A spot denoising method for a long-distance laser triangulation displacement measurement system

Abstract

This invention relates to the field of optical detection technology, and more particularly to a method for denoising light spots in a long-distance laser triangulation displacement measurement system. The method includes: constructing a dual-reflection long-distance laser triangulation measurement system; acquiring light spot image samples; performing grayscale conversion and region of interest extraction to obtain a region of interest image; performing a logarithmic transformation on the region of interest image; decomposing the transformed image into a pair of downsampled feature maps using a downsampling operator; constructing a loss function for a neural network; obtaining a trained lightweight neural network by minimizing the loss function; and inputting the original light spot image into the lightweight neural network to obtain a denoised light spot image. The light spot denoising method of this invention significantly suppresses noise in light spot images, solving the problem of excessive smoothing in traditional denoising methods. Furthermore, the main network structure uses only two convolutional layers, offering advantages such as lightweight design, few parameters, fast training speed, and excellent denoising effect.

Description

Technical Field

[0001] This invention relates to the field of optical detection technology, and in particular to a method for denoising the light spot in a long-distance laser triangulation displacement measurement system. Background Technology

[0002] With the continuous advancement of social technology and the rise of intelligent manufacturing, long-range laser measurement technology plays a vital role in industrial production and daily life. In industry, it is used to accurately measure the size, shape, and position of objects, such as parts inspection and 3D modeling in manufacturing, as well as design and monitoring in architecture and civil engineering. In daily life, it is used in fields such as interior design, urban planning, medicine, and environmental monitoring, providing high-precision and high-efficiency measurement solutions for various applications.

[0003] Laser triangulation, as a non-contact measurement method, boasts advantages such as high accuracy, good stability, long lifespan, fast response speed, and low cost. However, it suffers from decreased accuracy over long distances. Patent CN202123110079.8 proposes a long-distance laser triangulation displacement sensor based on a reflective structure. However, the introduction of new reflective units increases noise in the laser spot image, limiting the spot positioning accuracy. Therefore, a high-performance image denoising method is needed. Traditional denoising methods such as Gaussian filtering, median filtering, and Lee filtering all present a trade-off between noise removal and feature preservation. While filtering out noise and smoothing the image, they also lead to problems such as laser spot feature attenuation and softened edges, which reduces the spot positioning accuracy. Summary of the Invention

[0004] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is the problem of decreased accuracy, reduced system sensitivity and increased noise in spot imaging of laser triangulation displacement sensors when performing long-distance measurements as the measurement range increases.

[0005] To achieve the above objectives, the present invention provides a method for denoising the laser spot in a long-distance laser triangulation displacement measurement system, comprising the following steps:

[0006] Step 1: Construct a dual-reflection long-distance laser triangulation system, acquire light spot image samples, and perform grayscale conversion and region of interest extraction on the light spot image samples to obtain the region of interest image;

[0007] Step 2: Perform a logarithmic transformation on the region of interest image to make the noise level of the transformed image uniformly distributed;

[0008] Step 3: Decompose the transformed image into a pair of downsampled feature maps using a downsampling operator;

[0009] Step 4: Considering the impact of outliers, residual learning, and the principle of symmetric loss, construct the loss function for the neural network;

[0010] Step 5: By minimizing the loss function of the neural network, train a lightweight neural network consisting of two convolutional layers, map one downsampled feature map to another downsampled feature map, and obtain the trained lightweight neural network.

[0011] Step 6: Input the original spot image into the lightweight neural network trained in Step 5 to obtain the denoised spot image, and calculate the repeatability accuracy of the system.

[0012] In a preferred embodiment of the present invention, the dual-reflection long-distance laser triangulation system includes a first reflector, a second reflector, an imaging CMOS, a receiving lens, a paper target to be measured, a laser source focusing module, and a laser source.

[0013] Preferably, the laser beam emitted by the laser source is incident perpendicularly on the paper target being tested through the focusing module. The diffuse reflected light is focused by the receiving lens, and after being reflected by the first and second reflecting mirrors, it is imaged on the CMOS imaging device.

[0014] Preferably, step 1 includes the following steps:

[0015] Step 1.1: Place the object to be measured at 2250mm from the laser emitter, calculate the focal length of the receiving lens, and design the optical path;

[0016] Step 1.2: Based on the principle of reflection, since the imaging optical path after reflection by the reflector unit is symmetrical to the original optical path without the reflector unit, calculate the position parameters of the optical elements in the symmetrical optical path;

[0017] Step 1.3: Based on the focal length of the receiving lens and the position parameters of the optical elements in the symmetrical optical path, construct the actual optical path, set the camera frame rate to 10Hz, and continuously acquire 50 light spot images as experimental sample images.

[0018] Step 1.4: Convert the experimental sample image to grayscale to obtain a grayscale image sample;

[0019] Step 1.5: First, extract the contour and calculate the geometric center coordinates of the light spot on the grayscale image sample. Then, set an appropriate cropping box and place the light spot at the center of the image to obtain the region of interest image.

[0020] In a preferred embodiment of the present invention, the logarithmic transformation method in step 2 is as follows:

[0021]

[0022] Where m is the normalization parameter, i represents the x-coordinate of the pixel, and j represents the y-coordinate of the pixel.

[0023] In a preferred embodiment of the present invention, step 3 includes the following steps:

[0024] Step 3.1: Use a downsampling operator with a convolution kernel size of 2×2 and a stride of 2 to segment the 1041 pixel × 1053 pixel × 2 light spot image into non-overlapping blocks of size 2×2;

[0025] Step 3.2: Assign the average value of the sub-diagonal pixels of each block to the first pixel of a downsampled feature map D1(y), similar to average pooling. Similarly, assign the average value of the main diagonal pixels to another downsampled feature map D2(y).

[0026] Preferably, both downsampled feature maps are 520 pixels × 526 pixels × 2.

[0027] In a preferred embodiment of the present invention, the loss function for constructing the neural network includes the following steps:

[0028] Step 4.1: Based on the pair of downsampled feature maps obtained after downsampling the noisy image, fit a neural network f. To reduce the influence of outliers, the loss function is defined as:

[0029] L=||f(D1(y))-D2(y)||1

[0030] Step 4.2: Considering that in residual learning the network is optimized to adapt to noise rather than the image, the loss function is modified as follows:

[0031]

[0032] Step 4.3: To ensure that the results obtained by denoising the noisy image first and then downsampling are similar to those obtained by downsampling first and then denoising, and to improve consistency, the following loss function is adopted:

[0033]

[0034] In a preferred embodiment of the present invention, step 5 includes the following steps:

[0035] Step 5.1: Gradientization to minimize loss L = L res. +L cons. Set reasonable network training parameters and use the Adam optimizer to accelerate the network training process to obtain a trained neural network;

[0036] Step 5.2: Use the trained neural network to obtain the denoised image x = yf(y).

[0037] In a preferred embodiment of the present invention, the repeatability accuracy of the calculation system includes: using the square-weighted centroid method to calculate the position of the spot centroid, obtaining the standard deviation of the position of the spot centroid in the horizontal and vertical directions of 50 images acquired in a single acquisition, and then obtaining the repeatability accuracy of the system.

[0038] Technical effects of the present invention:

[0039] 1. The spot denoising method in this invention has a significant effect on suppressing noise in spot images, and solves the problem of excessive smoothing in traditional denoising methods such as Gaussian and median filtering. At the same time, the main network structure only uses two convolutional layers, which has the advantages of being lightweight, having few parameters, fast training speed, and excellent denoising effect.

[0040] 2. The spot noise reduction method in this invention can help improve the accuracy of spot center positioning, ensure the accuracy of measurement results and the stable operation of the system;

[0041] 3. The self-supervised lightweight neural network in this invention does not require pre-labeled data and can self-adjust and optimize during actual operation. The sensor can adapt to various operating environments, and system performance can be improved through software optimization without adding additional hardware. This method is more economical than physically improving the sensor system, and maintenance and upgrades are also more convenient, helping to reduce the overall system's operating and maintenance costs.

[0042] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the optical path principle of a dual-reflection long-distance laser triangulation system built from the acquisition of light spot images according to a preferred embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the light spot denoising network principle of a preferred embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram comparing the noise reduction performance of different methods according to a preferred embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram comparing the impact of different methods on the spot positioning accuracy of a preferred embodiment of the present invention;

[0047] Figure 5This is a schematic diagram illustrating the system repeatability accuracy at different working distances according to a preferred embodiment of the present invention.

[0048] In the diagram: 1. First reflecting mirror; 2. Second reflecting mirror; 3. Imaging CMOS; 4. Receiving lens; 5. Paper target under test; 6. Laser source focusing module; 7. Laser source. Detailed Implementation

[0049] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0050] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0051] refer to Figures 1-5 In one specific embodiment of the present invention, a method for denoising the light spot in a long-distance laser triangulation displacement measurement system is disclosed, comprising the following steps:

[0052] Step 1: Construct a dual-reflection long-distance laser triangulation system, acquire light spot image samples, and perform grayscale conversion and region of interest extraction on the light spot image samples to obtain the region of interest image;

[0053] Step 2: Perform a logarithmic transformation on the region of interest image to make the noise level of the transformed image uniformly distributed;

[0054] Step 3: Decompose the transformed image into a pair of downsampled feature maps using a downsampling operator;

[0055] Step 4: Considering the impact of outliers, residual learning, and the principle of symmetric loss, construct the loss function for the neural network;

[0056] Step 5: By minimizing the loss function of the neural network, train a lightweight neural network consisting of two convolutional layers, map one downsampled feature map to another downsampled feature map, and obtain the trained lightweight neural network.

[0057] Step 6: Input the original spot image into the lightweight neural network trained in Step 5 to obtain the denoised spot image, and calculate the repeatability accuracy of the system.

[0058] The spot denoising method in this invention significantly suppresses noise in spot images, solving the problem of excessive smoothing in traditional denoising methods such as Gaussian and median filtering. Furthermore, the main network structure uses only two convolutional layers, offering advantages such as lightweight design, few parameters, fast training speed, and excellent denoising performance. In a specific embodiment of this invention, an improved self-supervised lightweight neural network is used to suppress noise in spot images of a long-distance laser triangulation system, reducing the peak signal-to-noise ratio (PSNR) of the spot images and significantly improving the repeatability accuracy of the entire system.

[0059] like Figure 1 As shown, in a preferred embodiment of the present invention, the dual-reflection long-distance laser triangulation system includes a first reflecting mirror 1, a second reflecting mirror 2, an imaging CMOS 3, a receiving lens 4, a paper target 5 to be measured, a laser source focusing module 6, and a laser source 7. The laser beam emitted by the laser source 7 is perpendicularly incident on the paper target 5 to be measured through the focusing module 6. The resulting diffuse reflected light is focused by the receiving lens 4, and after being reflected by the first reflecting mirror 1 and the second reflecting mirror 2, it is imaged on the imaging CMOS 3.

[0060] Preferably, step 1 includes the following steps:

[0061] Step 1.1: Place the object to be measured at 2250mm from the laser emitter, calculate the focal length of the receiving lens, and design the optical path;

[0062] Step 1.2: Based on the principle of reflection, since the imaging optical path after reflection by the reflector unit is symmetrical to the original optical path without the reflector unit, calculate the position parameters of the optical elements in the symmetrical optical path;

[0063] Step 1.3: Construct the actual optical path based on the focal length of the receiving lens and the position parameters of the optical elements in the symmetrical optical path. Set the frame rate of the imaging CMOS 3 to 10Hz and continuously acquire 50 spot images as experimental sample images. Calculate the mean of these 50 images as the noise-free real spot image. Then, construct noisy spot image samples with different noise levels using this method.

[0064] Step 1.4: Convert the experimental sample image to grayscale to obtain a grayscale image sample;

[0065] Step 1.5: First, extract the contour and calculate the geometric center coordinates of the light spot on the grayscale image sample. Then, set an appropriate cropping box and place the light spot at the center of the image to obtain the region of interest image.

[0066] In a preferred embodiment of the present invention, the specific steps of the image logarithmic domain transformation in step 2 are as follows:

[0067] Step 2.1: The following logarithmic transformation process is adopted:

[0068]

[0069] Where m is the normalization parameter.

[0070] Step 2.2: To ensure the numerical stability of the logarithmic transformation and avoid infinity, 1 is added before the logarithmic transformation to ensure that all results after the transformation are greater than 0. This is modified as follows:

[0071]

[0072] Step 2.3: Transform the region of interest image obtained in Step 1 to the logarithmic domain.

[0073] Step 2 involves transferring the light spot image samples obtained in Step 1 and the generated real light spot image to the logarithmic domain through logarithmic transformation, which makes the noise level distribution in the light spot image more uniform and improves the network denoising efficiency.

[0074] In a preferred embodiment of the present invention, step 3 includes the following steps:

[0075] Step 3.1: Use a downsampling operator with a convolution kernel size of 2×2 and a stride of 2 to segment the 1041 pixel × 1053 pixel × 2 light spot image into non-overlapping blocks of size 2×2;

[0076] Step 3.2: Assign the average value of the sub-diagonal pixels of each block to the first pixel of a downsampled feature map D1(y), similar to average pooling. Similarly, assign the average value of the main diagonal pixels to another downsampled feature map D2(y). Preferably, both downsampled feature maps are 520 pixels × 526 pixels × 2.

[0077] In a preferred embodiment of the present invention, the loss function for constructing the neural network includes the following steps:

[0078] Step 4.1: Based on the pair of downsampled feature maps obtained after downsampling the noisy image, fit a neural network f. To reduce the influence of outliers, the loss function is defined as:

[0079] L=||f(D1(y))-D2(y)||1

[0080] Step 4.2: Considering that in residual learning the network is optimized to adapt to noise rather than the image, the loss function is modified as follows:

[0081]

[0082] Step 4.3: To ensure that the results obtained by denoising the noisy image first and then downsampling are similar to those obtained by downsampling first and then denoising, and to improve consistency, the following loss function is adopted:

[0083]

[0084] In a preferred embodiment of the present invention, step 5 includes the following steps:

[0085] Step 5.1: Gradientization to minimize loss L = L res. +L cons. Set reasonable network training parameters. Preferably, the learning rate is set to 0.001, the learning step size is 1500, the learning decay value is 0.5, the number of training rounds is 30 epochs, and the Adam optimizer is used to accelerate the network training process to obtain the trained neural network.

[0086] Step 5.2: Use the trained neural network to obtain the denoised image x = yf(y).

[0087] In a preferred embodiment of the present invention, the repeatability accuracy of the calculation system includes: calculating the position of the spot centroid using the square-weighted centroid method, obtaining the standard deviation of the spot centroid position in the horizontal and vertical directions for 50 images acquired in a single batch, and then determining the repeatability accuracy of the system. The calculation of the spot centroid position using the square-weighted centroid method is as follows:

[0088]

[0089] Calculate the standard deviation of the spot centroid position in the horizontal and vertical directions for 50 images to obtain the system's repeatability error. The spot centroid position offset curves are shown below when the working distances are 1750mm, 2000mm, 2250mm, 2500mm, and 2750mm. Figure 5 As shown, this is to analyze the performance of the denoising method. Figure 3 The graphs show the peak signal-to-noise ratio (PSNR) and root mean square error (RMSE) of images at different noise levels using different methods, compared to the real spot image. Figure 5 The graphs show the root mean square error of the light spot relative to the true position in the horizontal and vertical directions under different noise levels using different methods.

[0090] The spot noise reduction method in this invention can help improve the accuracy of spot center positioning, ensure the accuracy of measurement results and the stable operation of the system.

[0091] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for denoising the laser spot in a long-distance laser triangulation displacement measurement system, characterized in that, Includes the following steps: Step 1: Construct a dual-reflection long-distance laser triangulation system, acquire light spot image samples, and perform grayscale conversion and region of interest extraction on the light spot image samples to obtain the region of interest image. The dual-reflection long-distance laser triangulation system includes a first reflector, a second reflector, an imaging CMOS, a receiving lens, a paper target to be measured, a laser source focusing module, and a laser source. Step 2: Perform a logarithmic transformation on the region of interest image to make the noise level of the transformed image uniformly distributed; Step 3: Decompose the transformed image into a pair of downsampled feature maps using a downsampling operator; Step 4: Considering the impact of outliers, residual learning, and the principle of symmetric loss, construct the loss function for the neural network; The loss function for constructing the neural network includes the following steps: Step 4.1: Based on the pair of downsampled feature maps obtained after downsampling the noisy image, fit a neural network f. To reduce the influence of outliers, the loss function is defined as: Step 4.2: Considering that in residual learning the network is optimized to adapt to noise rather than the image, the loss function is modified as follows: Step 4.3: To ensure that the results obtained by denoising the noisy image first and then downsampling are similar to those obtained by downsampling first and then denoising, and to improve consistency, the following loss function is adopted: ； Step 5: By minimizing the loss function of the neural network, a lightweight neural network consisting of two convolutional layers is trained, and one downsampled feature map is mapped to another downsampled feature map to obtain the trained lightweight neural network. Step 6: Input the original spot image into the lightweight neural network trained in Step 5 to obtain the denoised spot image, and calculate the repeatability accuracy of the system.

2. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 1, characterized in that, The laser beam emitted by the laser source is incident perpendicularly on the paper target being tested through the focusing module. The diffuse reflected light is focused by the receiving lens, and after being reflected by the first and second reflecting mirrors, it is imaged on the CMOS imaging device.

3. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 2, characterized in that, Step 1 includes the following steps: Step 1.1: Place the object to be measured at 2250mm from the laser emitter, calculate the focal length of the receiving lens, and design the optical path; Step 1.2: Based on the principle of reflection, since the imaging optical path after reflection by the reflector unit is symmetrical to the original optical path without the reflector unit, calculate the position parameters of the optical elements in the symmetrical optical path; Step 1.3: Based on the focal length of the receiving lens and the position parameters of the optical elements in the symmetrical optical path, construct the actual optical path, set the camera frame rate to 10Hz, and continuously acquire 50 light spot images as experimental sample images. Step 1.4: Convert the experimental sample image to grayscale to obtain a grayscale image sample; Step 1.5: First, extract the contour and calculate the geometric center coordinates of the light spot on the grayscale image sample. Then, set an appropriate cropping box and place the light spot at the center of the image to obtain the region of interest image.

4. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 1, characterized in that, The logarithmic transformation in step 2 is as follows: Where m is the normalization parameter, i represents the x-coordinate of the pixel, and j represents the y-coordinate of the pixel.

5. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Use a downsampling operator with a convolution kernel size of 2×2 and a stride of 2 to segment the 1041 pixel × 1053 pixel × 2 light spot image into non-overlapping blocks of size 2×2; Step 3.2: Assign the average value of the sub-diagonal pixels of each block to the first pixel of a downsampled feature map D1(y), similar to average pooling. Similarly, assign the average value of the main diagonal pixels to another downsampled feature map D2(y).

6. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 5, characterized in that, Both downsampled feature maps have a shape of 520 pixels × 526 pixels × 2.

7. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 6, characterized in that, Step 5 includes the following steps: Step 5.1: Gradientize the minimum loss L = L res. + L cons. , set reasonable network training parameters, and use the Adam optimizer to accelerate the network training process to obtain the trained neural network; Step 5.2: Use the trained neural network to obtain the denoised image x = y - f(y).

8. The spot noise reduction method for the long-distance laser triangulation displacement measurement system as described in claim 1, characterized in that, The repeatability accuracy of the calculation system includes: using the square-weighted centroid method to calculate the position of the spot centroid, obtaining the standard deviation of the spot centroid position in the horizontal and vertical directions for 50 images acquired in a single acquisition, and then obtaining the repeatability accuracy of the system.

Images