Striped tube laser radar image fine-grained denoising method

Abstract

The present application relates to a kind of streak tube laser radar image fine-grained denoising method, belong to streak tube imaging laser radar denoising technical field, comprising: obtaining the streak image of target under different postures;The system signal-to-noise ratio of streak image under different postures is calculated;According to system signal-to-noise ratio, the streak image is handled to obtain noisy streak image and not noisy streak image;Noisy streak image and not noisy streak image are input into WACAFRN model as data set to train and obtain denoising model;The denoising model is used to complete the denoising of target streak image.The present application can effectively retain the details and edge information in streak image while suppressing noise by using WACAFRN model and training a large number of pairs of "noisy / noise-free streak" data.Compared with traditional filtering method, the present application can better avoid streak blur and detail loss, thereby improving the accuracy of subsequent phase calculation and target ranging.

Description

Technical Field

[0001] This invention relates to the field of striped tube imaging lidar denoising technology, specifically to a fine-grained denoising method for striped tube lidar images. Background Technology

[0002] As an advanced 3D imaging system, stripe tube lidar boasts advantages such as a large field of view, high sensitivity, and long detection range. Its core lies in utilizing stripe tubes to perform high-precision time-space conversion of optical signals, thereby obtaining the target's distance, shape, and even dynamic information. Based on the number of slits, stripe tube imaging lidar can be further divided into single-slit stripe tube imaging lidar and multi-slit stripe tube imaging lidar.

[0003] As can be seen from the working principle of multi-slit stripe LiDAR, the stripe image acquired by CCD contains various types of noise, which can interfere with image reconstruction. Therefore, denoising the stripe image is essential before image reconstruction. A stripe image consists of many light spots of varying brightness, with distinct center and edge features. These light spots contain intensity and depth information of the target. Therefore, the unique characteristics of stripe images place higher demands on their denoising techniques. Specifically, while removing noise from the stripe image, it is crucial to avoid blurring information points in the image and to preserve stripe details as much as possible for feature point extraction. Simultaneously, the algorithm should not be overly complex to avoid affecting the final imaging speed.

[0004] Currently, traditional algorithms, such as Gaussian filtering and median filtering, are still more commonly used for denoising striped images. These algorithms are relatively simple and fast, but their denoising effect is usually mediocre. While denoising, they damage the details and stripe structure of the original image, resulting in a less than ideal reconstructed image. Summary of the Invention

[0005] To address the aforementioned problems, the purpose of this invention is to provide a fine-grained denoising method for stripe tube lidar images.

[0006] A fine-grained denoising method for striped laser radar images includes:

[0007] Step 1: Obtain stripe images of the target in different poses;

[0008] Step 2: Calculate the system signal-to-noise ratio of the stripe images under different poses;

[0009] Step 3: Process the stripe image according to the system signal-to-noise ratio to obtain a noisy stripe image and a noise-free stripe image;

[0010] Step 4: Input the noisy stripe image and the noisy stripe image as datasets into the WACAFRN model for training to obtain the denoising model;

[0011] Step 5: Use the denoising model to denoise the target stripe image.

[0012] Preferably, step 2: calculating the system signal-to-noise ratio of the stripe image under different poses includes:

[0013] Step 4.1: Calculate the average number of photons produced by the receiver based on the number of photons falling on the receiver;

[0014] Step 4.2: Calculate the laser speckle noise variance based on the average number of photons generated by the receiver;

[0015] Step 4.3: Calculate the standard charge deviation based on the detector circuit temperature and the detector circuit capacitance;

[0016] Step 4.4: Calculate the number of photons generated by the background light based on the intensity of the background light;

[0017] Step 4.5: Determine the system signal-to-noise ratio based on the average number of photons generated by the receiver, the variance of laser speckle noise, the variance of thermal noise, and the number of photons generated by the background light.

[0018] Preferably, in step 4.1, the formula for calculating the average number of photons generated by the receiver is:

[0019]

[0020] in, Indicates the expectation. This represents the average number of photons produced by the receiver. This indicates the quantum efficiency of the receiver. This represents the number of photons that fall on the receiver.

[0021] Preferably, in step 4.2, the formula for calculating the variance of laser speckle noise is:

[0022]

[0023] in, This represents the variance of laser speckle noise. It is the number of photons detected by the system. Represents the degrees of freedom of a laser.

[0024] Preferably, in step 4.3, the formula for calculating the standard charge deviation is:

[0025]

[0026] in, It is the standard charge deviation of the number of thermal noise electrons. It is Boltzmann's constant. It is the temperature of the detector circuit. It is the capacitor of the detection circuit. It is the elementary charge.

[0027] Preferably, in step 4.4, the background noise variance is:

[0028]

[0029] in, The number of photons generated by the background light. The intensity of the target background light. The electromagnetic bandwidth of an optical electromagnetic bandwidth bandpass filter. For the target area, The surface reflectivity of the target. For quantum efficiency, The atmospheric transport coefficient, For the system to receive optical transmission coefficient, To receive the optical aperture, This refers to the distance between the lidar and the target. Let be Planck's constant. For the frequency of light, The number of photons generated by the receiver under the influence of dark current.

[0030] Preferably, the WACAFRN model includes: PFEB and ACA modules; PFEB consists of three convolutional / dilated convolutional layers for extracting initial features; the ACA module is used to apply adaptive average pooling to the input feature map in the horizontal and width directions respectively, then the results of the adaptive average pooling operations in the horizontal and width directions are adjusted by MLP and attention intensity factor, and finally, the results in the horizontal and width directions are summed and attention weights are generated by the sigmoid function.

[0031] Preferably, the loss function during the training of the WACAFRN model is:

[0032]

[0033]

[0034]

[0035]

[0036] in, Represents the loss function. Indicates the reconstruction loss. Represents gradient loss, Indicates wavelet domain loss, , and Indicates weight, For network output, For real labels, This represents the horizontal and vertical gradients calculated by the Sobel operator. These represent the four sub-band coefficients obtained from a single-layer discrete wavelet transform. Represents approximate components, Indicates the horizontal detail component. Indicates the vertical detail component. This represents the diagonal detail component.

[0037] The present invention also provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected via the bus, characterized in that the computer program, when executed by the processor, implements the steps in the above-described fine-grained denoising method for stripe laser radar images.

[0038] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps in the above-described method for fine-grained denoising of stripe laser radar images.

[0039] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] This invention relates to a fine-grained denoising method for stripe laser radar images. Compared with existing technologies, this invention utilizes the deep learning model WACAFRN, trained with a large number of paired "noisy / noisy stripe" data, which can effectively preserve details and edge information in the stripe image while suppressing noise. Compared with traditional filtering methods (such as mean filtering, Gaussian filtering, etc.), this method can better avoid the problems of stripe blurring and detail loss, thereby improving the accuracy of subsequent phase calculation and target ranging.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart of a fine-grained noise reduction method for stripe tube lidar images provided by the present invention;

[0044] Figure 2 This is a schematic diagram illustrating the acquisition of target depth image data provided by the present invention;

[0045] Figure 3 A schematic diagram of a noisy stripe image and a noise-free stripe image of a vehicle target in a certain posture, provided by the present invention.

[0046] Figure 4 Striped images of four target parts—car, house, airplane, and tree—in different poses, provided for this invention.

[0047] Figure 5 Denoising striped images of four target parts—car, house, airplane, and tree—in different poses, provided by the present invention.

[0048] Figure 6 Detailed structural diagram of PFEB provided for this invention;

[0049] Figure 7 This is a schematic diagram of the ACA module provided by the present invention;

[0050] Figure 8 Comparison of stripe images using different denoising methods provided by this invention;

[0051] Figure 9 This is a schematic diagram of a trapezoidal target provided by the present invention;

[0052] Figure 10 The trapezoidal target depth image provided by this invention. Detailed Implementation

[0053] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0054] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0055] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0056] Please see Figure 1 A fine-grained denoising method for striped laser radar images, comprising:

[0057] Step 1: Obtain stripe images of the target in different poses;

[0058] Step 2: Calculate the system signal-to-noise ratio of the stripe images under different poses;

[0059] In step 2, as a type of lidar, the stripe tube imaging lidar inevitably experiences various types of noise during operation, which severely affects the imaging quality of the stripe image. The sources of noise in the system are complex; any factor that may affect the transmitted laser waveform can be considered noise. Typically, the noise in a stripe tube imaging lidar mainly includes: photon counting noise, laser speckle noise, thermal noise, and background noise.

[0060] 1. Photon counting noise

[0061] Number of photons falling on the receiver It is a random variable, and its mean can be calculated as:

[0062] (2.1)

[0063] in, Indicates the expectation. It is the energy received by the detector. It is the integration time of the detection circuit. It is Planck's constant. It refers to the light frequency. The average number of photons produced by the receiver. The quantum efficiency of the receiver can be calculated as the number of photons and the receiver's quantum efficiency. The product of:

[0064] (2.2)

[0065] The expression for the variance of photocurrent is:

[0066] (2.3)

[0067] in, It is the elementary charge. It is the bandwidth of the detection circuit.

[0068] 2. Laser speckle noise

[0069] Laser speckle refers to the irregular intensity distribution of a laser beam on a target surface due to diffuse reflection or passing through a transparent scattering body, resulting in randomly distributed bright and dark spots. Since laser speckle noise occurs randomly, it can only be studied using probabilistic and statistical methods. Goodman argues that during lidar detection, laser speckle noise follows a negative binomial distribution, and its variance can be expressed as:

[0070] (2.4)

[0071] in, It is the number of photons detected by the system. The degrees of freedom representing laser light, for perfectly coherent light, The value of is 1, and for completely incoherent light, its value is close to infinity. Therefore, As a physical quantity characterizing laser coherence during lidar detection, it includes both temporal coherence and spatial coherence.

[0072] 3. Thermal noise

[0073] Any temperature greater than Objects of size K radiate photons and generate thermal noise. That is, thermal noise exists in all electronic components and propagation media. During operation, the detector generates heat, and therefore also emits a certain number of photons. These photons constitute the detector's thermal noise, and their charge variance can be expressed as:

[0074] (2.5)

[0075] in, It is the standard charge deviation of the number of thermal noise electrons. It is Boltzmann's constant. It is the temperature of the detector circuit. It is the capacitor in the detection circuit.

[0076] 4. Background noise

[0077] Background noise in lidar refers to the light or signals received by the receiver that are not from the laser emitter, and it exists in most optical systems. In long-range lidar detection, background noise is the dominant noise. Its noise variance is equal to the number of photoelectrons present in the background radiation, and its statistical characteristics typically follow a Poisson distribution. The expression for the expected number of electrons is:

[0078] (2.6)

[0079] in, The number of photons generated by the background light. The intensity of the target background light. The electromagnetic bandwidth of an optical electromagnetic bandwidth bandpass filter. For the target area, The surface reflectivity of the target. For quantum efficiency, The atmospheric transport coefficient, For the system to receive optical transmission coefficient, To receive the optical aperture, This refers to the distance between the lidar and the target. is Planck's constant. For the frequency of light, The number of photons generated by the receiver under the influence of dark current.

[0080] 5. System signal-to-noise ratio

[0081] The signal-to-noise ratio (SNR) of a stripe tube imaging lidar is the ratio of the number of effective photons in the system to the number of photons generated by noise. Since the various noise sources are statistically independent, the standard deviation of the noise can be approximated as the square root of the sum of the variances of the individual noise sources. Therefore, the expression for the system's signal-to-noise ratio (SNR) is:

[0082]

[0083] Step 3: Process the stripe image according to the system signal-to-noise ratio to obtain a noisy stripe image and a noise-free stripe image;

[0084] In deep learning-based image denoising methods, the dataset is a core element, and its quality can even determine the denoising result. Therefore, to improve the denoising effect of the system, this invention constructs a dataset suitable for this system based on the system signal-to-noise ratio. This striped image dataset mainly consists of two parts: the original striped image and the noisy striped image.

[0085] To obtain as many stripe images as possible of different targets commonly seen in real life, a data acquisition experiment was designed.

[0086] When illuminating different objects, their postures are not the same. In order to collect different postures of different targets, this invention builds a data acquisition platform, places the stripe tube lidar directly above the target, and then rotates around the object according to a certain pattern to take pictures. This change can obtain images of different postures of the target in sequence.

[0087] Considering the characteristics and application scenarios of stripe tube imaging lidar, 150 common outdoor targets will be selected as the base sample for the dataset. The dataset required for this study will be established by acquiring stripe images of different targets. After the targets are determined, data acquisition is required. The stripe tube imaging lidar system will be used to individually illuminate and image each selected target. Detailed parameters of the stripe tube imaging lidar system are shown in Table 1.

[0088] Table 1

[0089] System parameters numerical values System parameters numerical values Pulse energy 0.2J Number of slits 16 laser wavelength 532nm Distance gating 25.6m Laser pulse width 10ns Atmospheric transport 0.95 Laser beam divergence angle 0.012° Background noise <![CDATA[100W / m 2 ]]> Dark current <![CDATA[10 -9 A]]> detector quantum efficiency 0.075 Fiber diameter 125μm Reconstructing image pixels 64×64

[0090] In real-world operating environments, the relative position between the stripe tube imaging lidar system and the target is uncertain. Taking a car as an example, the system's illumination position might be in front of the car, or on top of it, or to its side. Therefore, to comprehensively simulate the scenario where the target is detected by the lidar system from different positions, the following scheme was designed, such as... Figure 2As shown, the stripe tube imaging lidar system is placed directly above the target, and then the target is rotated in an orderly manner to change its attitude. Theoretically, any attitude of the target can be obtained by rotating it around the three mutually perpendicular axes X, Y, and Z.

[0091] Under normal circumstances, the lidar cannot illuminate the bottom of the target, therefore the rotation angle range of the three axes is set, with the yaw angle being... arrive ( ), pitch angle is arrive ( ), roll angle is arrive ( To obtain images of the target being illuminated from different directions, the target is rotated sequentially around three mutually perpendicular axes, with each rotation being an angle of _____. An image is created every time the angle of an axis changes. Figure 3 These are noisy and noiseless striped images of a car target in a specific pose.

[0092] The stripe image dataset corresponding to some targets, such as Figure 4-5 As shown. Figure 4 The striped images correspond to different poses of four target parts: cars, houses, airplanes, and trees. Figure 5 These are the denoised stripe images corresponding to them.

[0093] Step 4: Input the noisy stripe image and the noisy stripe image as datasets into the WACAFRN model for training to obtain the denoising model;

[0094] Traditional denoising methods rely on total variation, sparse coding, and gradient priors, while techniques such as BM3D and nonlocal methods utilize self-similarity in images, which can sometimes lead to loss of detail. In contrast, modern advancements in artificial intelligence have introduced neural network-based methods, such as DnCNN, HINet, and CBDNet, which utilize nonlinear activation functions like ReLU and LeakyReLU to better preserve image details and mitigate gradient vanishing. Furthermore, attention mechanisms in models like RIDNet and MIRNet, and Transformer-based methods like SwinIR, further improve performance by effectively utilizing global image information. Denoising network architectures vary from single-level [7,10,13,14] to multi-level [9,11,15], each with its own advantages in handling noise and preserving detail. They are typically built upon the encoder-decoder framework exemplified by the U-net structure, which captures multi-scale information. Furthermore, as shown in ResNet and DANet, the integration of skip connections enables the network to retain the low-level features necessary for image reconstruction and enhance detail restoration. To improve the system's denoising capability, convolutional neural networks (CNNs) have achieved significant performance in image denoising, particularly considering the characteristics of striped images. However, most existing CNNs cannot accurately capture and remove minute noise during denoising, easily losing edge detail information. To mitigate this shortcoming, this invention selects the latest WACAFRN (Wavelet and Adaptive Coordinate Attention Guided Fine-Grained Residual Network) network for image denoising. This model employs an adaptive coordinate attention mechanism and combines it with cascaded Res2Net residual blocks to form an encoder network for more accurate noise removal. Secondly, the model employs a wavelet attention mechanism, which combines global and local residual blocks to form a decoder network, aiming to address the problem of edge detail loss. Figure 6 As shown, FEB consists of three convolutional / dilated convolutional layers for extracting initial features. The ACA module applies adaptive average pooling to the input feature map in both the horizontal and width directions. Subsequently, the results of these two adaptive average pooling operations undergo transformations such as MLP and attention intensity factor adjustment. Finally, the results in both directions are summed, and attention weights are generated using the sigmoid function.

[0095] The first part of WACAFRN is the main feature extraction block, which provides initial shallow feature information. The main feature extraction block consists of three layers: a convolutional layer, a dilated convolutional layer with a dilation factor of 2, and a convolutional layer. A detailed structural diagram of PFEB is shown below. Figure 7 As shown. In Figure 7 In this model, the PFEB block takes a noisy image as input and outputs the extracted main feature map. To provide a more intuitive view of the extracted features, the output feature map is displayed as a thermodynamic image. By observing... Figure 7It can be observed that the main feature maps extracted by PFEB can cover most of the noise locations and provide better shallow features for the model. This can be expressed by the following formula:

[0096]

[0097] in, The input is a noisy image. The function representing the first convolutional layer, This is the function of the second convolutional layer, and also the performance of the first extended convolutional layer. This is the function of the second convolutional layer.

[0098] To effectively extract finer-grained noise information and remove as much noise as possible, this invention designs a fine-grained encoder consisting of three fine-grained coding layers (FGELs). Each FGEL includes the following operations: convolution with altered feature channels, a fine-grained feature extraction block consisting of stacked Res2Net blocks and an adaptive coordinate attention mechanism, and downsampling to reduce image resolution. The processing of the FGE can be represented as follows:

[0099]

[0100] in, fS1 is the output of PFEB, f1, f2 and f3 are convolution functions, representing the number of channels from 64 to 64, 64 to 128 and 128 to 256 respectively, fS1, fS2 and fS3 are fine-grained feature extraction blocks, and OFGE is the output of FGE.

[0101] Loss function; Joint loss function

[0102]

[0103] (1) Reconstruction Loss )

[0104] Function: To ensure that the denoised image is consistent with the real noise-free image (Ground Truth) at the pixel level, it is a fundamental constraint.

[0105]

[0106] For network output, For accurate labeling, the mean absolute error (MAE) of the depth map and intensity map must be calculated separately and then summed.

[0107] (2) Gradient Loss )

[0108] Specifically optimized for the ACA (Adaptive Coordinate Attention) mechanism in this patent. This forces the network to learn the gradient information (edges and texture) of the image, preventing edge blurring during denoising.

[0109]

[0110] This represents the horizontal and vertical gradients calculated by the Sobel operator.

[0111] (3) Wavelet Domain Loss )

[0112] This echoes the core innovation of this patent, the "wavelet attention mechanism." It constrains high-frequency components (details / noise) and low-frequency components (structure) in the frequency domain to ensure accurate coefficient distribution after wavelet decomposition.

[0113]

[0114] These represent the four sub-band coefficients obtained from a single-layer discrete wavelet transform.

[0115] Step 5: Use the denoising model to denoise the target stripe image.

[0116] To verify the effectiveness of the proposed method, it was compared with traditional stripe image denoising methods, median filtering, and BM3D. In the DCNN denoising algorithm, the dataset used was also the dataset constructed in this invention. This invention images a tank target and then uses different methods for denoising, with the results shown below. Figure 8 As shown in the figure, a is the original image, b is the denoised image obtained by the median filtering algorithm, c is the image obtained by BM3D, and d is the image obtained by the algorithm of this invention.

[0117] Figure 8This section presents a comparison of stripe images using different denoising methods. Screenshots from three methods—BM3D, CNN, and the method disclosed herein—are shown for comparison. It can be seen that the red-framed areas in each stripe image represent magnified local images. After denoising using the three different methods, the noise in the images is significantly removed. To demonstrate the effectiveness of the denoising method, the four stripe images are reconstructed to obtain intensity and depth images. Figure ef represents the intensity image, and Figure il represents the depth image. It can be seen that the differences between the intensity images are relatively small; however, Figure a, obtained through the network model, has a significantly better visual effect. The differences between the depth images are more pronounced. The depth image reconstructed from the original stripe image without denoising processing has very poor image quality, only showing the general outline of the target. The depth images reconstructed from the denoised stripe images reveal the target details, and the target image obtained by the network model method a has the clearest details. Overall, the intensity and depth images corresponding to the method proposed in this invention are the best.

[0118] However, the aforementioned tank imaging experiment only demonstrated the effectiveness of the method of this invention in denoising striped images from a visual perspective. To more objectively verify the effectiveness of the proposed method, this invention designed another trapezoidal target imaging experiment, as shown in the figure:

[0119] Table 2 Reflectivity of Four Different Materials

[0120] Material lime cork iron plate Aluminum plate reflectivity 0.11 0.19 0.31 0.68

[0121] The target is a trapezoidal target, consisting of four layers, each a rectangle 4m long and 1m wide, with a 1m interval between layers. The reflectivities are a, b, c, and d, respectively. The trapezoidal target is then illuminated, and the resulting fringe pattern is shown below. Figure 10 As shown in (a), denoising was then performed using four methods: median filtering, BM3D, DnCNN, and WACAFRN, corresponding to stripe images 10(b)-(e). Compared to the tank experiment, DnCNN network model denoising was added to verify the superiority of the proposed method based on neural network models. Then, image reconstruction was performed on the denoised images obtained by the different methods. The reconstruction process used peak power analysis, where the maximum value in each light spot was used as the feature point of that spot. The intensity image and depth image were reconstructed using the brightness and distance information of the feature points. The final intensity images corresponding to different methods are shown below. Figure 10 (f)-(j), depth images such as Figure 10 (k)-(o).

[0122] This invention utilizes the deep learning model WACAFRN, trained with a large number of paired "noisy / noisy stripe" data, to effectively preserve details and edge information in stripe images while suppressing noise. Compared to traditional filtering methods (such as mean filtering and Gaussian filtering), this method better avoids stripe blurring and detail loss, thereby improving the accuracy of subsequent phase calculation and target ranging.

[0123] The present invention also provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor. The transceiver, the memory, and the processor are connected via the bus. The computer program, when executed by the processor, implements the steps in the above-described fine-grained denoising method for striped laser radar images. Compared with the prior art, the beneficial effects of the electronic device provided by the present invention are the same as those of the fine-grained denoising method for striped laser radar images described above, and will not be elaborated further here.

[0124] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps in the above-described method for fine-grained denoising of striped laser radar images. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present invention are the same as the beneficial effects of the above-described method for fine-grained denoising of striped laser radar images, and will not be elaborated here.

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fine-grained noise reduction method for stripe tube lidar images, characterized in that, include: Step 1: Obtain stripe images of the target in different poses; Step 2: Calculate the system signal-to-noise ratio of the stripe images under different poses; Step 3: Process the stripe image according to the system signal-to-noise ratio to obtain a noisy stripe image and a noise-free stripe image; Step 4: Input the noisy stripe image and the noisy stripe image as datasets into the WACAFRN model for training to obtain the denoising model; Step 5: Use the denoising model to denoise the target stripe image.

2. The method for fine-grained noise reduction of stripe tube lidar images according to claim 1, characterized in that, Step 2: Calculate the system signal-to-noise ratio of the stripe image under different poses, including: Step 4.1: Calculate the average number of photons produced by the receiver based on the number of photons falling on the receiver; Step 4.2: Calculate the laser speckle noise variance based on the average number of photons generated by the receiver; Step 4.3: Calculate the standard charge deviation based on the detector circuit temperature and the detector circuit capacitance; Step 4.4: Calculate the number of photons generated by the background light based on the intensity of the background light; Step 4.5: Determine the system signal-to-noise ratio based on the average number of photons generated by the receiver, the variance of laser speckle noise, the variance of thermal noise, and the number of photons generated by the background light.

3. The fine-grained noise reduction method for stripe tube lidar images according to claim 2, characterized in that, In step 4.1, the formula for calculating the average number of photons generated by the receiver is: in, Indicates the expectation. This represents the average number of photons produced by the receiver. This indicates the quantum efficiency of the receiver. This represents the number of photons that fall on the receiver.

4. The fine-grained noise reduction method for stripe tube lidar images according to claim 3, characterized in that, In step 4.2, the formula for calculating the variance of laser speckle noise is: in, This represents the variance of laser speckle noise. It is the number of photons detected by the system. Represents the degrees of freedom of a laser.

5. The fine-grained noise reduction method for stripe tube lidar images according to claim 4, characterized in that, In step 4.3, the formula for calculating the standard charge deviation is: in, It is the standard charge deviation of the number of thermal noise electrons. It is Boltzmann's constant. It is the temperature of the detector circuit. It is the capacitor of the detection circuit. It is the elementary charge.

6. The fine-grained noise reduction method for stripe tube lidar images according to claim 5, characterized in that, In step 4.4, the background noise variance is: in, The number of photons generated by the background light. The intensity of the target background light. The electromagnetic bandwidth of an optical electromagnetic bandwidth bandpass filter. For the target area, The surface reflectivity of the target. For quantum efficiency, The atmospheric transport coefficient, For the system to receive optical transmission coefficient, To receive the optical aperture, This refers to the distance between the lidar and the target. is Planck's constant. For the frequency of light, The number of photons generated by the receiver under the influence of dark current.

7. The method for fine-grained denoising of striped laser radar images according to claim 1, characterized in that, The WACAFRN model includes the PFEB and ACA modules. The PFEB consists of three convolutional / dilated convolutional layers used to extract initial features. The ACA module is used to apply adaptive average pooling to the input feature map in the horizontal and width directions respectively. Then, the results of the adaptive average pooling operations in the horizontal and width directions are adjusted by MLP and attention intensity factor. Finally, the results in the horizontal and width directions are summed and attention weights are generated by the sigmoid function.

8. The fine-grained noise reduction method for stripe tube lidar images according to claim 7, characterized in that, The loss function used in training the WACAFRN model is: in, Represents the loss function. Indicates the reconstruction loss. Represents gradient loss, Indicates wavelet domain loss, , and Indicates weight, For network output, For real labels, This represents the horizontal and vertical gradients calculated by the Sobel operator. These represent the four sub-band coefficients obtained from a single-layer discrete wavelet transform. Represents approximate components, Indicates the horizontal detail component. Indicates the vertical detail component. This represents the diagonal detail component.

9. An electronic device comprising a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected via the bus, characterized in that, When the computer program is executed by the processor, it implements the steps in the fine-grained denoising method for stripe tube lidar images as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the fine-grained denoising method for stripe tube lidar images as described in any one of claims 1-8.

Images