A method and system for aspheric surface topography measurement based on deep learning

Abstract

The present application relates to the field of optical measurement, more particularly, it relates to a kind of aspherical surface topography measurement method and system based on deep learning, its technical points are, including: S1, preparation simulation image data set, image data set is divided into training set and test set according to the proportion of 8:2;S2, image data set is loaded into deep learning network and is trained, saves best weight parameter;S3, construction system, obtains multi-spot image data;S4, best weight parameter is loaded into deep learning network, and then using deep learning network is positioned to multi-spot image data center;S5, using Fourier transform method measures the three-dimensional topography of aspherical surface object.The present application solves the positioning precision decline caused by the influence of uneven illumination, spot deformation and other factors when traditional spot center positioning method faces the complex surface of aspherical lens, compared with traditional center positioning and deep learning positioning method, it is more efficient and more optimal in precision.

Description

Technical Field

[0001] This invention relates to the field of optical measurement, and more specifically, to a method and system for measuring aspherical topography based on deep learning. Background Technology

[0002] In fields such as optical engineering, precision manufacturing, and optical instrument calibration, the measurement of the three-dimensional surface topography of aspherical lenses is crucial. Traditional measurement methods generally face challenges such as expensive equipment, complex operation, and sensitivity to the environment, especially when dealing with aspherical lenses with large apertures, high curvature variations, or microstructures, where accuracy and efficiency are difficult to meet requirements.

[0003] In recent years, three-dimensional topography measurement technology based on optical reflection principles and image processing has attracted attention due to its non-contact, fast, and flexible characteristics. Among them, the key to three-dimensional topography measurement technology lies in multi-spot center localization. Traditional spot center localization methods are easily affected by factors such as uneven illumination, spot deformation, and image noise when dealing with complex surfaces of aspherical lenses, resulting in a decrease in localization accuracy.

[0004] The rapid development of deep learning technology has provided new insights into spot center localization, demonstrating powerful capabilities in image processing, feature extraction, and predictive analysis. Existing research shows that deep learning models can automatically learn complex features in spot images and achieve accurate spot center localization. However, current deep learning-based multi-spot center localization methods are still in the research stage and suffer from problems such as high algorithm complexity and insufficient real-time performance. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for measuring aspherical topography based on deep learning, so as to solve the above problems.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a method for measuring aspherical topography based on deep learning, comprising the following steps:

[0007] S1. Prepare a simulation image dataset, and divide the image dataset into a training set and a test set in an 8:2 ratio;

[0008] S2. Load the image dataset into the deep learning network for training and save the optimal weight parameters;

[0009] S3. Construct the system to acquire multi-spot image data;

[0010] S4. Load the optimal weight parameters into the deep learning network, and then use the deep learning network to perform center localization on the multi-spot image data;

[0011] S5. Use the Fourier transform method to measure the three-dimensional shape of an aspherical object.

[0012] In one embodiment, step S1 further includes:

[0013] S11. Generate a standard Gaussian spot using a two-dimensional Gaussian function;

[0014] S12. Generate the standard Gaussian spots at different positions in the image, denoted as a standard Gaussian multi-spot image, and record the center coordinates of each standard Gaussian spot;

[0015] S13. Add Gaussian noise and salt-and-pepper noise to the standard Gaussian multi-spot image, and record it as a noisy multi-spot image;

[0016] S14. Using the center coordinates of each of the standard Gaussian spots as labels and the noisy multi-spot images as corresponding data, the noisy multi-spot images are divided into a dataset and a test set in an 8:2 ratio.

[0017] In one embodiment, the two-dimensional Gaussian function is:

[0018]

[0019] Where A is the total energy of the light spot, (x0, y0) is the center coordinate of the light spot, and σ x and σ y The standard deviations along the x-axis and y-axis, respectively, determine the width of the light spot in the corresponding directions.

[0020] In one embodiment, in step S2, the deep learning network includes a backbone, an FPN-Neck section, and a detector section that integrates a localization attention module;

[0021] Step S2 specifically includes the following steps:

[0022] S21. Input the simulated image dataset into the backbone. The input images are processed by the first Focus convolution layer. Each image is divided into four sub-images by taking a value every other pixel. The information of each sub-image is concentrated into the channel space. The input channels are expanded by 4 times to obtain a stitched image with 12 channels.

[0023] The stitched image is then sequentially input into the first CBS module and the second CBS module. The first CBS module uses 16 convolutional kernels of size 3x3 with a stride of 1 to expand the channels of the stitched image to 16 channels. The second CBS module uses 32 convolutional kernels of size 3x3 with a stride of 2 to expand the channels of the stitched image to 32 channels and halve the size of the stitched image.

[0024] The stitched image is processed by the C3 module to obtain the feature map Feature_C0;

[0025] S22. Input the feature map Feature_C0 into the third CBS module of the backbone, amplify the shallow features by channel expansion, and then input it into the C3 module for feature learning, and output the feature map Feature_C1.

[0026] S23. The feature map Feature_C1 is passed through the fourth CBS module and C3 module of the main body to obtain the feature map Feature_C2. The feature map Feature_C2 is passed through the fifth CBS module and C3 module of the main body, and then through spatial pyramid pooling to obtain the feature map Feature_C3.

[0027] S24. Using the feature map Feature_C3 as the input to the multi-scale fusion feature FPN-Neck part, the feature map Feature_Up1 is obtained after passing through the sixth CBS module of the neural network and undergoing an upsampling operation.

[0028] S25. The feature map Feature_Up1 and the feature map Feature_C2 are fused to obtain the fused feature map Feature_Fuse1. Then, through the C3 module and the seventh CBS module of the neural network, the feature map Feature_Up2 is obtained after an upsampling operation.

[0029] S26. The feature map Feature_Up2 and the feature map Feature_C1 are fused to obtain the fused feature map Feature_Fuse2. Then, through the C3 module and the eighth CBS module of the neural network, the feature map Feature_Up3 is obtained after an upsampling operation.

[0030] S27. The feature map Feature_Up3 and the feature map Feature_C0 are fused to obtain the feature map Feature_Fuse3, and then passed through the C3 module with a convolution kernel size of 3*3 and a channel number of 64 to obtain the feature map Feature_C4.

[0031] S28. Input the feature map Feature_C4 into the detector of the integrated localization attention module to obtain a multidimensional array containing the center coordinates of the light spot.

[0032] In one embodiment, the detector with integrated localization attention module includes a localization attention module and a detection module;

[0033] The positioning attention module consists of a channel attention module and a spatial attention module;

[0034] The execution steps of the channel attention module are as follows:

[0035] Input feature map Obtain the one-dimensional channel weight matrix The one-dimensional channel weight matrix With input feature map Multiplication produces a multi-scale feature map with channel attention.

[0036] The execution steps of the spatial attention module are as follows:

[0037] Input the feature map Obtain spatial attention feature maps containing location information The feature map With the feature map Multiplication yields the feature map Defined as:

[0038]

[0039] The channel attention module generates a channel attention weight matrix by analyzing the channel correlations in the feature map. The specific formula is as follows:

[0040]

[0041] F ex (F sq ,w)=σ(ω2δ(ω1F sq ))

[0042] F C (F,F ex )=F·F ex

[0043] Where i and j are respectively the feature maps In the i-th row and j-th column, σ is the sigmoid activation function, δ is the ReLU activation function, and ω1 and ω2 represent the weights of two different fully connected layers;

[0044] The spatial attention module receives input feature maps To extract spatial mapping relationships, specifically, depthwise separable convolutions are used to capture the relevant weights between different regions, preserving channel correlations while reducing computational overhead. The formula is defined as follows:

[0045]

[0046] Among them, DConv i×i This indicates a depthwise separable convolution with a kernel size of i×i;

[0047] After passing through the localization attention module, the feature map Feature_C4 inputs multi-scale features into the detection module;

[0048] feature map The data is input into the CBS module and residual connections are introduced to facilitate cross-layer information flow. Finally, two parallel 1x1 convolutions are used to obtain the regression parameters and IoU values ​​related to the detection target.

[0049] In one embodiment, step S2 further includes: designing the loss function as a center loss function;

[0050] The central loss function is derived by using the bounding box loss L. ciou And confidence loss L conf Composition, which is defined as:

[0051] L = L ciou +L conf

[0052] Loss=(1-IoU)×L ciou +(1-IoU)×L ciou +IoU×L Center

[0053]

[0054] in, The sample label is 1 for the point and 0 for the background. P i It is the probability that a sample is predicted as a point.

[0055] B i These are the predicted coordinates of the center of point i. p is the center coordinate of the label corresponding to the i-th point; IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box. 2 (B,B gt ) represents the Euclidean distance between the predicted point and the true point, c is the diagonal length of the smallest outer rectangle containing the predicted and true values, α is the weight parameter, and v is used to measure the similarity of the aspect ratio between the predicted and true values. α and v are specifically defined as follows:

[0056]

[0057] Among them, w gt h gt w and h are the width and height of the ground truth bounding box and the predicted bounding box, respectively.

[0058] In one embodiment, step S3 includes the following steps:

[0059] S31. Install the sensor on the dual-axis moving platform;

[0060] S32. Position the rotating platform below the sensor;

[0061] S33. Install the computer and electrically connect it to the sensor, the dual-axis moving platform, and the rotating platform respectively.

[0062] In one embodiment, step S5 includes the following steps:

[0063] S51. Based on the center coordinates of each spot in the multi-spot image, the angle difference is calculated by the corresponding spot coordinate positions on two adjacent multi-spot images.

[0064] S52. Based on the relationship between the Fourier series coefficients of the angle difference and the surface morphology, reconstruct the three-dimensional morphology information of the aspherical optical element.

[0065] An aspherical surface topography measurement system includes a dual-axis moving platform, a sensor, a rotating platform, and a computer;

[0066] The sensor is mounted on the dual-axis moving platform;

[0067] The rotating platform is positioned below the sensor and is configured correspondingly to the sensor;

[0068] The computer is electrically connected to the sensor, the dual-axis moving platform, and the rotating platform, respectively.

[0069] The aforementioned deep learning-based method and system for measuring aspherical topography has the following advantages:

[0070] First, it is more efficient and more accurate than traditional center-based localization and deep learning localization methods;

[0071] Secondly, we perform spot feature extraction and feature fusion to obtain advanced semantic information, improve the accuracy of spot center localization and detection, and enhance the network's ability to extract spot features and suppress noise.

[0072] Third, by increasing the weight of the point center loss in the later training stage, the positioning accuracy is further optimized, resulting in higher accuracy of the measured surface topography of the aspherical optical element. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the measurement method in this embodiment;

[0074] Figure 2 This is a schematic diagram of the deep learning network structure in this embodiment;

[0075] Figure 3 This is a schematic diagram of the attention module structure in this embodiment;

[0076] Figure 4 This is a schematic diagram of the positioning attention module structure in this embodiment;

[0077] Figure 5 This is a schematic diagram of the detection module structure in this embodiment;

[0078] Figure 6 This is a schematic diagram of the measurement system in this embodiment;

[0079] Figure 7 This is a comparison diagram of the spot center localization results of each method in this embodiment;

[0080] Figure 8 This is a comparison chart of the accuracy of each method in this embodiment;

[0081] Figure 9 This is a stability comparison chart of the various methods in this embodiment;

[0082] Figure 10 This is a diagram showing the three-dimensional topography measurement results in this embodiment;

[0083] Figure 11 This is the effect diagram of the generated standard Gaussian spot in this embodiment.

[0084] In the diagram: 1. Sensor; 2. Dual-axis moving platform; 3. Rotating platform; 4. Computer. Detailed Implementation

[0085] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0086] A deep learning-based method for measuring aspherical topography, such as Figure 1 As shown, it includes the following steps:

[0087] S1. Prepare a simulation image dataset, and divide the image dataset into a training set and a test set in an 8:2 ratio;

[0088] S2. Load the image dataset into the deep learning network for training and save the optimal weight parameters;

[0089] In this embodiment, the deep learning network of the present invention was trained for 300 epochs, using a batch size of 4 and a monotonically decreasing learning rate from 0.01 to 0.0001;

[0090] The SGD optimizer was chosen for optimization, with momentum decay and weight decay values ​​set to 0.937 and 0.0005, respectively. This optimization strategy helps the model avoid local minima, enhances training diversity, and prevents overfitting to the optimal result.

[0091] S3. Construct the system to acquire multi-spot image data;

[0092] S4. Load the optimal weight parameters into the deep learning network, and then use the deep learning network to locate the center of the multi-spot image data.

[0093] In this embodiment, the PyTorch deep learning framework is used on a PC, and the PyCharm integrated development environment is used to perform center localization on multi-spot image data.

[0094] S5. Use the Fourier transform method to measure the three-dimensional shape of an aspherical object.

[0095] Step S1 also includes:

[0096] S11. Generate a standard Gaussian spot using a two-dimensional Gaussian function;

[0097] S12. Generate standard Gaussian spots at different locations in the image, denoted as a standard Gaussian multi-spot image, and record the center coordinates of each standard Gaussian spot.

[0098] In this embodiment, as Figure 11 As shown, standard Gaussian spots are generated at different locations on the image using Matlab software, and their center coordinates are recorded.

[0099] S13. Add Gaussian noise and salt-and-pepper noise to the standard Gaussian multi-spot image, and denot it as a noisy multi-spot image;

[0100] S14. Using the center coordinates of each standard Gaussian spot as labels and the noisy multi-spot images as corresponding data, divide the noisy multi-spot images into a dataset and a test set in an 8:2 ratio.

[0101] The two-dimensional Gaussian function is:

[0102]

[0103] Where A is the total energy of the light spot, (x0, y0) is the center coordinate of the light spot, and σ x and σ y The standard deviations along the x-axis and y-axis, respectively, determine the width of the light spot in the corresponding directions.

[0104] like Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown, in step S2, the deep learning network includes a backbone, an FPN-Neck part, and a detector part that integrates a localization attention module;

[0105] Step S2 specifically includes the following steps:

[0106] S21. Input the simulated image dataset into the backbone. The input images are processed by the first Focus convolution layer. Each image is divided into four sub-images by taking a value every other pixel. The information of each sub-image is concentrated into the channel space. The input channels are expanded by 4 times to obtain a stitched image with 12 channels.

[0107] The stitched image is then sequentially input into the first CBS module and the second CBS module. The first CBS module uses 16 convolutional kernels of size 3x3 with a stride of 1 to expand the channels of the stitched image to 16 channels. The second CBS module uses 32 convolutional kernels of size 3x3 with a stride of 2 to expand the channels of the stitched image to 32 channels and halve the size of the stitched image.

[0108] The stitched images are processed by module C3 to obtain the feature map Feature_C0;

[0109] S22. Input the feature map Feature_C0 into the third CBS module of the backbone, amplify the shallow features by channel expansion, and then input it into the C3 module for feature learning, outputting the feature map Feature_C1.

[0110] S23. After passing the feature map Feature_C1 through the fourth CBS module and C3 module of the backbone, we obtain the feature map Feature_C2. After passing the feature map Feature_C2 through the fifth CBS module and C3 module of the backbone, we obtain the feature map Feature_C3 through spatial pyramid pooling.

[0111] S24. Using the feature map Feature_C3 as the input to the multi-scale fusion feature FPN-Neck part, it passes through the sixth CBS module of the neural network and undergoes an upsampling operation to obtain the feature map Feature_Up1.

[0112] S25. Feature map Feature_Up1 and feature map Feature_C2 are fused to obtain fused feature map Feature_Fuse1. Then, through module C3 and the seventh CBS module of the neural network, feature map Feature_Up2 is obtained after upsampling.

[0113] S26. Feature map Feature_Up2 and feature map Feature_C1 are fused to obtain fused feature map Feature_Fuse2. Then, through module C3 and the eighth CBS module of the neural network, feature map Feature_Up3 is obtained after upsampling.

[0114] S27. Feature map Feature_Up3 and feature map Feature_C0 are fused to obtain feature map Feature_Fuse3, and then passed through C3 module with a convolution kernel size of 3*3 and 64 channels to obtain feature map Feature_C4.

[0115] S28. Input the feature map Feature_C4 into the detector of the integrated localization attention module to obtain a multidimensional array containing the center coordinates of the spot.

[0116] In this embodiment, as Figure 2 , Figure 3 , Figure 4 , Figure 5 As shown, the specific steps for preparing the simulated image dataset and loading it into the deep learning network for training are as follows:

[0117] A standard Gaussian spot is generated using a two-dimensional Gaussian function. Based on the standard Gaussian spot at different locations in the image, an array of Gaussian spots is generated and the true center coordinates of each array of Gaussian spots are recorded. Based on the above operations, a standard Gaussian multi-spot image is generated.

[0118] To make standard Gaussian multi-spot images closer to realistic imaging and increase the diversity of the dataset, Gaussian noise with a variance of less than or equal to 25 is added to the standard Gaussian multi-spot images to simulate the noise generated by the thermal motion of electronic components and circuit noise in the image sensor. In addition, salt-and-pepper noise with a density of less than 0.025 is added to simulate defective and damaged pixels in the image sensor, as well as pixels that may be lost during transmission. Irregular brightness changes and local information loss are introduced, enabling the model to better learn and process local anomalies in the image.

[0119] Next, the prepared simulation image dataset is used for model training at an 8:2 ratio. The dataset is then loaded into the model for training.

[0120] Model training is mainly divided into three parts: the backbone part responsible for multi-scale feature extraction, the FPN-Neck part that fuses multi-scale features, and the detector part containing the localization attention module. The specific training process is as follows:

[0121] In the main body:

[0122] The input image is segmented into four sub-images through the first Focus convolution layer, which takes a value every other pixel, similar to downsampling. The four sub-images are similar without losing information, and the information is concentrated in the channel space, expanding the input channels by 4 times. That is, the channels of each sub-image are concatenated into 12 channels. Then the image is sequentially input into two CBS modules. The convolutional layer of the first CBS module uses 16 convolutional kernels of size 3x3 with a stride of 1 to expand the image channels to 16 channels. The convolutional layer of the second CBS module uses 32 convolutional kernels of size 3x3 with a stride of 2 to expand the number of channels of the image to 32 channels and halve the image size. Finally, it is processed by the C3 module to obtain the feature map Feature_C0, which achieves the effect of reducing the number of parameters and improving the training speed.

[0123] After passing through the third CBS module, the feature map Feature_C0 enters the C3 module, where channel expansion is used to amplify the shallow features. By expanding the channels of the input features, high-dimensional to low-dimensional feature mapping is achieved, rich shallow information is obtained, and the output feature map Feature_C1 is generated.

[0124] Feature map Feature_C1 is processed by the fourth CBS module and the C3 module to obtain feature map Feature_C2. Then, after processing by the fifth CBS module and the C3 module, it is processed by spatial pyramid pooling to obtain multi-scale feature map Feature_C3.

[0125] In the FPN-Neck section:

[0126] Upsample represents the upsampling operation. The final feature map, Feature_C3, of the backbone responsible for multi-scale feature extraction enters the FPN-Neck part of multi-scale feature fusion. It first passes through the sixth CBS module of the neural network and undergoes an upsampling operation to obtain the feature map, Feature_Up1.

[0127] The feature map Feature_Up1 is fused with the feature map Feature_C2 from the feature extraction part to obtain the fused feature map Feature_Fuse1. The fused feature map Feature_Fuse1 is then passed through the C3 module and the seventh CBS module of the neural network, and then upsampled to obtain the feature map Feature_Up2.

[0128] Feature map Feature_Up2 is fused with feature map Feature_C1 of the extracted feature part to obtain feature map Feature_Fuse2. The above operation is repeated to obtain feature map Feature_Up3 through C3 module, the eighth CBS module of the neural network and upsampling operation.

[0129] Feature map Feature_Up3 and feature map Feature_C0 are fused to obtain feature map Feature_Fuse3, which is then passed through C3 module with a convolution kernel size of 3*3 and 64 channels to obtain feature map Feature_C4.

[0130] In the detector section containing the localization attention module:

[0131] The final feature map, Feature_C4, obtained from the FPN-Neck part enters the localization attention module, and then a multidimensional array containing the center coordinate data of the light spot is obtained.

[0132] like Figure 3 , Figure 4 and Figure 5 As shown, the detector (LAD) with integrated localization attention module includes a localization attention module (LAM) and a detection module (DM);

[0133] The Localization Attention Module (LAM) consists of the Channel Attention Module (CAM) and the Spatial Attention Module (SAM).

[0134] The execution steps of the Channel Attention Module (CAM) are as follows:

[0135] Input feature map Obtain the one-dimensional channel weight matrix One-dimensional channel weight matrix With input feature map Multiplication produces a multi-scale feature map with channel attention.

[0136] The execution steps of the Spatial Attention Module (SAM) are as follows:

[0137] Input feature map Obtain spatial attention feature maps containing location information Feature map With feature map Multiplication yields the feature map Defined as:

[0138]

[0139] The Channel Attention Module (CAM) generates a channel attention weight matrix by analyzing the channel correlations in the feature map. The specific formula is as follows:

[0140]

[0141] F ex (F sq,w)=σ(ω2δ(ω1F sq ))

[0142] F C (F,F ex )=F·F ex

[0143] Where i and j are the feature maps. In the i-th row and j-th column, σ is the sigmoid activation function, δ is the ReLU activation function, and ω1 and ω2 represent the weights of two different fully connected layers;

[0144] The Spatial Attention Module (SAM) draws from the input feature map To extract spatial mapping relationships, specifically, depthwise separable convolutions are used to capture the relevant weights between different regions, preserving channel correlations while reducing computational overhead. The formula is defined as follows:

[0145]

[0146] Among them, DConv i×i This indicates a depthwise separable convolution with a kernel size of i×i;

[0147] After passing through the Localization Attention Module (LAM), the feature map Feature_C4 inputs multi-scale features into the Detection Module (DM).

[0148] feature map The data is input into the CBS module and residual connections are introduced to facilitate cross-layer information flow. Finally, two parallel 1x1 convolutions are used to obtain the regression parameters and IoU values ​​related to the detection target.

[0149] Step S2 also includes: designing the loss function as the center loss function;

[0150] The center loss function is derived from the bounding box loss L. ciou And confidence loss L conf Composition, which is defined as:

[0151] L = L ciou +L conf

[0152] Loss=(1-IoU)×L ciou +(1-IoU)×L ciou +IoU×L Center

[0153]

[0154] in, The sample label is 1 for the point and 0 for the background. P iB is the probability that a sample is predicted as a point. i These are the predicted coordinates of the center of point i. p is the center coordinate of the label corresponding to the i-th point; IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box. 2 (B,B gt ) represents the Euclidean distance between the predicted point and the true point, c is the diagonal length of the smallest outer rectangle containing the predicted and true values, α is the weight parameter, and v is used to measure the similarity of the aspect ratio between the predicted and true values. α and v are specifically defined as follows:

[0155]

[0156] Among them, w gt h gt w and h are the width and height of the ground truth bounding box and the predicted bounding box, respectively.

[0157] Step S3 includes the following steps:

[0158] S31. Install sensor 1 on dual-axis moving platform 2;

[0159] S32. Position the rotating platform 3 below the sensor 1;

[0160] S33. Install computer 4 and electrically connect computer 4 to sensor 1, dual-axis moving platform 2 and rotating platform 3 respectively.

[0161] In practical applications, the object to be tested is placed on the rotating platform 3. The computer 4 controls the dual-axis moving platform 2 to move the sensor 1 above the object to be tested. The computer 4 then controls the rotating platform 3 to rotate the object to be tested. The sensor 1 is used to collect data and obtain multi-spot images. The images are then input into the deep learning network of this invention for prediction to obtain the center data of each spot.

[0162] Step S5 includes the following steps:

[0163] S51. Based on the center coordinates of each spot in the multi-spot image, the angle difference is calculated by using the corresponding spot coordinates on two adjacent multi-spot images.

[0164] S52. Based on the relationship between the Fourier series coefficients of the angle difference and the surface morphology, the three-dimensional morphology information of the aspherical optical element is reconstructed.

[0165] In this embodiment, as Figure 10As shown, the three-dimensional morphology of an aspherical object is measured using the Fourier transform method. Specifically, plano-convex cylindrical lenses with curvature radii of 500 mm, 700 mm, and 1000 mm are used as the workpieces. The angular difference between two points on the aspherical optical surface is obtained using the spot center data. Finally, based on the relationship between the Fourier series coefficients of the angular difference and the surface morphology, the three-dimensional morphology information of the aspherical optical element is reconstructed. The measurement process is described for the surfaces of plano-convex cylindrical lenses with curvature radii of 500 mm, 700 mm, and 1000 mm.

[0166] For reconstructing the deviation PV and STD values ​​of plano-convex cylindrical lenses with different radii of curvature using different spot center positioning methods, analysis shows that for workpieces with curvature radii of curvature of 500mm, 700mm, and 1000mm, the deep learning network of this invention improves the PV value by at least 31.64%, 44.87%, and 31.09% respectively, and the STD value by at least 27.78%, 30.00%, and 25.45% respectively, compared to other methods. The results indicate that this method can accurately measure aspherical optical elements, and the larger the curvature radius of the workpiece and the smaller the curvature change, the higher the measurement accuracy. Detailed results are shown in Table 1.

[0167]

[0168] Table 1

[0169] The performance of various deep learning networks was compared through a series of experiments.

[0170] To evaluate the efficiency of the deep learning network of this invention in locating the center of each point in a multi-point image, a comparative analysis was conducted with typical traditional image processing methods such as the centroid method (GCM), Gaussian fitting method (GFM), circle fitting method (CFM), and geometric method (GM). The deep learning network of this invention processes multi-point images in 20.3 ms, which is 34.82 times that of the centroid method, 35.59 times that of the geometric method, 43.90 times that of the circle fitting method, and 271.36 times that of the Gaussian fitting method.

[0171] To evaluate the localization accuracy of the deep learning network of this invention, it was compared with typical traditional image processing methods. Different methods were used to locate the center of the blob. The Euclidean distance between the obtained center and the accurate center was used as the localization deviation E. d The calculation formula is as follows:

[0172]

[0173] Among them, (x i ,y i(x0, y0) represents the coordinates of the spot center located by different methods, where (x0, y0) is the accurate spot center. Due to spot shape variations caused by noise, the maximum localization error of the Circle Fitting Method (CFM) is 1.372 pixels. The Gaussian Fitting Method (GFM) is sensitive to noise, resulting in a fitting offset and localization error of 0.679 pixels. The Geometric Method (GM) estimates the center from the pixel average, but noise-induced offset leads to a localization error of 0.531 pixels. The Centroid Method (GCM) is sensitive to uneven pixel distribution, with a localization error of 0.223 pixels. In contrast, the deep learning network model of this invention, trained on spot images with various levels of noise, achieves a localization accuracy of 0.052 pixels, effectively resisting various noises while maintaining high accuracy.

[0174] To comprehensively compare the performance of the deep learning network of this invention, it was compared with other convolutional neural networks DSCNet and SCPNet, and also with mainstream object detection algorithms Faster R-CNN, YOLOv5, and YOLOv8. Experiments were conducted using the same dataset under consistent settings. Evaluation metrics included mean bias. Maximum deviation (MD), standard deviation (STD), and processing time (including data transfer time from CPU to GPU). Detailed results are shown in Table 2.

[0175]

[0176]

[0177] Table 2

[0178] The average bias of the deep learning network of this invention The average deviation is 0.052 pixels, the maximum deviation (MD) is 0.063 pixels, the standard deviation (STD) is 0.011 pixels, and the processing time is 20.3ms, demonstrating good positioning accuracy and reliability. YOLOv8 has a shorter processing time, but its center positioning accuracy is poorer. Although the average deviation of DSCNet is... While smaller in size, the process of first segmenting each point in a multi-spot image, then inputting it into the network for processing, and finally calculating the center of each point using Gaussian fitting is time-consuming. Therefore, the deep learning network of this invention has better performance in locating the center of each spot in a multi-spot image.

[0179] like Figure 7As shown in the figure, the spot center localization results of the deep learning network of the present invention are compared with those of different methods. It can be seen from the figure that the deep learning network of the present invention has better localization accuracy and better performance than traditional algorithms such as centroid method (GCM) and Gaussian fitting method (GFM) and deep learning algorithms such as Faster RCNN and YOLOv8.

[0180] like Figure 8 As shown, to evaluate the robustness of the deep learning network of the present invention to the center localization accuracy of light spots of different sizes, a comparative experiment on the center localization accuracy of light spots was conducted within a range of 35 pixels (±10%). In the actual experiment, the size of the light spot fluctuated within a certain range, and the average value of the light spot radius was approximately 35 pixels. To investigate the robustness of the deep learning network of the present invention to the center localization accuracy of light spots of different sizes, a comparative experiment on the center localization accuracy of light spots was conducted within a range of light spot radius of 35 pixels (±10%). The deep learning network of the present invention still maintains good center localization accuracy within this range, demonstrating good robustness.

[0181] like Figure 9 As shown, to evaluate the stability and noise resistance of the deep learning network of this invention, it is compared with commonly used spot center localization algorithms under Gaussian noise (variance of 5-45), salt-and-pepper noise (density of 0.005-0.045), and combined noise of Gaussian noise (variance of 5-45) and salt-and-pepper noise (density of 0.005-0.045), specifically traditional algorithms such as Gaussian fitting (GFM), circle fitting (CFM), and geometric method (GM), as well as deep learning methods such as YOLOv8 and DSCNet.

[0182] As noise levels increase, the localization error of traditional algorithms rises significantly, while deep learning methods exhibit better stability. Further analysis shows that the deep learning network of this invention maintains an error below 0.3 pixels, with a Gaussian noise variance of less than 45 and a salt-and-pepper noise density of less than 0.045, outperforming the deep learning methods YOLOv8 and DSCNet. The deep learning network of this invention demonstrates good stability when the Gaussian noise variance is less than 60 and the salt-and-pepper noise density is less than 0.060. Experimental results show that the deep learning network of this invention can accurately locate the center of the light spot even under high noise conditions, exhibiting good anti-interference capabilities.

[0183] To evaluate the impact of different improvement techniques on the spot center localization accuracy, stability, and computational efficiency of the deep learning network of this invention, ablation experiments were conducted. In the first experiment, a convolutional neural network without a localization attention module (LAM) was used as the baseline model. The average deviation of the baseline model was... The mean deviation was 0.314 pixels, the maximum deviation (MD) was 0.524 pixels, and the standard deviation (STD) was 0.093 pixels, indicating poor stability. In the second experiment, the introduction of a localization attention module (LAM) significantly improved the center localization accuracy, with the mean deviation... The maximum deviation (MD) and standard deviation (STD) decreased by 68.15% and 57.06%, respectively, while the standard deviation (STD) decreased by 34.41%. Although the number of parameters only increased by 1.15%, the computational cost increased by 13.78%, and the processing time increased by 14.04%. In the third experiment, the center loss was used as the loss function of the second model. During later training of the model, the center localization accuracy was further improved by increasing the weight of the center loss. This improved the mean deviation... The maximum deviation (MD) and standard deviation (STD) decreased by 48.00% and 72.00% respectively, and the standard deviation (STD) decreased by 81.97%, further improving the stability of the model. Detailed results are shown in Table 3.

[0184]

[0185] Table 3

[0186] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A deep learning-based method for measuring aspherical topography, characterized in that, Includes the following steps: S1. Prepare a simulation image dataset, and divide the image dataset into a training set and a test set in an 8:2 ratio; S2. Load the image dataset into the deep learning network for training and save the optimal weight parameters; S3. Construct the system to acquire multi-spot image data; S4. Load the optimal weight parameters into the deep learning network, and then use the deep learning network to perform center localization on the multi-spot image data; S5. Use the Fourier transform method to measure the three-dimensional shape of an aspherical object; In step S2, the deep learning network includes a backbone, an FPN-Neck, and a detector part that integrates a localization attention module; Step S2 specifically includes the following steps: S21. Input the simulated image dataset into the backbone. The input images are processed by the first Focus convolution layer. Each image is divided into four sub-images by taking a value every other pixel. The information of each sub-image is concentrated into the channel space. The input channels are expanded by 4 times to obtain a stitched image with 12 channels. The stitched image is then sequentially input into the first CBS module and the second CBS module. The first CBS module uses 16 convolutional kernels of size 3x3 with a stride of 1 to expand the channels of the stitched image to 16 channels. The second CBS module uses 32 convolutional kernels of size 3x3 with a stride of 2 to expand the channels of the stitched image to 32 channels and halve the size of the stitched image. The stitched image is processed by the C3 module to obtain the feature map Feature_C0; S22. Input the feature map Feature_C0 into the third CBS module of the backbone, amplify the shallow features by channel expansion, and then input it into the C3 module for feature learning, and output the feature map Feature_C1. S23. The feature map Feature_C1 is passed through the fourth CBS module and C3 module of the main body to obtain the feature map Feature_C2. The feature map Feature_C2 is passed through the fifth CBS module and C3 module of the main body, and then through spatial pyramid pooling to obtain the feature map Feature_C3. S24. Using the feature map Feature_C3 as input to the multi-scale fusion feature FPN-Neck part, the feature map Feature_Up1 is obtained after passing through the sixth CBS module of the deep learning network and undergoing an upsampling operation. S25. The feature map Feature_Up1 and the feature map Feature_C2 are fused to obtain the fused feature map Feature_Fuse1. Then, through the C3 module and the seventh CBS module of the deep learning network, the feature map Feature_Up2 is obtained by upsampling. S26. The feature map Feature_Up2 and the feature map Feature_C1 are fused to obtain the fused feature map Feature_Fuse2. Then, through the C3 module and the eighth CBS module of the deep learning network, the feature map Feature_Up3 is obtained by upsampling. S27. The feature map Feature_Up3 and the feature map Feature_C0 are fused to obtain the feature map Feature_Fuse3, and then passed through the C3 module with a convolution kernel size of 3*3 and a channel number of 64 to obtain the feature map Feature_C4. S28. Input the feature map Feature_C4 into the detector of the integrated localization attention module to obtain a multidimensional array containing the center coordinates of the light spot; The detector of the integrated localization attention module includes a localization attention module and a detection module; The positioning attention module consists of a channel attention module and a spatial attention module; The execution steps of the channel attention module are as follows: Input feature map The one-dimensional channel weight matrix is ​​obtained. The one-dimensional channel weight matrix With input feature map Multiplication produces a multi-scale feature map with channel attention. ; The execution steps of the spatial attention module are as follows: Input the feature map This yields a spatial attention feature map containing location information. The feature map With the feature map Multiplication yields the feature map Defined as: The channel attention module generates a channel attention weight matrix by analyzing the channel correlations in the feature map. The specific formula is as follows: in, , These are the feature maps The row and number List, It is the sigmoid activation function. It is the ReLU activation function. and These represent the weights of two different fully connected layers; The spatial attention module receives input feature maps To extract spatial mapping relationships, specifically, depthwise separable convolutions are used to capture the relevant weights between different regions, preserving channel correlations while reducing computational overhead. The formula is defined as follows: in, Indicates the kernel size as Depth-separable convolution; After passing through the localization attention module, the feature map Feature_C4 inputs multi-scale features into the detection module; feature map The input is incorporated into the CBS module, and residual connections are introduced to facilitate cross-layer information flow. Finally, regression parameters related to the detection target are obtained through two parallel 1x1 convolutions. value.

2. The aspherical shape measurement method based on deep learning according to claim 1, characterized in that, Step S1 further includes: S11. Generate a standard Gaussian spot using a two-dimensional Gaussian function; S12. Generate the standard Gaussian spots at different positions in the image, denoted as a standard Gaussian multi-spot image, and record the center coordinates of each standard Gaussian spot; S13. Add Gaussian noise and salt-and-pepper noise to the standard Gaussian multi-spot image, and record it as a noisy multi-spot image; S14. Using the center coordinates of each of the standard Gaussian spots as labels and the noisy multi-spot images as corresponding data, the noisy multi-spot images are divided into a dataset and a test set in an 8:2 ratio.

3. The aspherical shape measurement method based on deep learning according to claim 2, characterized in that, The two-dimensional Gaussian function is: in, The total energy of the light spot, ( , () represents the center coordinates of the light spot. and The standard deviations along the x-axis and y-axis, respectively, determine the width of the light spot in the corresponding directions.

4. The aspherical shape measurement method based on deep learning according to claim 1, characterized in that, Step S2 further includes: designing the loss function as the center loss function; The central loss function is composed of the bounding box loss. and confidence loss Composition, which is defined as: in, The sample label is 1 for the dot and 0 for the background. It is the probability that a sample is predicted as a point. It is the first Predicted coordinates of the point center These are the center coordinates of the label corresponding to point i; The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. This represents the Euclidean distance between the predicted point and the actual point. The length of the diagonal of the smallest outer rectangle containing both the predicted and actual values. For weight parameters, This is used to measure the similarity between the aspect ratios of predicted and actual values. and Specifically defined as: in, , and These represent the width and height of the ground truth bounding box and the predicted bounding box, respectively.

5. The aspherical shape measurement method based on deep learning according to claim 1, characterized in that, Step S3 includes the following steps: S31. Install the sensor on the dual-axis moving platform; S32. Position the rotating platform below the sensor; S33. Install the computer and electrically connect it to the sensor, the dual-axis moving platform, and the rotating platform respectively.

6. The aspherical shape measurement method based on deep learning according to claim 1, characterized in that, Step S5 includes the following steps: S51. Based on the center coordinates of each spot in the multi-spot image, the angle difference is calculated by the corresponding spot coordinate positions on two adjacent multi-spot images. S52. Based on the relationship between the Fourier series coefficients of the angle difference and the surface morphology, reconstruct the three-dimensional morphology information of the aspherical optical element.

7. An aspherical topography measurement system for implementing the deep learning-based aspherical topography measurement method according to any one of claims 1-6, characterized in that: Includes a dual-axis moving platform, sensors, a rotating platform, and a computer; The sensor is mounted on the dual-axis moving platform; The rotating platform is positioned below the sensor and is configured correspondingly to the sensor; The computer is electrically connected to the sensor, the dual-axis moving platform, and the rotating platform, respectively.

Images