MATLAB is used to calculate the clipping matrix B of the single spot image after the eight connected regions ROI segmentation. i , i = 1, 2, ..., N, the size of the matrix is ​​the area C of each spot connection after segmentation i The smallest rectangle of The MATLAB function imcrop() is used to extract the mask image corresponding to the cropping matrix of the original denoised image. The specific operation is as follows: mask i =imcrop(I2(x,y),B i ) Among them, mask i Represents the corresponding cropping matrix B in the multi-spot denoised image I2(x,y) i Each light spot connected area image, I2(x, y) is the denoised image after training with the MSIDNet model; The steps are based on the iterative method to perform three-dimensional reconstruction to obtain the three-dimensional morphology of the original object: The coordinates of the corresponding sub-spot in the rotation plane are reconstructed through the slope difference. Taking the point on the rotation axis as the center of the circle, the coordinates of the first point are first determined. The calculation method is as follows: Among them, (x1, y1) is the coordinate of the first point, th is the rotation angle, and r is the distance from the starting point to the center of the circle; The slope k i Calculated by the following formula: Among them, y i is the ordinate of the i-th point, x i is the horizontal coordinate of the i-th point, k i is the slope; The point (x1, y1) and the point (x i ,y i ) is: y i =tan(θ i-1 +th)x i (2) Among them, (x1, y1) is the coordinate of the first point, th is the rotation angle, and r is the distance from the starting point to the center of the circle; Taking π / 2+th / 2 as the initial angle of the first point, we can get the following formula: i i-1 =π / 2+i×th-th / 2 (3) By combining formulas (1), (2), and (3), we can obtain the following relationship between the two points: Among them, th is the rotation angle, y i is the ordinate of the i-th point, x i is the horizontal coordinate of the i-th point, k i is the slope; All points are calculated using the above formula, and the three-dimensional shape contour of the cylindrical shaft object can be reconstructed based on the calculated points.

2. A three-dimensional reconstruction method according to claim 1, characterized in that: The steps for obtaining a multi-spot image with noise are specifically as follows: Use Zemax software to simulate and obtain multi-spot images; Matlab software is used to add various noises to the multi-spot image to obtain the multi-spot image with noise.

3. A three-dimensional reconstruction method according to claim 1, characterized in that: The step of inputting the noisy multi-spot image into the MSIDNet model to obtain the denoised image is specifically as follows: The noisy multi-spot image is input into an encoder, the multiple convolution layers extract the abstract features of the noisy multi-spot image through convolution kernels, the pooling layer removes redundant information of the noisy multi-spot image, and the multiple abstract features are fused based on a splicing operation of feature mapping to obtain a feature map; The feature map is input into the decoder, the upsampling layer performs upsampling operation by using bilinear interpolation technology, the upsampling layer performs splicing operation before performing the second and third upsampling operations, and the convolution layer outputs the denoised image after identifying the noise in the feature map.

4. A three-dimensional reconstruction method according to claim 1, characterized in that: The step uses a Gaussian fitting algorithm to extract the fitting formula of the center coordinates of the plurality of sub-spots: Among them, f(x, y) is the grayscale distribution function of the sub-spot, and the sub-pixel spot center coordinates (x0, y0) can be obtained by least squares fitting.

5. A three-dimensional reconstruction method according to claim 1, characterized in that: The step of matching the corresponding sub-spot based on the moving window is specifically as follows: A threshold value T is selected to screen the corresponding sub-light spots. When the sub-light spot at the next angle is located in a window with the threshold value T as the radius window and the sub-light spot as the center, the two sub-light spots are identified as corresponding sub-light spots. Starting from the initial angle, according to the offset of the sub-spot, the position of the window is moved, and the corresponding sub-spot is identified in a continuous cycle until a rotation is completed. Here, the value of T is as follows: Where P is the microlens array pitch, d s is the CMOS pixel size.