Deep learning-based method for extracting center of light stripe in complex background under online structured light

Abstract

This invention provides a deep learning-based method for extracting the center of structured light stripes under complex backgrounds, relating to the field of industrial vision measurement. Specifically, it includes the following steps: acquiring workpiece images with different materials, surface reflectivity, and morphology under various measurement conditions to obtain images containing light stripes; creating a light stripe region annotation dataset from the images using open-source software; constructing a semantic segmentation deep learning model and training it using the dataset to obtain a trained semantic segmentation deep learning model; applying the trained semantic segmentation deep learning model to segment the image under test; and combining an improved Steger algorithm for sub-pixel level extraction of the light stripe center. The technical solution of this invention overcomes the problems of decreased accuracy, broken center lines, and unstable extraction in existing technologies under conditions of environmental noise, workpiece surface reflection, and complex surface morphology.

Description

Technical Field

[0001] This invention relates to the field of industrial vision measurement, specifically to a method for extracting the center of an upper structured light stripe against a complex background based on deep learning. Background Technology

[0002] Line structured light measurement methods based on optical triangulation offer advantages such as simple structure, fast measurement speed, and high accuracy, and are widely used in dimensional measurement and surface topography reconstruction of industrial parts. Accurate extraction of the light stripe center is a key technology for ensuring the performance of a line structured light measurement system; its accuracy directly affects the accuracy and reliability of the measurement results. Existing sub-pixel level extraction methods for light stripe centers mainly include the gray-scale centroid method and the Steger algorithm. The gray-scale centroid method is computationally simple and fast, but it struggles to guarantee extraction accuracy when there is high noise or complex surface topography of the measured object. The Steger algorithm can estimate the line direction and has some adaptability to surface undulations, but it is prone to extraction instability when the light stripe width varies significantly, and its multiple convolution calculations result in a slower processing speed.

[0003] Therefore, there is a need for a deep learning-based method for extracting the center of an online structured light stripe under complex background conditions, which can stably extract the center of the light stripe with high accuracy under conditions of environmental noise, workpiece surface reflection and complex surface morphology. Summary of the Invention

[0004] The main objective of this invention is to provide a method for extracting the center of a structured light stripe under complex backgrounds based on deep learning, in order to solve the problems of decreased extraction accuracy, centerline breakage, and unstable extraction in the prior art under conditions of environmental noise, workpiece surface reflection, and complex surface morphology.

[0005] To achieve the above objectives, this invention provides a method for extracting the center of an online structured light stripe against a complex background based on deep learning, specifically including the following steps:

[0006] S1 acquires images of workpieces with different materials, surface reflection characteristics, and morphology under various measurement conditions, and obtains images containing light stripes.

[0007] S2. The images obtained in step S1 are used to create a light bar region annotation dataset using the open-source software Labelme, and then divided into training set, validation set and test set according to a certain ratio.

[0008] S3. Construct a semantic segmentation deep learning model and train it using the dataset created in step S2 to obtain a trained semantic segmentation deep learning model.

[0009] S4. Apply the trained semantic segmentation deep learning model to segment the image under test.

[0010] S5. Using the image obtained in step S4 as input, the improved Steger algorithm is used to extract the center of the light stripe at the sub-pixel level.

[0011] Furthermore, step S2 specifically includes the following steps:

[0012] S2.1, The light stripe image obtained in step S1 is labeled with regions, wherein the part of the workpiece surface irradiated by the laser is labeled as the light stripe, and the rest is labeled as the background.

[0013] S2.2, and save the annotation results as PNG format images. Each sample contains an original light stripe image and its corresponding light stripe mask image, which are used for training, validation and testing of the semantic segmentation deep learning model.

[0014] Furthermore, the semantic segmentation deep learning model in step S3 includes an encoding path and a decoding path, wherein the decoding path is symmetrical to the encoding path.

[0015] Furthermore, the encoding path comprises five layers. The first layer of the encoding path consists of two 3×3 convolutional layers, each followed by BN and ReLU activation functions, then embedding a CA attention module, and finally achieving downsampling through 2×2 max pooling. The second layer of the encoding path has the same structure as the first layer. The third layer of the encoding path consists of three 3×3 convolutional layers, each followed by BN and ReLU activation functions, then embedding a CA attention module, and finally achieving downsampling through 2×2 max pooling. The fourth layer of the encoding path has the same structure as the third layer. The fifth layer of the encoding path consists of three 3×3 convolutional layers, each followed by BN and ReLU activation functions.

[0016] Furthermore, the decoding path comprises five layers. In the fifth layer of the decoding path, the features output from the fifth layer of the encoding path are upsampled, and then adjusted for the number of channels via a 1×1 convolution. In the fourth layer of the decoding path, the output from the fifth layer of the decoding path is fused with the features from the fourth layer of the encoding path, followed by two 3×3 convolutions, each followed by a BN and ReLU activation function, then upsampled again and adjusted for the number of channels via a 1×1 convolution. In the third layer of the decoding path, the output from the fourth layer of the decoding path is fused with the features from the third layer of the encoding path, followed by two 3×3 convolutions, each followed by a BN. The signal is first processed with a ReLU activation function, then upsampled and followed by a 1×1 convolution to adjust the number of channels. In the second layer of the decoding path, the result of the third layer of the decoding path is fused with the features of the second layer of the encoding path, followed by two 3×3 convolutions, each followed by a BN and ReLU activation function, then upsampled and followed by a 1×1 convolution to adjust the number of channels. In the first layer of the decoding path, the result of the second layer of the decoding path is fused with the features of the first layer of the encoding path, followed by two 3×3 convolutions, each followed by a BN and ReLU activation function, and finally output as the final segmentation result image through a 1×1 convolution.

[0017] Furthermore, during the training process in step S3, the loss function Using binary cross-entropy loss:

[0018] ;

[0019] in, This represents the output value of the semantic segmentation deep learning model. This represents the Sigmoid function. This is the tag value.

[0020] Furthermore, in step S3, the accuracy is evaluated using the Dice coefficient:

[0021] ;

[0022] Where X represents the prediction result and Y represents the label.

[0023] Furthermore, step S5 specifically includes the following steps:

[0024] S5.1, Extract the Region of Interest (ROI) from the image obtained in step S4, and calculate the centerline only within the effective area of ​​the light stripe.

[0025] S5.2, use the Steger algorithm within the ROI to obtain the initial center point of the light stripe.

[0026] S5.3, Non-maximum suppression (NMS) is performed on the center points of the light stripe obtained in step S5.2 to remove redundant and isolated noise points; a local comparison window is established along the normal direction of the light stripe, and the gray values ​​of the center points within the window are compared, retaining only the local maximum points. The NMS process is expressed as follows:

[0027] ;

[0028] in, For the first The gray intensity of each candidate center point It is half the width of the local comparison window.

[0029] S5.4, Perform Akima interpolation on the center points filtered by NMS in step S5.3 to obtain the interpolation curve.

[0030] S5.5, Savitzky-Golay filtering is applied to the interpolation curve obtained in step S5.4, and the curve is smoothed by fitting a local polynomial through a sliding window, further eliminating high-frequency noise.

[0031] S5.6, using the standard deviation σ to evaluate the accuracy of the smooth curve extraction obtained in S5.5:

[0032] ;

[0033] in, , Indicates the first The center point of each light bar For the first The coordinates of the center point The total number of extracted center points, The slope of the line obtained by fitting the center point using the least squares method. It is the intercept of the line.

[0034] Furthermore, step S5.2 specifically includes the following steps:

[0035] S5.2.1 Perform two-dimensional Gaussian differential convolution on the ROI image obtained in step S5.1, and calculate the second-order partial derivative of each pixel.

[0036] S5.2.2, construct the Hessian matrix, and determine the direction of the light stripe normal based on the eigenvalues ​​and eigenvectors of the Hessian matrix.

[0037] S5.2.3 Extract the local grayscale distribution along the normal direction and perform a second-order Taylor expansion on the grayscale distribution.

[0038] S5.2.4, by solving for the points where the first derivative of the function is zero, the extreme position of the light stripe grayscale is obtained, that is, the sub-pixel center point.

[0039] The present invention has the following beneficial effects:

[0040] The improved U-Net model used in this invention can enhance the robustness and accuracy of light stripe segmentation under complex lighting conditions, workpiece surface reflection, and complex surface morphology. Testing showed that the Dice coefficient of the model in this invention reached 90.71%, an improvement of approximately 2% compared to the original U-Net model's 88.74%, while maintaining the integrity of the light stripe structure and improving stripe continuity.

[0041] In the centerline extraction stage, this invention optimizes the Steger algorithm, combining ROI, NMS, Akima interpolation, and Savitzky-Golay smoothing algorithms to process the centerline, resulting in more continuous extraction results and enhanced noise resistance. Verification shows that compared to the Steger algorithm, the standard deviation of extraction accuracy decreased from 0.355 to 0.302, a reduction of approximately 15%, effectively improving the stability and accuracy of light stripe centerline extraction.

[0042] By combining a semantic segmentation deep learning model with an optimized Steger centerline extraction method, this invention can achieve high-precision extraction of the light stripe center in complex backgrounds and environments with multiple interferences, with center positioning accuracy reaching the sub-pixel level. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0044] Figure 1 The flowchart illustrates a method for extracting the center of an online structured light stripe in a complex background based on deep learning, according to the present invention.

[0045] Figure 2 A schematic diagram of a structured light measurement system is shown.

[0046] Figure 3 The image shows a light stripe image acquired synchronously using an industrial camera.

[0047] Figure 4 A label image produced according to the present invention is shown.

[0048] Figure 5 A schematic diagram of the semantic segmentation model structure of the present invention is shown.

[0049] Figure 6 The image to be tested is shown.

[0050] Figure 7 The image shows the light stripe segmentation result obtained after step S4.

[0051] Figure 8 The image shows the center point of the light stripe extracted using the Steger algorithm.

[0052] Figure 9 The center point of the light stripe extracted using the method of the present invention is shown. Detailed Implementation

[0053] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figure 1The method for extracting the center of an upper structured light stripe in a complex background based on deep learning, shown below, includes the following steps:

[0055] S1 acquires images of workpieces with different materials, surface reflection characteristics, and morphologies under various measurement conditions, obtaining images containing light streaks. For example... Figure 2 As shown, during the acquisition process, the laser emits line structured light that illuminates the workpiece surface, and the industrial camera simultaneously acquires the light stripe image, such as... Figure 3 As shown, the acquired images are then transmitted to the computer processing module.

[0056] S2. The images obtained in step S1 are used to create a light bar region annotation dataset using the open-source software Labelme, and then divided into training set, validation set and test set according to a certain ratio.

[0057] S3. Construct a semantic segmentation deep learning model and train it using the dataset created in step S2 to obtain a trained semantic segmentation deep learning model.

[0058] S4. Apply the trained semantic segmentation deep learning model to segment the image under test.

[0059] S5. Using the image obtained in step S4 as input, the improved Steger algorithm is used to perform sub-pixel level precise extraction of the light stripe center.

[0060] Specifically, step S2 includes the following steps:

[0061] S2.1, The light stripe image obtained in step S1 is labeled with regions, wherein the part of the workpiece surface irradiated by the laser is labeled as the light stripe, and the rest is labeled as the background.

[0062] S2.2, and save the annotation results as a PNG image, such as Figure 4 As shown, each sample contains an original light stripe image and its corresponding light stripe mask image, used for training, validation, and testing of the semantic segmentation deep learning model. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0063] Specifically, the semantic segmentation deep learning model in step S3 includes an encoding path and a decoding path, wherein the decoding path is symmetrical to the encoding path.

[0064] Specifically, such as Figure 5As shown, the encoding path consists of five layers. The first layer of the encoding path includes two 3×3 convolutional layers. Each convolutional layer is followed by BN and ReLU activation functions, then a CA attention module (Coordinate attention) is embedded, and downsampling is achieved through 2×2 max pooling. The input feature map of this layer has a size of 512×512 and a channel number of 1; the output feature map has a size of 256×256 and a channel number of 64.

[0065] The second layer of the encoding path has the same structure as the first layer. The size of the input feature map in this layer is 256×256, and the number of channels is 64. The size of the output feature map is 128×128, and the number of channels is 128.

[0066] The third layer of the encoding path consists of three 3×3 convolutional layers. Each convolutional layer is followed by BN and ReLU activation functions, then a CA attention module is embedded, and downsampling is achieved through 2×2 max pooling. The input feature map of this layer has a size of 128×128 and 128 channels; the output feature map has a size of 64×64 and 256 channels.

[0067] The fourth layer of the encoding path has the same structure as the third layer. The input feature map of this layer has a size of 64×64 and 256 channels; the output feature map has a size of 32×32 and 512 channels.

[0068] The fifth layer of the encoding path consists of three 3×3 convolutions, each followed by BN and ReLU activation functions. The input feature map of this layer has a size of 32×32 and 512 channels; the output feature map has a size of 32×32 and 1024 channels.

[0069] Specifically, such as Figure 5 As shown, the decoding path consists of five layers. In the fifth layer of the decoding path, the features output by the fifth layer of the encoding path are upsampled, and then adjusted by a 1×1 convolution (Conv1×1). The size of the input feature map in this layer is 32×32, and the number of channels is 1024. The size of the output feature map is 32×32, and the number of channels is 512.

[0070] In the fourth layer of the decoding path, the output of the fifth layer of the decoding path is fused with the features of the fourth layer of the encoding path. This is done through two 3×3 convolutions (Conv3×3), each followed by BN and ReLU activation functions, then upsampling, and finally a 1×1 convolution to adjust the number of channels. The input feature map of this layer has a size of 32×32 and 1024 channels; the output feature map has a size of 64×64 and 512 channels.

[0071] In the third layer of the decoding path, the output of the fourth layer of the decoding path is fused with the features of the third layer of the encoding path. This is done through two 3×3 convolutions, each followed by BN and ReLU activation functions, then upsampling, and finally a 1×1 convolution to adjust the number of channels. The input feature map of this layer has a size of 64×64 and 512 channels. The output feature map has a size of 128×128 and 256 channels.

[0072] In the second layer of the decoding path, the results of the third layer of the decoding path are fused with the features of the second layer of the encoding path. This is done through two 3×3 convolutions, each followed by BN and ReLU activation functions, then upsampling, and finally a 1×1 convolution to adjust the number of channels. The input feature map of this layer has a size of 128×128 and 256 channels; the output feature map has a size of 256×256 and 128 channels.

[0073] In the first layer of the decoding path, the results of the second layer of the decoding path are fused with the features of the first layer of the encoding path (i.e., Figure 5 The process involves copying and concatenating the data, followed by two 3×3 convolutions, each followed by BN and ReLU activation functions, and finally outputting the final segmentation result image through a 1×1 convolution. The input feature map of this layer has a size of 256×256 and 128 channels; the output feature map has a size of 512×512 and 1 channel.

[0074] Specifically, during the training process in step S3, the loss function Using binary cross-entropy loss:

[0075] ;

[0076] in, This represents the output value of the semantic segmentation deep learning model. This represents the Sigmoid function. The value is the label. The optimizer uses Adam with a learning rate of 1×10⁻5, a batch size of 4, and 600 training epochs.

[0077] Specifically, in step S3, the accuracy is evaluated using the Dice coefficient:

[0078] ;

[0079] Where X represents the prediction result and Y represents the label.

[0080] Specifically, in step S4, the light stripe region is segmented as follows: Figure 6 The image to be tested is input into a trained semantic segmentation deep learning model, which outputs the light stripe segmentation result. The model can automatically remove complex backgrounds, reflection interference, and non-light stripe regions, retaining only the light stripe pixels. The segmentation result is as follows: Figure 7As shown.

[0081] Specifically, step S5 includes the following steps:

[0082] S5.1, Extract the Region of Interest (ROI) from the image obtained in step S4, and calculate the centerline only within the effective area of ​​the light stripe. This step can significantly reduce the amount of computation.

[0083] S5.2, use the Steger algorithm within the ROI to obtain the initial center point of the light stripe.

[0084] The specific process of the Steger algorithm in step S5.2 is as follows:

[0085] S5.2.1 Perform two-dimensional Gaussian differential convolution on the ROI image obtained in step S5.1, and calculate the second-order partial derivative of each pixel.

[0086] S5.2.2, construct the Hessian matrix, and determine the direction of the light stripe normal based on the eigenvalues ​​and eigenvectors of the Hessian matrix.

[0087] S5.2.3 Extract the local grayscale distribution along the normal direction and perform a second-order Taylor expansion on the grayscale distribution.

[0088] S5.2.4, by solving for the point where the first derivative of the function is zero, the extreme position of the grayscale of the light stripe is obtained, that is, the sub-pixel center point, thus achieving sub-pixel level positioning.

[0089] S5.3, Non-maximum suppression (NMS) is performed on the center points of the light stripe obtained in step S5.2 to remove redundant and isolated noise points. A local comparison window is established along the normal direction of the light stripe, and the gray values ​​of the center points within the window are compared, retaining only the local maxima. The NMS process is expressed as follows:

[0090] ;

[0091] in, For the first The gray intensity of each candidate center point The width of the local comparison window is set to half, and w = 4. This step effectively removes isolated points and erroneous points, improving the continuity and accuracy of the centerline.

[0092] S5.4, Akima interpolation is performed on the center points after NMS filtering in step S5.3 to obtain the interpolation curve; this method achieves smooth transition and shape preservation of the curve by establishing a piecewise cubic polynomial function between adjacent sampling points and performing weighted calculation based on the slope of adjacent intervals.

[0093] S5.5, Savitzky-Golay filtering is applied to the interpolation curve obtained in step S5.4. A local polynomial is fitted through a sliding window to smooth the curve and further eliminate high-frequency noise. Savitzky-Golay filtering fits a local polynomial within the sliding window to smooth the curve while preserving peak values ​​and geometric features. The window length and polynomial order can be adaptively set according to the noise level; in this embodiment, they are set to 7 and 3, respectively. After processing, a continuous, smooth light stripe center curve with sub-pixel accuracy is obtained.

[0094] Extraction results using only the Steger algorithm are as follows Figure 8 As shown, the results extracted using the method of the present invention are as follows: Figure 9 As shown.

[0095] S5.6, using the standard deviation σ to evaluate the accuracy of the smooth curve extraction obtained in S5.5:

[0096] ;

[0097] in, , Indicates the first The center point of each light bar For the first The coordinates of the center point The total number of extracted center points, The slope of the line obtained by fitting the center point using the least squares method. It is the intercept of the line.

[0098] Testing and verification show that the Dice coefficient of the semantic segmentation deep learning model provided by this invention is 90.71%, which is about 2% higher than the 88.74% of the original U-Net model. It can maintain the integrity and continuity of the light stripe structure under complex lighting and strong reflection conditions. Furthermore, when extracting the centerline of the segmentation results, the standard deviation of the traditional Steger algorithm is 0.355, while the standard deviation of the method of this invention is 0.302, a reduction of about 15%. This invention can effectively improve the stability and accuracy of light stripe centerline extraction.

[0099] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for extracting the center of an online structured light stripe against a complex background based on deep learning, characterized in that, Specifically, the steps include the following: S1, under various measurement conditions, acquire images of workpieces with different materials, surface reflection characteristics and morphology, and obtain images containing light stripes; S2, the images obtained in step S1 are used to create a light bar region annotation dataset using the open-source software Labelme, and then divided into training set, validation set and test set according to a certain ratio; S3. Construct a semantic segmentation deep learning model and train it using the dataset created in step S2 to obtain a trained semantic segmentation deep learning model. S4. Apply the trained semantic segmentation deep learning model to segment the image under test; S5, using the image obtained in step S4 as input, perform sub-pixel level extraction of the light stripe center using the improved Steger algorithm; The semantic segmentation deep learning model in step S3 includes: an encoding path and a decoding path, wherein the decoding path is symmetrical to the encoding path; Step S5 specifically includes the following steps: S5.1, Extract the region of interest (ROI) from the image obtained in step S4, and calculate the centerline only within the effective area of ​​the light stripe; S5.2, use the Steger algorithm within the ROI to obtain the initial center point of the light stripe; S5.3, Non-maximum suppression (NMS) is performed on the center points of the light stripe obtained in step S5.2 to remove redundant points and isolated noise points; a local comparison window is established along the normal direction of the light stripe, and the gray values ​​of the center points within the window are compared, retaining only the local maximum points; the NMS process is expressed as follows: ； in, For the first The gray intensity of each candidate center point The width is half the width of the local comparison window; S5.4, Perform Akima interpolation on the center points filtered by NMS in step S5.3 to obtain the interpolation curve; S5.5, Apply Savitzky-Golay filtering to the interpolation curve obtained in step S5.4, and use a sliding window to fit a local polynomial to smooth the curve and further eliminate high-frequency noise; S5.6, using the standard deviation σ to evaluate the accuracy of the smooth curve extraction obtained in S5.5: ； in, , Indicates the first The center point of each light bar For the first The coordinates of the center point The total number of extracted center points, The slope of the line obtained by fitting the center point using the least squares method. It is the intercept of the line.

2. The method for extracting the center of an online structured light stripe under complex backgrounds based on deep learning, as described in claim 1, is characterized in that... Step S2 specifically includes the following steps: S2.1, Mark the regions of the light stripe image obtained in step S1, wherein the part of the workpiece surface irradiated by the laser is marked as the light stripe, and the rest is marked as the background; S2.2, and save the annotation results as PNG format images. Each sample contains an original light stripe image and its corresponding light stripe mask image, which are used for training, validation and testing of the semantic segmentation deep learning model.

3. The method for extracting the center of an online structured light stripe under complex backgrounds based on deep learning according to claim 1, characterized in that, The encoding path consists of five layers. The first layer consists of two 3×3 convolutional layers, each followed by BN and ReLU activation functions, then embedding a CA attention module, and finally achieving downsampling through 2×2 max pooling. The second layer has the same structure as the first layer. The third layer consists of three 3×3 convolutional layers, each followed by BN and ReLU activation functions, then embedding a CA attention module, and finally achieving downsampling through 2×2 max pooling. The fourth layer has the same structure as the third layer. The fifth layer consists of three 3×3 convolutional layers, each followed by BN and ReLU activation functions.

4. The method for extracting the center of an online structured light stripe under complex backgrounds based on deep learning according to claim 1, characterized in that, The decoding path consists of five layers. In the fifth layer of the decoding path, the features output from the fifth layer of the encoding path are upsampled and then adjusted by a 1×1 convolution. In the fourth layer of the decoding path, the output from the fifth layer of the decoding path is fused with the features from the fourth layer of the encoding path. This is done by two 3×3 convolutions, each followed by BN and ReLU activation functions, then upsampled and then adjusted by a 1×1 convolution. In the third layer of the decoding path, the output of the fourth layer of the decoding path is fused with the features of the third layer of the encoding path, followed by two 3×3 convolutions, each followed by BN and ReLU activation functions, then upsampled and then adjusted by a 1×1 convolution; in the second layer of the decoding path, the result of the third layer of the decoding path is fused with the features of the second layer of the encoding path, followed by two 3×3 convolutions, each followed by BN and ReLU activation functions, then upsampled and then adjusted by a 1×1 convolution. In the first layer of the decoding path, the results of the second layer of the decoding path are fused with the features of the first layer of the encoding path. After two 3×3 convolutions, each convolution is followed by BN and ReLU activation functions. Finally, the final segmentation result image is output through a 1×1 convolution.

5. The method for extracting the center of an online structured light stripe against a complex background based on deep learning, as described in claim 1, is characterized in that... During the training process in step S3, the loss function Using binary cross-entropy loss: ； in, This represents the output value of the semantic segmentation deep learning model. This represents the Sigmoid function. This is the tag value.

6. The method for extracting the center of an online structured light stripe under complex backgrounds based on deep learning according to claim 1, characterized in that, In step S3, the accuracy is evaluated using the Dice coefficient: ； Where X represents the prediction result and Y represents the label.

7. The method for extracting the center of an upper structured light stripe in a complex background based on deep learning according to claim 1, characterized in that, Step S5.2 specifically includes the following steps: S5.2.1 Perform two-dimensional Gaussian differential convolution on the ROI image obtained in step S5.1, and calculate the second-order partial derivative of each pixel; S5.2.2, Construct the Hessian matrix, and determine the direction of the light stripe normal based on the eigenvalues ​​and eigenvectors of the Hessian matrix; S5.2.3 Extract the local grayscale distribution along the normal direction and perform a second-order Taylor expansion on the grayscale distribution; S5.2.4, by solving for the points where the first derivative of the function is zero, the extreme position of the light stripe grayscale is obtained, that is, the sub-pixel center point.

Images