Line laser 3D contourgraph working method integrating multiple stray light suppression algorithms

Abstract

The invention discloses a line laser 3D contourgraph working method integrating various stray light suppression algorithms, and the method comprises the steps: constructing a stray light sensitivity weight map based on a multi-exposure image sequence through the combination of a deep convolutional neural network and a point spread function; guiding Gaussian Laplacian pyramid adaptive filtering by using the weight map, and adjusting the filtering intensity by using a residual stability index to generate a purification intensity image; performing one-dimensional profile analysis based on the purified image and the original height map, calculating a morphological distortion index, extracting an energy concentration ratio through three-level wavelet packet decomposition, and triggering morphological restoration; performing center line extraction and motion blur correlation degree calculation on the repaired image, and performing anisotropic smoothing on the height map when a threshold value is exceeded; and registering the processed multi-frame height map, constructing geometric fidelity as a feedback driving genetic algorithm to optimize an initial parameter, and restarting a whole process until the fidelity meets a precision requirement. According to the invention, the precision of 3D imaging is improved through various stray light suppression algorithms.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of 3D measurement, in particular to a working method of line laser 3D profiler integrating multiple stray light suppression algorithms. BACKGROUND

[0002] One of the main challenges faced by line laser 3D profilers in practical applications is stray light interference, which includes environmental light, multi-path reflection and non-target area scattered light, etc. Stray light not only increases image noise, reduces signal-to-noise ratio, affects height measurement accuracy, but also causes false edges, stripe artifacts and other abnormal structures, which mislead center line extraction and three-dimensional reconstruction. Especially in dynamic scanning scenarios, stray light coupled with motion easily produces time-varying artifacts and dynamic blur, which seriously damages the geometric reality and stability of the reconstruction results. There are many technologies that try to solve the above problems, but they each have limitations. Given that a single method cannot comprehensively and effectively solve the stray light problem, it is necessary to integrate multiple advanced algorithms. Based on this, the present application proposes a working method of line laser 3D profiler integrating multiple stray light suppression algorithms. SUMMARY

[0003] The present application provides a working method of line laser 3D profiler integrating multiple stray light suppression algorithms, comprising: S10, based on the multi-exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the best exposure interval are predicted through a deep convolutional neural network, and a stray light sensitivity weight map is constructed in combination with a point spread function; S20, using the weight map to perform adaptive threshold filtering on each layer image after Gaussian Laplacian pyramid decomposition, and counting the stable index of each layer after filtering the residual characteristics, the stable index is used as an energy attenuation coefficient to dynamically adjust the filtering strength of the next layer, and a purified intensity image is generated while optimizing the edge retention rate; S30, taking the purified intensity image and the original height map as input, using one-dimensional profile analysis to calculate the sharpness gradient and background flatness to obtain a morphological distortion index sequence, performing three-level wavelet packet decomposition on the sequence, extracting the energy concentration degree, and using the energy concentration degree as a criterion to determine whether to start a row-by-row stripe repair program based on morphological opening and closing operation; S40, performing center line extraction on the stripe image processed by the repair program, calculating the motion blur correlation degree index using a three-dimensional reconstruction algorithm, and if the index exceeds a preset dynamic threshold, using an inverse diffusion equation to perform anisotropic smoothing on the current frame height map to eliminate artifacts caused by dynamic stray light; S50, registering the processed multi-frame height maps based on an iterative closest point method, extracting a normal vector consistent cosine mean, generating a geometric fidelity by combining a global height fluctuation coefficient, and taking the geometric fidelity as a feedback signal to drive a genetic algorithm to optimize initial weights and restart the whole process until the geometric fidelity converges to its theoretical peak value and meets a preset accuracy requirement.

[0004] The working method of the line laser 3D profiler integrating multiple stray light suppression algorithms as described above, wherein, based on the multi-exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the optimal exposure interval are predicted by a deep convolutional neural network, and a stray light sensitivity weight map is constructed combining a point spread function. Specifically, the following sub-steps are included: After extracting the original height map from the original image sequence and denoising and correcting, the optimal exposure interval is determined based on the multi-exposure response and physical noise, the signal-to-noise ratio gain factor and the local intensity asymmetry are calculated, and the two are jointly predicted and optimized by a double-branch convolutional neural network to realize accurate quantification of laser stripe quality and stray light interference; Combined with physical optics simulation and experimental calibration, the imaging light path of the line laser 3D profiler is accurately mathematically characterized to construct a point spread function, and the function is used to simulate the light energy diffusion effect caused by stray light; Based on the signal-to-noise ratio gain factor and the local intensity asymmetry, a stray light sensitivity weight map is constructed combining a point spread function.

[0005] The working method of the line laser 3D profiler integrating multiple stray light suppression algorithms as described above, wherein, the weight map is used to perform adaptive threshold filtering on each layer image after Gaussian-Laplacian pyramid decomposition, and the stable index of each layer after filtering is calculated, the stable index is taken as an energy attenuation coefficient to dynamically adjust the filtering strength of the next layer, and a purified intensity image is generated while optimizing the edge retention rate. Specifically, the following sub-steps are included: The original intensity image is decomposed by Gaussian-Laplacian pyramid to obtain a high-frequency layer sequence containing multi-scale detail information and a low-frequency base layer retaining overall structure; Based on the stray light sensitivity weight map, adaptive threshold filtering is designed for each layer of Laplacian detail image to effectively suppress stray light and protect edges; The contrast and information entropy of each layer of residual image after adaptive filtering are calculated and fused into an energy attenuation coefficient, and the filtering strength of the next layer is dynamically adjusted according to the coefficient to reconstruct a purified intensity image that suppresses noise and retains edges.

[0006] The line laser 3D profiler working method integrating multiple stray light suppression algorithms as described above, wherein, characterized by taking the purified intensity image and the original height map as input, using one-dimensional profile analysis to calculate the sharpness gradient and background flatness to obtain a morphological distortion index sequence, performing three-level wavelet packet decomposition on the sequence, extracting the energy concentration degree, and using the energy concentration degree as a criterion to determine whether to start the line-by-line stripe repair program based on morphological opening and closing operation. Specifically divided into the following sub-steps: Based on the purified intensity image, the main peak value of the laser stripe is located by Gaussian fitting combined with the original height map, and the peak sharpness gradient variance and background flatness standard deviation in the neighborhood are calculated with the main peak value as the center, and a morphological distortion index sequence is constructed; The third layer high-frequency detail coefficient of the morphological distortion index sequence is extracted by three-level wavelet packet decomposition, and the energy concentration degree is calculated, which accurately identifies severe distortion at the sequence level and provides a reliable criterion for subsequent repair; Based on the energy concentration degree result, the energy focusing criterion is introduced to comprehensively evaluate the stripe fracture degree by filtering difference and curvature enhancement item, and when the energy focusing criterion is greater than the set threshold, structure-guided interpolation repair is started combined with geometric prior to restore the image stripe continuity.

[0007] The line laser 3D profiler working method integrating multiple stray light suppression algorithms as described above, wherein, characterized by starting structure-guided interpolation repair combined with geometric prior to restore image stripe continuity. Specifically divided into the following sub-steps: Based on the purified intensity image and the morphological distortion index, a fracture positioning mask is generated using the low value area in the morphological distortion index and the detected energy concentration area; The purified intensity image is selectively smoothed and bridged in the scanning direction by multi-scale morphological opening and closing and closing opening filtering of the image fracture area, and a weight map is constructed using the difference between the two to highlight the fracture position; Under the guidance of the mask, according to the gradient direction, width characteristics of the stripes on both sides and the geometric constraints of the original height map, linear interpolation guided by structure tensor is used to adaptively reconstruct the fracture area, and the repaired pixels are updated to the global image to effectively restore the stripe continuity and structural integrity.

[0008] The line laser 3D profiler working method integrating multiple stray light suppression algorithms as described above, wherein, characterized by performing center line extraction on the stripe image processed by the repair program, using a three-dimensional reconstruction algorithm to calculate a motion blur correlation degree index, and if the index exceeds a preset dynamic threshold, using an inverse diffusion equation to perform anisotropic smoothing on the current frame height map to eliminate artifacts caused by dynamic stray light. Specifically divided into the following sub-steps: Sub-pixel level peak detection and cubic spline interpolation are performed on the repaired stripes to construct a smooth and continuous center line trajectory, thereby providing a geometric basis for high-precision three-dimensional reconstruction. Based on the center line stripe trajectory, the curvature change accumulation and the normal offset between adjacent frames are calculated by three-dimensional reconstruction to construct a motion blur correlation degree index, which quantifies the intensity of the artifacts caused by dynamic stray light. Based on the motion blur correlation degree index value, a spatiotemporal anisotropic inverse diffusion equation is used to effectively filter out the height field noise caused by dynamic stray light while preserving the geometric edges.

[0009] The line laser 3D profiler working method integrating multiple stray light suppression algorithms as described above, wherein the processed multiple frames of height maps are registered based on the iterative closest point method, the consistent angle cosine mean of normal vectors is extracted, and the geometric fidelity is generated in combination with the global height fluctuation coefficient, and the geometric fidelity is used as a feedback signal to drive the genetic algorithm to optimize the initial weight and restart the whole process until the geometric fidelity converges to its theoretical peak value and meets the preset accuracy requirement. Specifically, the following sub-steps are included: The iterative closest point method based on three-dimensional scale invariant feature transformation key point matching optimization is used to perform high-precision spatiotemporal registration on the denoised multiple frames of height maps, effectively overcoming the interference of motion blur and stray light, and forming a registered point cloud sequence; The point cloud sequence is surface reconstructed to generate a three-dimensional model, and the geometric fidelity index is calculated through the consistent mean of normal vectors of static area point clouds and the height fluctuation coefficient of variation, thereby evaluating the overall reconstruction quality; The geometric fidelity index is used as a feedback to dynamically optimize the key parameters using the genetic algorithm and restart the whole process, and through convergence judgment and double verification, the adaptive suppression accuracy and reconstruction robustness of stray light interference are significantly improved.

[0010] The application also provides a line laser 3D profiler working system integrating multiple stray light suppression algorithms, which comprises: A sensitivity weight map module: based on the multiple exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the best exposure interval are predicted through a deep convolutional neural network, and a stray light sensitivity weight map is constructed in combination with a point spread function; A purification intensity map module: the weight map is used to perform adaptive threshold filtering on each layer of image decomposed by a Gaussian Laplacian pyramid, and the stable index of each layer of filtered residual features is calculated, which is used as an energy attenuation coefficient to dynamically adjust the filtering intensity of the next layer, thereby generating a purification intensity image and optimizing the edge retention rate; Stripe repair module: take the purification intensity image and the original height map as input, adopt one-dimensional profile analysis to calculate the sharpness gradient and the background flatness to obtain the morphological distortion index sequence, carry out three-level wavelet packet decomposition on the sequence, extract the energy concentration degree, and determine whether to start the row-by-row stripe repair program based on the morphological opening and closing operation according to the energy concentration degree; Anisotropic smoothing module: perform center line extraction on the stripe image processed by the repair program, calculate the motion blur correlation degree index by using a three-dimensional reconstruction algorithm, if the index exceeds the preset dynamic threshold, perform anisotropic smoothing on the current frame height map by using an inverse diffusion equation to eliminate the artifacts caused by dynamic stray light; Registration optimization module: register the processed multiple height maps based on the iterative closest point method, extract the consistent cosine mean of normal vectors, generate the geometric fidelity by combining the global height fluctuation coefficient, and drive the genetic algorithm to optimize the initial weight by taking the geometric fidelity as the feedback signal and restarting the whole process until the geometric fidelity converges to the theoretical peak value and meets the preset accuracy requirement.

[0011] The present application has the following beneficial effects: the present application improves the accuracy of 3D imaging by using multiple stray light suppression algorithms, and ensures the imaging accuracy of the line laser 3D profiler. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0013] Figure 1 is a line laser 3D profiler working method flow chart provided by the first embodiment of the present application, which integrates multiple stray light suppression algorithms.

[0014] Figure 2 is a line laser 3D profiler working system schematic diagram provided by the second embodiment of the present application, which integrates multiple stray light suppression algorithms. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] Embodiment one As Figure 1As shown, the embodiment one of the present application provides a working method of line laser 3D profiler integrating multiple stray light suppression algorithms, comprising: S10, based on the multi-exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the best exposure interval are predicted through a deep convolutional neural network, and a stray light sensitivity weight map is constructed in combination with a point spread function.

[0017] Specifically, the stray light sensitivity weight map is constructed, specifically comprising the following sub-steps: S11, after extracting the original height map and denoising correction of the original image sequence, the best exposure interval is determined based on the multi-exposure response and the physical noise, the signal-to-noise ratio gain factor and the local intensity asymmetry are calculated, and the two are optimized by joint prediction through a double-branch convolutional neural network, so as to realize accurate quantification of laser stripe quality and stray light interference.

[0018] Based on the multi-exposure original image collected by the line laser 3D profiler, the profile analysis is performed on each row of scanning lines, the center position of the laser stripe is accurately extracted by using the centroid method, and according to the calibration parameters, the pixel displacement amount of the stripe center is converted into the actual height value by using the geometric triangular relationship, thereby generating a two-dimensional original height map , which represents the initial three-dimensional topography of the measured object surface.

[0019] The multi-exposure original image sequence is denoised by using a bilateral filtering method, and based on the calibration parameters, the camera intrinsic matrix and the distortion coefficient are used to perform geometric correction on the image to eliminate lens distortion, and the brightness consistency correction is performed through the radiation response function, so as to ensure that the laser stripe structure is clear and the gray value is accurate.

[0020] The deep convolutional neural network is used to extract features from the preprocessed image sequence, and the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the best exposure interval are accurately extracted, the signal-to-noise ratio gain factor reflects the intensity improvement degree of the signal relative to the background noise, and the local intensity asymmetry measures the brightness distribution asymmetry of the surrounding area of the pixel point, which is used to identify potential stray light, and the specific process is as follows: a double-branch full convolutional neural network is designed, the corrected image sequence is input, and two output channels correspond to the signal-to-noise ratio gain factor and the local intensity asymmetry, the network structure includes: the encoder extracts multi-scale features by layer-by-layer downsampling, and fuses different exposure level information; the decoder recovers the spatial resolution by layer-by-layer upsampling, fuses the high-dimensional features from the encoder, and retains the edges and fine structures; the loss function is the sum of the L1 regularization terms of the signal-to-noise ratio gain factor prediction error, the local intensity asymmetry prediction error and the asymmetry gradient.

[0021] On this basis, the network is divided into two independent output branches, one branch Extract its values ​​at K different exposure times The grayscale values ​​below are used to form a multi-exposure response curve. The exposure response curve is linearized using the calibrated camera response function to restore the true irradiance estimate. Based on this, the optimal exposure range for each pixel is determined by comprehensively considering factors such as whether the grayscale value is within the ideal dynamic range, whether the gradient is significant, and whether the changes between adjacent frames are stable. The exposure level that maximizes the signal-to-noise ratio is selected as the optimal imaging reference. The specific formula is as follows: , Index the selected optimal exposure level, which is the frame that provides the highest signal-to-noise ratio for this pixel (x,y) among all eligible exposure frames. For pixels The exposure time corresponding to the kth exposure level The grayscale response value below, Let represent photon shot noise, q represent the probability that each incident photon generates a detectable photogenerated electron under illumination, and E represent the incident irradiance. To read out noise, it is obtained from sensor calibration. To measure the maximum quality gain that a pixel can achieve throughout the entire exposure sequence, a signal-to-noise ratio (SNR) gain factor is defined, expressed on a logarithmic scale as the factor by which the SNR is improved under optimal conditions relative to the worst available imaging conditions.

[0022] Another branch takes a local neighborhood window centered on each pixel and analyzes the spatial symmetry of its grayscale distribution along the normal direction of the laser stripe. It calculates the mirror brightness difference of the corresponding position in the region about the central axis and sums them by weight to obtain the local symmetry error. Gaussian weight is used to highlight the influence of the central region. Then, the difference between the local peak intensity and the background intensity is normalized to eliminate the interference of low contrast area on the asymmetry evaluation and obtain the local intensity asymmetry. This is used to quantify the degree to which the light intensity distribution around the pixel deviates from the ideal symmetry shape, thereby reflecting the risk of stripe distortion caused by stray light.

[0023] The two branches share coding features but independently complete the regression task, thereby achieving refined, pixel-by-pixel joint prediction of the signal-to-noise ratio gain factor and local intensity asymmetry.

[0024] S12. Combining physical optics simulation and experimental calibration, the imaging optical path of the line laser 3D profilometer is accurately mathematically characterized to construct a point spread function, and this function is used to simulate the light energy diffusion effect caused by stray light.

[0025] Based on the actual imaging optical path structure of the line laser 3D profiler, the imaging optical path is accurately constructed to determine its point spread function, and the influence of stray light is improved from empirical observation to predictable physical process. Specifically, the light ray tracing method is used to simulate the light energy diffusion distribution of the ideal point light source on the image plane after passing through the lens, sensor and internal structure. In the process, the actual parameters such as the numerical aperture, focal length, aberration characteristics of the lens, light path obstruction, lens reflectivity and scattering characteristics of the non-ideal surface inside the cavity are comprehensively considered. In order to improve the accuracy, a small pinhole target is used as an approximate point light source in combination with the experimental calibration method, and the actual imaging results at different field positions are collected under the condition of no main beam interference, and the two-dimensional distribution of light energy diffusion is recorded. The measured image and the simulation result are registered and fused, and the unknown parameters such as scattering coefficient and surface roughness in the simulation are corrected in reverse through least squares fitting, and a high-fidelity point spread function distribution is obtained.

[0026] The point spread function is applied to the multi-exposure image sequence layer by layer as a convolution kernel to simulate the effects of non-local light halo, diffusion spot and background lifting of stray light in the imaging process. By analyzing the diffusion range, intensity overflow degree and neighborhood interference mode of the pixel gray value after convolution, the potential degree of different regions affected by stray light is quantified.

[0027] S13, based on the signal-to-noise ratio gain factor and the local intensity asymmetry, the point spread function is used to construct a stray light sensitivity weight map.

[0028] The stray light sensitivity weight map is defined to quantify the risk of each pixel affected by stray light. The specific formula is , wherein, The signal-to-noise ratio gain factor is greater, the original signal is stronger, and the noise resistance is better, The local intensity asymmetry is greater, the pixel is more affected by the lateral stray light, represents the original image The image obtained after applying the point spread function is used to measure the influence degree of stray light on the pixel. is the control weight of the point spread function term, is a preset lower threshold value, which is ensured not to be too small by taking the maximum value operation to prevent the weight map from having an abnormally high value. The obtained is a two-dimensional weight map with the same size as the original image, and each pixel position in the image is the calculated value, and the greater the value, the more susceptible to stray light interference.

[0029] S20, use the weight map to perform adaptive threshold filtering on each layer image after Gaussian Laplace pyramid decomposition, and count the stable index of each layer after filtering. The stable index is used as an energy attenuation coefficient to dynamically adjust the filtering strength of the next layer to generate a purification intensity image while optimizing the edge retention rate.

[0030] Specifically, the purification intensity image is generated, specifically including the following sub-steps: S21, Gaussian Laplace pyramid decomposition is performed on the original intensity image to obtain a high-frequency layer sequence containing multi-scale detail information and a low-frequency base layer retaining the overall structure.

[0031] An original intensity image collected by a line laser 3D profiler is obtained , the image is subjected to multi-level Gaussian Laplace pyramid decomposition, the image is layered according to different spatial scales, the separation of high-frequency details and low-frequency structures is realized, and accurate identification and adaptive filtering of local abnormalities caused by stray light are facilitated under multi-resolution. The number of decomposition layers is set to L layers, and the Gaussian image of each layer is obtained by performing Gaussian low-pass filtering and down-sampling on the image of the previous layer, wherein the image of the Lth layer is the low-frequency base layer of the top layer, which retains the overall structure and background information of the image. Subsequently, a Laplace pyramid is constructed based on the Gaussian pyramid: from the 1st layer to the Lth layer, the Laplace image of each layer is obtained by difference calculation of the Gaussian image of the current layer and the up-sampled result of the Gaussian image of the previous layer to the same size. The difference extracts high-frequency detail information at the corresponding scale, including edges, textures and local mutations, thereby forming a set of multi-scale detail layers. Finally, the original image is decomposed into a multi-resolution representation system composed of high-frequency detail layers and a low-frequency base layer. As the 0th layer, Gaussian low-pass filtering and down-sampling are performed layer by layer to obtain a series of smooth images with resolution halved layer by layer, wherein the image of the Lth layer is the low-frequency base layer of the top layer, which retains the overall structure and background information of the image. Subsequently, a Laplace pyramid is constructed based on the Gaussian pyramid: from the 1st layer to the Lth layer, the Laplace image of each layer is obtained by difference calculation of the Gaussian image of the current layer and the up-sampled result of the Gaussian image of the previous layer to the same size. The difference extracts high-frequency detail information at the corresponding scale, including edges, textures and local mutations, thereby forming a set of multi-scale detail layers. Finally, the original image is decomposed into a multi-resolution representation system composed of high-frequency detail layers and a low-frequency base layer.

[0032] S22, adaptive threshold filtering is designed for each layer of the Laplace detail image based on the stray light sensitivity weight map to effectively suppress stray light and protect edges.

[0033] First, initialize the parameters, set an empirical threshold for each layer , which is determined based on a statistical method and represents an estimate of the noise level of the layer. And determine the adjustment coefficient , which is used to control the influence strength of the stray light sensitivity weight map on the threshold. Then, calculate the adaptive threshold. For each pixel position (x, y), adjust the threshold according to its value in the weight map. For the Laplace image of the layer , the adaptive threshold is calculated as .

[0034] For each pixel (x, y) of each layer , use the indicator function I to judge whether its absolute value exceeds the adaptive threshold: is the Laplace image of the layer after adaptive threshold filtering, if , the indicator function value is 1, the original value of the pixel is reserved; otherwise, the indicator function value is 0, and the pixel value is 0. The adaptive threshold design makes it possible to use a higher threshold for filtering in areas with high signal-to-noise ratio and low stray light risk, thereby effectively removing noise without damaging real edges and details. In areas with high stray light risk, the threshold is reduced to allow more potential useful signals to pass through.

[0035] S23, the contrast and information entropy of each layer of the residual image after adaptive filtering are calculated and fused into an energy attenuation coefficient, and the filtering strength of the next layer is dynamically adjusted according to the coefficient, and a purified intensity map that suppresses noise and retains edges is reconstructed.

[0036] After adaptive threshold filtering is completed, the filtering effect of each layer is analyzed, and feedback adjustment basis is provided for the processing of the next layer. First, the Laplacian detail image of each layer is calculated The residual image after filtering: is the reserved part after adaptive threshold processing, and the residual represents the high-frequency components filtered out, including noise, abnormal textures caused by stray light, and weak signals mistakenly removed.

[0037] In order to quantify the local structural disorder degree reflected by these residuals, the gray level co-occurrence matrix of each layer of residual image is calculated, and two key texture features are extracted from it: contrast , reflecting the intensity of local brightness change in the image; information entropy , measuring the irregularity and complexity of the image gray scale distribution. Take the geometric mean of the contrast and information entropy, define it as the energy attenuation coefficient of the first layer, which reflects the structural disorder degree contained in the energy removed in the filtering process of the current layer. The larger the value is, the more non-random, non-noise complex textures there are in the layer residual, there is incomplete suppression of stray light interference or loss of effective details, which needs to be adjusted in the subsequent layer level.

[0038] The energy attenuation coefficient is used to dynamically adjust the filtering strategy of the next layer, and the empirical threshold of the first layer is set to be positively correlated with the energy attenuation coefficient of the first layer, that is is the preset baseline empirical threshold, is an adjustable gain coefficient used to control the feedback strength. The larger the value is, the stronger the stray light interference is, and the stronger the filtering strength of the next layer is, realizing layer-by-layer adaptive optimization and ensuring effective differentiation between real edges and stray light artifacts on different scales.

[0039] All the adaptive filtered detail layers are up-sampled and added to the low frequency base layer of the top layer to complete the reconstruction process of Laplacian pyramid, and the purification intensity image is obtained. The image effectively suppresses stray light noise while preserving the real edges and geometric structure of the laser stripe to the greatest extent.

[0040] S30, taking the purification intensity image and the original height map as inputs, using one-dimensional profile analysis to calculate the sharpness gradient and background flatness to obtain a morphological distortion index sequence, performing three-level wavelet packet decomposition on the sequence, extracting the energy concentration degree, and using the energy concentration degree as a criterion to determine whether to start the row-by-row stripe repair program based on morphological opening and closing operation.

[0041] Specifically, step S30 specifically includes the following sub-steps: S31, based on the purification intensity image, the main peak value of the laser stripe is located by Gaussian fitting combined with the original height map, and the peak sharpness gradient variance and the background flatness standard deviation in the neighborhood are calculated with the main peak value as the center to construct the morphological distortion index sequence.

[0042] In the purification intensity image, one-dimensional profile analysis is performed along each row of pixels, and the position of the maximum light intensity on each scanning line is extracted as the main peak value position by using the Gaussian fitting method. Due to the existence of multiple local maxima caused by stray light, it is difficult to determine the main peak value, and the height value corresponding to the current pixel provided by the original height map is used to assist in judgment. The maximum value at a reasonable height is the main peak value position. A local neighborhood window is selected with the main peak value as the center, and two indicators are calculated in the neighborhood: the peak sharpness gradient variance, and the background flatness standard deviation. The gradient is obtained by taking the first derivative of the stripe profile, which is the gradient, and then the variance of the absolute value of the gradient is calculated to measure the steepness of the stripe edge. If the variance value is small, it means that the stripe becomes flat and wide, which is caused by stray light diffusion. The background flatness standard deviation is the standard deviation of the background gray value in the region away from the peak value on both sides of the main peak of the stripe, which reflects the stability of the local background and the noise level. If the standard deviation increases, it is affected by the overflow light of the adjacent highlight area or by noise.

[0043] The sum of the peak sharpness gradient variance and the background flatness standard deviation is calculated along each row of scanning lines point by point, and a minimum constant is added to obtain the morphological distortion index sequence, wherein the minimum constant is used to prevent the denominator from being zero. Therefore, the morphological distortion index can sensitively represent the morphological integrity of the laser stripe at each spatial position: low value area corresponds to stripe fracture, severe widening or being submerged by stray light, which can be used as a trigger basis for subsequent repair mechanism.

[0044] S32, the third layer high frequency detail coefficient of the morphological distortion index sequence is extracted by three-level wavelet packet decomposition, and the energy concentration degree is calculated, which accurately identifies the severe distortion at the sequence level and provides a reliable criterion for subsequent repair.

[0045] The morphological distortion index sequence is used to identify abnormal regions such as stripe expansion, blur or background interference. However, relying only on point-by-point threshold judgment is easily affected by local noise or slight fluctuations, and it is difficult to distinguish between true structural distortion and gentle changes. Therefore, the morphological distortion index sequence is subjected to three-level wavelet packet decomposition to extract more discriminative indicators. In the first layer, the sequence is decomposed into low-frequency components and high-frequency components by low-pass and high-pass filters, respectively. The former represents the overall trend of the sequence, and the latter captures local mutations. In the second layer, the low-frequency and high-frequency components obtained in the first layer are further decomposed to obtain four more refined frequency bands. In the third layer, the four sub-bands in the second layer are further decomposed to obtain eight frequency bands. The detail coefficients in the deepest high-frequency path in the third layer are extracted , which reflect the mutations, edges or local sharp fluctuations in the morphological distortion index sequence, and correspond to the positions where the laser stripes are broken, jumped or severely expanded.

[0046] In order to quantify whether these high-frequency components are concentrated in certain specific regions, the energy concentration degree is calculated as follows: is the sum of the absolute values of all third-layer detail coefficients, is the sum of the squares of all third-layer detail coefficients, and N is the number of detail coefficients. When the energy of the detail coefficients is concentrated in a few positions, i.e., there is a significant mutation, the concentration degree increases significantly; otherwise, the energy is evenly distributed without obvious distortion, and the concentration degree is close to 1.

[0047] An empirical threshold is set. If , it indicates that there is a high-energy mutation in the scanning line, suggesting that there is a stripe break or severe expansion, and the line is marked as a line to be analyzed in depth. Otherwise, it is considered that the overall continuity of the stripe is good and no processing is needed, and the repair process is directly skipped.

[0048] S33, based on the energy concentration degree result, an energy focusing criterion is introduced to comprehensively evaluate the stripe break degree by filtering difference and curvature enhancement. When the energy focusing criterion is greater than a set threshold, structure-guided interpolation repair is started in combination with geometric prior to restore the continuity of the image stripe.

[0049] To accurately identify the distortion position and prevent misjudgment, an energy focusing criterion based on multi-scale morphological response is proposed to analyze the overall continuity and local irregularity of the stripe from a wider perspective and to judge whether repair is needed. The specific formula is: , where is the integral region, i.e., the effective pixels in the whole line, is the value of the morphological distortion index at position x , which is the result of opening operation followed by closing operation, is the value of The results of performing the closing operation first and then the opening operation show different responses to discontinuous regions, with significant differences at the fracture point. The absolute difference between the two highlights the true structural gap. To integrate the original morphological distortion exponent over the entire scan line, use it as the denominator. It becomes a relative metric to avoid misjudgment due to changes in overall brightness. for The variance of the second derivative measures the severity of local curvature fluctuations. for The L1 norm of the gradient measures the total edge intensity of the entire line signal and is used for normalization to prevent over-response in textured regions. This is the adjustment coefficient.

[0050] like If the value exceeds a preset threshold, it indicates significant discontinuity distortion in the current scan line. A line-by-line stripe repair procedure based on morphological opening and closing operations is then initiated to significantly enhance the continuity and structural integrity of the stripes. Specifically, based on the cleansing intensity image and the morphological distortion index, a fracture localization mask is generated using the low-value regions in the morphological distortion index and the energy concentration regions detected in step S32. The cleansing intensity image is selectively smoothed and bridged along the scanning direction using multi-scale morphological opening and closing and closing-opening filters. A weight map is constructed using the difference between the two to highlight the fracture location. Guided by the mask, and based on the gradient direction and width characteristics of the stripes on both sides, as well as the geometric constraints of the original height map, linear interpolation guided by the structural tensor is used to adaptively reconstruct the fracture region. Finally, the repaired pixels are updated to the global image, effectively restoring the continuity and structural integrity of the stripes.

[0051] S40. Perform centerline extraction on the striped image after the repair procedure, and calculate the motion blur correlation index using the three-dimensional reconstruction algorithm. If the index exceeds the preset dynamic threshold, use the inverse diffusion equation to perform anisotropic smoothing on the current frame height map to eliminate artifacts caused by dynamic stray light.

[0052] Specifically, anisotropic smoothing is performed on the current frame height map, which includes the following sub-steps: S41. Subpixel-level peak detection and cubic spline interpolation are performed on the repaired stripes to construct a smooth and continuous centerline trajectory, providing a geometric basis for high-precision 3D reconstruction.

[0053] Sub-pixel level center line extraction is performed on the repaired fringe image to obtain high-precision laser fringe geometric position. Specifically, the intensity peak of the repaired fringe is detected along each row scanning direction to preliminarily locate the pixel-level center point. A number of adjacent pixel points are selected near the peak for cubic spline interpolation to construct a continuous intensity distribution curve, and the sub-pixel accuracy is achieved by solving the extreme point of the curve. All the accurately positioned center points are connected in the scanning order to form a spatially continuous and smooth two-dimensional trajectory where s is the arc length parameter along the trajectory, and a represents the current scanning row.

[0054] S42, based on the center line fringe trajectory, the curvature change accumulation and the normal offset between adjacent frames are calculated by three-dimensional reconstruction to construct a motion blur correlation index, which quantifies the intensity of the artifacts caused by dynamic stray light.

[0055] To evaluate the influence of dynamic stray light on the reconstruction result, the three-dimensional reconstruction algorithm is used to perform geometric consistency analysis on the fringe center line at the same spatial position between adjacent frames. First, based on the sub-pixel center line trajectory and , the curvatures and at the corresponding points are calculated, and the absolute value of the curvature change is calculated. Then, the absolute cumulative sum of the curvature change is calculated along the entire trajectory, which reflects the degree of non-rigid distortion of the fringe shape between consecutive frames and is sensitive to local distortion caused by dynamic stray light. At the same time, the normal offset distance between the corresponding points of the center lines of adjacent frames is calculated, and the root mean square value is calculated to represent the stability of the overall displacement of the fringe. On this basis, the motion blur correlation index is constructed as , to prevent the denominator from being zero. This index effectively amplifies the pseudo-influence caused by dynamic light changes by normalizing the shape distortion intensity to the overall displacement level. If the value exceeds the preset dynamic threshold, it indicates that there is non-physical local jitter or blur caused by dynamic stray light.

[0056] S43, based on the motion blur correlation index value, a spatiotemporal anisotropic inverse diffusion equation is used to effectively filter out the high-field noise caused by dynamic stray light while preserving the geometric edges.

[0057] Based on the motion blur correlation index value, a spatiotemporal anisotropic inverse diffusion smoothing equation is proposed to suppress the influence of such artifacts on the three-dimensional reconstruction result. The specific process is as follows: the height map ​The following partial differential equation is introduced for spatiotemporal co-optimization: ,in, height map The rate of evolution with virtual time t, This is a two-dimensional height map of the current frame, calculated from the stripe center line using a three-dimensional reconstruction algorithm. This is a spatially anisotropic diffusion term, responsible for smoothing noise and preserving edges within the spatial domain. For motion-based fuzzy correlation The diffusion rate control function, For adaptive diffusion rate control coefficients, It is a divergence operator used to construct nonlinear diffusion processes and achieve local adaptive smoothing. Let H be the gradient vector field of the height map. For anisotropic diffusion-type edge preservation functions, strong spatial smoothing is allowed in flat regions to effectively smooth noise, while smoothing is suppressed in edge regions to protect the boundary. This is a time consistency constraint used to ensure a smooth transition of the height field between adjacent frames. Adjust the strength of the time constraint. For time consistency, it pushes the current frame heightmap H to the previous frame. By bringing the frames closer together, we can prevent abrupt changes caused by single-frame artifacts and improve sequence stability. This is the dynamic adjustment coefficient.

[0058] S50. The processed multi-frame height map is registered based on the iterative nearest point method. The mean cosine of the normal vector cosine is extracted and combined with the global height fluctuation coefficient to generate geometric fidelity. The geometric fidelity is used as a feedback signal to drive the genetic algorithm to optimize the initial weights and restart the entire process until the geometric fidelity converges to its theoretical peak value and meets the preset accuracy requirements.

[0059] Specifically, step S50 includes the following sub-steps: S51. The iterative nearest point method based on the key point matching optimization of three-dimensional scale-invariant feature transformation is adopted to perform high-precision spatiotemporal registration on the denoised multi-frame height map, effectively overcoming motion blur and stray light interference, and forming a registered point cloud sequence.

[0060] To realize high-precision three-dimensional dynamic reconstruction, the multi-frame height maps after pre-sequencing denoising and repair processing are spatio-temporally registered. Due to the existence of small deformation, motion blur or local loss in single-frame height field, direct splicing will lead to cumulative error and structural misplacement, so the iterative closest point method based on feature matching optimization is used for accurate registration. First, generate the corresponding three-dimensional point cloud for each frame of height map, and extract the three-dimensional scale invariant feature transform key points on its surface. These key points are concentrated in the areas with significant curvature and rich geometric features, effectively resisting local intensity disturbance caused by stray light. Then, the three-dimensional scale invariant feature transform descriptors of each key point are calculated, and the corresponding point set between adjacent frames is established through descriptor matching, and the initial rigid transformation matrix containing rotation and translation is solved based on the corresponding point set.

[0061] On this basis, the iterative closest point method process is executed: first, search for the nearest neighbor point in the target frame for each point in the source frame according to the current pose, and establish the corresponding relationship; then construct a weighted registration error term, use the point-to-plane distance minimization strategy to improve the convergence accuracy, and assign higher weights to the high-confidence feature matching points provided by the three-dimensional scale invariant feature transform to enhance the guiding role of reliable correspondence on overall transformation estimation; at the same time, combined with distance threshold, normal angle and other geometric consistency tests, dynamically remove the mis-matching points to suppress abnormal disturbance; then solve the optimal rigid transformation based on the weighted error function, and update the pose of the current frame; finally, judge whether the transformation increment is less than the preset threshold or the maximum iteration number is reached, if satisfied, exit the iteration, otherwise return to continue optimization until high-precision registration is achieved. The whole process is pushed frame by frame in time sequence, ensuring that each frame of height map is accurately aligned in a unified coordinate system to form a continuous, stable and detailed point cloud sequence.

[0062] S52, surface reconstruction is performed on the point cloud sequence to generate a three-dimensional model, and a geometric fidelity index is calculated based on the normal vector consistency mean and height fluctuation coefficient of variation of the static area point cloud to evaluate the overall reconstruction quality.

[0063] After registration, the registered point cloud sequence is reconstructed by Poisson surface reconstruction to convert the discrete point cloud into a continuous and closed triangular mesh model while preserving the original geometric details. A geometric fidelity index is calculated to evaluate the geometric reliability and stability of the overall reconstruction result. Static regions are extracted from the registered time-series point cloud to exclude the interference of dynamic object displacement on the evaluation. Within these static regions, two key geometric features are calculated: one is the mean value of the cosine of the normal vector consistency angle, which is obtained by calculating the cosine value of the normal vector angle between adjacent frames and averaging it, reflecting the time-series stability of the surface orientation. The closer the value is to 1, the smaller the normal change, and the more consistent the reconstruction. The other is the coefficient of variation of the global height fluctuation after segmented linear transformation correction. This index is obtained by first fitting the segmented linear background trend of each frame's height field to eliminate systematic tilt or distortion, and then calculating the ratio of the standard deviation to the mean of the residual. It is used to quantify the degree of non-physical fluctuation of the height value. Both are normalized to the interval [0, 1] and weighted to obtain the geometric fidelity index. The closer the index value is to 1, the higher the reconstruction quality.

[0064] S53, using the geometric fidelity index as feedback, dynamically optimizing key parameters using a genetic algorithm and restarting the entire process, significantly improving the adaptive suppression accuracy of stray light interference and the robustness of reconstruction through convergence judgment and double verification.

[0065] To improve the suppression accuracy of dynamic stray light on the model, a closed-loop optimization mechanism is introduced using a genetic algorithm, with the geometric fidelity index as the fitness feedback signal to drive global optimization of key parameters. By iteratively adjusting the core parameters of denoising and repair effects, and restarting the entire process for verification, the geometric fidelity index gradually converges to the peak value, ensuring that the reconstruction result achieves an optimal balance between geometric consistency and detail fidelity. This process significantly enhances the adaptive ability of stray light suppression, especially in scenes with severe changes in light or local strong interference, it can automatically match the optimal filtering strategy, effectively avoiding the problem of residual artifacts due to under-smoothing or loss of details due to over-smoothing. And use double verification mechanism of outlier detection and physical simulation. Through outlier detection, local distortion, over-smoothing areas or false edges in the reconstructed point cloud are identified, and parameter combinations with high index but poor detail are excluded. Then combined with physical simulation verification, the time-series deformation trend of the reconstruction is compared with the pre-set simple harmonic vibration model to check whether there are false fluctuations that do not conform to the actual motion behavior. For static scenes, the displacement of each point is required to be close to zero; for periodic vibration scenes, the reconstructed trajectory is required to be highly consistent with the sinusoidal change rule. If the deviation exceeds the set threshold, the parameter combination is invalid. The two verification methods together ensure that the optimized parameters not only suppress stray light interference, but also have geometric reasonableness and physical authenticity, thus achieving high-precision and high-robustness dynamic three-dimensional reconstruction.

[0066] Embodiment Two As Figure 2As shown, the second embodiment of the present application provides a line laser 3D profiler working system integrating multiple stray light suppression algorithms, comprising: A sensitivity weight map module: based on the multi-exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the optimal exposure interval are predicted through a deep convolutional neural network, and a stray light sensitivity weight map is constructed in combination with a point spread function. Specifically, it is divided into the following sub-modules: A double-branch sub-module: after extracting the original height map from the original image sequence and correcting the noise, the optimal exposure interval is determined based on the multi-exposure response and physical noise, the signal-to-noise ratio gain factor and the local intensity asymmetry are calculated, and the two are jointly predicted and optimized through a double-branch convolutional neural network to realize accurate quantification of laser stripe quality and stray light interference.

[0067] A point spread function sub-module: in combination with physical optics simulation and experimental calibration, the imaging light path of the line laser 3D profiler is accurately mathematically characterized to construct a point spread function, and the function is used to simulate the light energy diffusion effect caused by stray light.

[0068] A construction sub-module: based on the signal-to-noise ratio gain factor and the local intensity asymmetry in combination with the point spread function, a stray light sensitivity weight map is constructed.

[0069] A purification intensity map module: using the weight map, adaptive threshold filtering is performed on each layer of image after Gaussian-Laplacian pyramid decomposition, and the stable index of each layer of filtered residual feature is calculated. The stable index is used as an energy attenuation coefficient to dynamically adjust the filtering intensity of the next layer to generate a purified intensity image while optimizing the edge retention rate. Specifically, it is divided into the following sub-modules: A decomposition sub-module: the original intensity image is decomposed by Gaussian-Laplacian pyramid to obtain a high-frequency layer sequence containing multi-scale detail information and a low-frequency base layer retaining overall structure.

[0070] A filtering sub-module: based on the stray light sensitivity weight map, an adaptive threshold filter is designed for each layer of Laplacian detail image to achieve effective suppression of stray light and edge protection.

[0071] A reconstruction sub-module: the contrast and information entropy of each layer of residual image after adaptive filtering are calculated and fused into an energy attenuation coefficient, and the filtering intensity of the next layer is dynamically adjusted according to the coefficient to reconstruct a purified intensity image that suppresses noise and retains edges.

[0072] A stripe repair module: taking the purified intensity image and the original height map as input, a morphological distortion index sequence is obtained by calculating the sharpness gradient and background flatness using one-dimensional profile analysis, and the energy concentration degree is extracted by three-level wavelet packet decomposition. The energy concentration degree is used as a criterion to determine whether to start the row-by-row stripe repair program based on morphological opening and closing operation. Specifically, it is divided into the following sub-modules: Morphological distortion sub-module: Based on the purification intensity image, the main peak value of the laser stripe is located by Gaussian fitting combined with the original height map, and the peak sharpness gradient variance and background flatness standard deviation in the neighborhood are calculated with the main peak value as the center to construct a morphological distortion index sequence.

[0073] Energy concentration sub-module: The third layer high frequency detail coefficient of the morphological distortion index sequence is extracted by three-level wavelet packet decomposition, and the energy concentration degree is calculated to accurately identify severe distortion at the sequence level and provide reliable criteria for subsequent repair.

[0074] Interpolation repair sub-module: Based on the energy concentration degree result, the energy focusing criterion is introduced to comprehensively evaluate the stripe fracture degree by filtering difference and curvature enhancement, and when the energy focusing criterion is greater than the set threshold, structure-guided interpolation repair is started combined with geometric prior to restore the stripe continuity of the image.

[0075] Anisotropic smoothing module: The center line extraction is performed on the stripe image processed by the repair program, the motion blur correlation index is calculated by using the three-dimensional reconstruction algorithm, and if the index exceeds the preset dynamic threshold, the anisotropic smoothing is performed on the current frame height map by using the inverse diffusion equation to eliminate the artifacts caused by dynamic stray light. It is specifically divided into the following sub-modules: Center line sub-module: The peak value detection and cubic spline interpolation at the sub-pixel level are performed on the repaired stripe to construct a smooth and continuous center line trajectory, which provides a geometric basis for high-precision three-dimensional reconstruction.

[0076] Motion correlation sub-module: Based on the center line stripe trajectory, the curvature change accumulation and normal offset between adjacent frames are calculated by three-dimensional reconstruction to construct a motion blur correlation index, which quantifies the artifact intensity caused by dynamic stray light.

[0077] Anisotropic sub-module: Based on the motion blur correlation index value, the spatiotemporal anisotropic inverse diffusion equation is used to effectively filter out the height field noise caused by dynamic stray light while preserving the geometric edges.

[0078] Registration optimization module: The processed multiple height maps are registered based on the iterative closest point method, the consistent cosine mean of normal vectors is extracted, and the geometric fidelity is generated combined with the global height fluctuation coefficient. The geometric fidelity is used as a feedback signal to drive the genetic algorithm to optimize the initial weight and restart the whole process until the geometric fidelity converges to its theoretical peak value and meets the preset accuracy requirement. It is specifically divided into the following sub-modules: Registration sub-module: The iterative closest point method based on three-dimensional scale invariant feature transform key point matching optimization is used to perform high-precision spatiotemporal registration on the denoised multiple height maps, effectively overcoming the motion blur and stray light interference, and forming a registration point cloud sequence.

[0079] Geometric fidelity sub-module: surface reconstruction is performed on the point cloud sequence to generate a three-dimensional model, and a geometric fidelity index is calculated based on the normal vector consistency mean and height fluctuation coefficient of the static area point cloud to evaluate the overall reconstruction quality.

[0080] Feedback optimization sub-module: taking the geometric fidelity index as feedback, the key parameters are dynamically optimized using a genetic algorithm, and the whole process is restarted. Through convergence judgment and double verification, the adaptive suppression precision of stray light interference and the reconstruction robustness are significantly improved.

[0081] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions of the present application should be included within the scope of protection of the present application.

Claims

1. A working method of a line laser 3D profilometer integrating multiple stray light suppression algorithms, characterized in that, The method comprises the following steps: S10, based on the multi-exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the best exposure interval are predicted through a deep convolutional neural network, and a stray light sensitivity weight map is constructed in combination with a point spread function; S20, using the weight map to perform adaptive threshold filtering on each layer image after Gaussian-Laplacian pyramid decomposition, and counting the residual feature stability index after filtering each layer, taking the stability index as the energy attenuation coefficient to dynamically adjust the filtering strength of the next layer, generating a purification intensity image while optimizing the edge retention rate; S30, taking the purification intensity image and the original height map as input, using one-dimensional profile analysis to calculate the sharpness gradient and background flatness to obtain a morphological distortion index sequence, performing three-level wavelet packet decomposition on the sequence, extracting the energy concentration degree, and taking the energy concentration degree as the criterion to determine whether to start the row-by-row stripe repair program based on morphological opening and closing operation; S40, performing center line extraction on the stripe image processed by the repair program, calculating the motion blur correlation degree index using a three-dimensional reconstruction algorithm, and if the index exceeds the preset dynamic threshold, performing anisotropic smoothing on the current frame height map using an inverse diffusion equation to eliminate artifacts caused by dynamic stray light; S50, registering the processed multi-frame height map based on the iterative closest point method, extracting the consistent angle cosine mean of the normal vector, and generating the geometric fidelity combining the global height fluctuation coefficient, and taking the geometric fidelity as the feedback signal to drive the genetic algorithm to optimize the initial weight and restart the whole process until the geometric fidelity converges to its theoretical peak value and meets the preset accuracy requirement.

2. The method of claim 1, wherein the integrated multiple stray light rejection algorithm linear laser 3D profilometer working method is characterized in that, Based on the multi-exposure original image sequence collected by the line laser 3D profiler, the signal-to-noise ratio gain factor and the local intensity asymmetry of each pixel point in the best exposure interval are predicted through a deep convolutional neural network, and a stray light sensitivity weight map is constructed in combination with a point spread function, which comprises the following sub-steps: After extracting the original height map and denoising correction from the original image sequence, the best exposure interval is determined based on the multi-exposure response and physical noise, the signal-to-noise ratio gain factor and the local intensity asymmetry are calculated, and the two are optimized through a double-branch convolutional neural network to realize accurate quantification of laser stripe quality and stray light interference; In combination with physical optics simulation and experimental calibration, the imaging light path of the line laser 3D profiler is accurately mathematically characterized to construct a point spread function, and the light energy diffusion effect caused by stray light is simulated using the function; Based on the signal-to-noise ratio gain factor and the local intensity asymmetry in combination with the point spread function, a stray light sensitivity weight map is constructed.

3. The method of claim 1, wherein the integrated multiple stray light rejection algorithm linear laser 3D profilometer working method is characterized by, Using the weight map to perform adaptive threshold filtering on each layer image after Gaussian-Laplacian pyramid decomposition, and counting the residual feature stability index after filtering each layer, taking the stability index as the energy attenuation coefficient to dynamically adjust the filtering strength of the next layer, generating a purification intensity image while optimizing the edge retention rate, which comprises the following sub-steps: The original intensity image is subjected to Gaussian-Laplacian pyramid decomposition to obtain a high-frequency layer sequence containing multi-scale detail information and a low-frequency base layer retaining overall structure; An adaptive threshold filter is designed for each layer Laplacian detail image based on stray light sensitivity weight map to effectively suppress stray light and protect edges; The contrast and information entropy of each layer residual image after adaptive filtering are calculated and fused into energy attenuation coefficient, and the filtering strength of the next layer is dynamically adjusted according to the coefficient to reconstruct the purification intensity map which suppresses noise and preserves edges.

4. The method of claim 1, wherein the integrated multiple stray light rejection algorithm linear laser 3D profilometer working method is characterized by, With the purification intensity image and the original height map as input, the morphological distortion index sequence is obtained by calculating the sharpness gradient and background flatness using one-dimensional profile analysis, and the sequence is decomposed by three-level wavelet packet to extract the energy concentration degree, which is used as a criterion to determine whether to start the line-by-line stripe repair program based on morphological opening and closing operation, which is specifically divided into the following sub-steps: Based on the purification intensity image, the main peak value of the laser stripe is located by Gaussian fitting combined with the original height map, and the peak sharpness gradient variance and background flatness standard deviation in the neighborhood are calculated with the main peak value as the center to construct the morphological distortion index sequence; The third layer high-frequency detail coefficient of the morphological distortion index sequence is extracted by three-level wavelet packet decomposition, and the energy concentration degree is calculated to accurately identify severe distortion at the sequence level and provide a reliable criterion for subsequent repair; Based on the energy concentration degree result, the energy focusing criterion is introduced to comprehensively evaluate the degree of stripe breakage by filtering difference and curvature enhancement term, and when the energy focusing criterion is greater than the set threshold, the structure-guided interpolation repair is started combined with geometric prior to restore the continuity of image stripes.

5. The method of claim 4, wherein the integrated multiple stray light rejection algorithm linear laser 3D profilometer working method is characterized by, The structure-guided interpolation repair is started combined with geometric prior to restore the continuity of image stripes, which is specifically divided into the following sub-steps: Based on the purification intensity image and the morphological distortion index, a broken positioning mask is generated using the low value area in the morphological distortion index and the detected energy concentration area; The purification intensity image is selectively smoothed and bridged in the broken area by multi-scale morphological opening and closing and closing opening filtering along the scanning direction, and a weight map is constructed using the difference between the two to highlight the broken position; Under the guidance of the mask, the broken area is adaptively reconstructed by linear interpolation guided by structure tensor according to the gradient direction, width characteristics of the stripes on both sides and the geometric constraints of the original height map, and the repaired pixels are updated to the global image to effectively restore the continuity of stripes and the integrity of the structure.

6. The method of claim 1, wherein the integrated multiple stray light rejection algorithm linear laser 3D profilometer working method is characterized by, The centerline extraction is performed on the stripe image processed by the repair program, the motion blur correlation index is calculated using the three-dimensional reconstruction algorithm, and if the index exceeds the preset dynamic threshold, the anisotropic smoothing is performed on the current frame height map using the inverse diffusion equation to eliminate the artifacts caused by dynamic stray light, which is specifically divided into the following sub-steps: The peak value detection and cubic spline interpolation are performed on the repaired stripes at the sub-pixel level to construct a smooth and continuous centerline trajectory, which provides a geometric basis for high-precision three-dimensional reconstruction; Based on the centerline stripe trajectory, the curvature change accumulation and normal offset between adjacent frames are calculated by three-dimensional reconstruction to construct the motion blur correlation index, which quantifies the intensity of artifacts caused by dynamic stray light; Based on the motion blur correlation index value, the time-space anisotropic inverse diffusion equation is used to effectively filter out the height field noise caused by dynamic stray light while preserving the geometric edges.

7. The method of claim 1, wherein the integrated multiple stray light rejection algorithm linear laser 3D profilometer working method is characterized by, The processed multi-frame height map is registered based on an iterative closest point method, a consistent angle cosine mean of normal vectors is extracted, a global height fluctuation coefficient is combined to generate a geometric fidelity, and the geometric fidelity is taken as a feedback signal to drive a genetic algorithm to optimize initial weights and restart the whole process until the geometric fidelity converges to a theoretical peak value and meets a preset accuracy requirement, and the process is specifically divided into the following sub-steps: An iterative closest point method based on three-dimensional scale invariant feature transform key point matching optimization is adopted to perform high-precision space-time registration on the denoised multi-frame height map, motion blur and stray light interference are effectively overcome, and a registered point cloud sequence is formed; A three-dimensional model is generated by surface reconstruction on the point cloud sequence, a geometric fidelity index is calculated through a consistent mean of normal vectors of static area point clouds and a height fluctuation coefficient of variation, and overall reconstruction quality is evaluated; The geometric fidelity index is taken as feedback, a genetic algorithm is used to dynamically optimize key parameters and restart the whole process, and through convergence judgment and double verification, adaptive suppression precision and reconstruction robustness to stray light interference are significantly improved.

8. A working system of a line laser 3D profilometer integrating multiple stray light suppression algorithms, characterized in that, It includes: A sensitivity weight map module: based on a multi-exposure original image sequence collected by a line laser 3D profiler, a depth convolutional neural network is used to predict a signal-to-noise ratio gain factor and a local intensity asymmetry of each pixel point in an optimal exposure interval, and a stray light sensitivity weight map is constructed in combination with a point spread function; A purification intensity map module: using the weight map, adaptive threshold filtering is performed on each layer of image decomposed by a Gaussian Laplacian pyramid, and a stable index of residual features after filtering of each layer is calculated, the stable index is taken as an energy attenuation coefficient to dynamically adjust the filtering intensity of the next layer, a purification intensity image is generated, and the edge retention rate is optimized; A stripe repair module: taking the purification intensity image and the original height map as inputs, a morphological distortion index sequence is obtained by calculating sharpness gradient and background flatness through one-dimensional profile analysis, the sequence is subjected to three-level wavelet packet decomposition, an energy concentration degree is extracted, and the energy concentration degree is taken as a criterion to determine whether to start a row-by-row stripe repair program based on morphological opening and closing operation; An anisotropic smoothing module: center line extraction is performed on the stripe image processed by the repair program, a motion blur correlation index is calculated by a three-dimensional reconstruction algorithm, if the index exceeds a preset dynamic threshold, an anisotropic smoothing is performed on the current frame height map by an inverse diffusion equation to eliminate artifacts caused by dynamic stray light; A registration optimization module: the processed multi-frame height map is registered based on an iterative closest point method, a consistent angle cosine mean of normal vectors is extracted, a global height fluctuation coefficient is combined to generate a geometric fidelity, and the geometric fidelity is taken as a feedback signal to drive a genetic algorithm to optimize initial weights and restart the whole process until the geometric fidelity converges to a theoretical peak value and meets a preset accuracy requirement.

Images