Laser radar signal denoising method and system based on double-residual neural network

Abstract

The invention provides a laser radar signal de-noising method and system based on a double-residual neural network, and belongs to the technical field of CCD imaging laser radar, and the method comprises the steps: inputting a simulation image into a double-residual neural network model, extracting features through a plurality of residual convolution modules, obtaining a feature map of an expanded receptive field through a first convolution module, and obtaining a second convolution module; and further extracting deep-level features from the feature map with the enlarged receptive field, and fusing the deep-level features with shallow-level features to obtain a de-noised image. For the echo information in the laser radar signal, the residual connection can ensure that the key information of the echo signal is not completely eliminated or distorted during the de-noising operation, thereby protecting the effectiveness of the echo signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of CCD imaging lidar, and particularly relates to a lidar signal denoising method and system based on a dual residual neural network. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] The vertical profile of the extinction coefficient of aerosols is crucial for evaluating the distribution and transmission of air pollution. Lidar remote sensing is the most commonly used technology for detecting the vertical structure of the atmosphere at present. As an active remote sensing tool, lidar can achieve all-day, all-weather, and fully automatic detection, and is widely used in the fields of meteorology, environmental protection, aviation, etc. The technology based on CCD detection of laser beam imaging is a new detection technology. It places the transmitting optical path and the receiving CCD at two different locations, avoiding the generation of the transition zone and blind zone in the backscatter lidar, and has high measurement accuracy in the short-distance section. From the lidar equation, the effective detection distance of the lidar system is mainly limited by the signal-to-noise ratio, and the signal-to-noise ratio will rapidly decrease as the detection distance increases. At the same time, during the image sensing process, noise pollution seriously reduces the visual quality of the acquired images. Removing noise from the observed images is an important method to improve the signal-to-noise ratio. The purpose of image denoising is to recover the noise-free image from the noise observations that follow the image degradation model. Therefore, it is of great significance to study and optimize image denoising algorithms.

[0004] Common denoising methods include wavelet denoising, filter denoising, and neural network noise prediction. Wavelet transform has the advantages of low entropy, multi-resolution analysis, decorrelation, etc. However, the wavelet threshold denoising algorithm faces many problems, and the selection and calculation process of the threshold are relatively complex. Wiener filtering is a commonly used linear filtering method for restoring signals contaminated by noise. It convolves the observed signal with the filter by minimizing the mean square error to restore the original signal. Its disadvantages include high requirements for the statistical characteristics of signals and noise, and poor performance in dealing with non-linear and non-Gaussian noise. Summary of the Invention

[0005] To overcome the deficiencies of the above-mentioned prior art, the present invention proposes a lidar signal denoising method and system based on a dual residual neural network. The CCD imaging lidar is used to obtain the lateral lidar signal and restore the original simulation image according to the beam width and Gaussian distribution characteristics, and then the convolutional neural network is used to achieve the denoising effect on the simulation image.

[0006] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, a method for denoising lidar signals based on a dual residual neural network is disclosed, including: Collect lidar signals; Preprocess the lidar signals to obtain a simulation image; Use a dual residual neural network model to denoise the simulation image; Specifically, the simulation image is input into the dual residual neural network model, and the first layer of feature maps is obtained through the first residual convolution module. The first layer of feature maps is input into the second residual convolution module to obtain the second layer of feature maps. The second layer of feature maps is input into the third residual convolution module to obtain the third layer of feature maps. The third layer of feature maps is input into the first convolution module to obtain the third layer of feature maps with an enlarged receptive field. The third layer of feature maps with an enlarged receptive field is fused with the third layer of feature maps passing through the first attention module and then input into the fourth residual convolution module to obtain the fourth layer of feature maps. The fourth layer of feature maps is fused with the second layer of feature maps passing through the first attention module and then input into the fifth residual convolution module to obtain the fifth layer of feature maps. The fifth layer of feature maps is fused with the second layer of feature maps passing through the first attention module and then input into the sixth residual convolution module to obtain the denoised image.

[0007] In a second aspect, a lidar signal denoising system based on a dual residual neural network is disclosed, including: A collection unit configured to collect lidar signals; A preprocessing unit configured to preprocess lidar signals to obtain a simulation image; A denoising unit configured to use a dual residual neural network model to denoise the simulation image; Specifically, the simulation image is input into the dual residual neural network model, and the first layer of feature maps is obtained through the first residual convolution module. The first layer of feature maps is input into the second residual convolution module to obtain the second layer of feature maps. The second layer of feature maps is input into the third residual convolution module to obtain the third layer of feature maps. The third layer of feature maps is input into the first convolution module to obtain the third layer of feature maps with an enlarged receptive field. The third layer of feature maps with an enlarged receptive field is fused with the third layer of feature maps passing through the first attention module and then input into the fourth residual convolution module to obtain the fourth layer of feature maps. The fourth layer of feature maps is fused with the second layer of feature maps passing through the first attention module and then input into the fifth residual convolution module to obtain the fifth layer of feature maps. The fifth layer of feature maps is fused with the second layer of feature maps passing through the first attention module and then input into the sixth residual convolution module to obtain the denoised image.

[0008] In a third aspect, an electronic device is disclosed, including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps of the above-mentioned method for denoising lidar signals based on a dual residual neural network are completed.

[0009] Fourthly, a computer-readable storage medium is disclosed, which is used to store computer instructions. When the computer instructions are executed by a processor, the steps of the above lidar signal denoising method based on the dual residual neural network are completed.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a lidar signal denoising method based on the dual residual neural network, which uses a deeper network structure to enhance the non-linear expression ability, designs a robust loss function for non-Gaussian noise, and simulates complex noise mixing in the training data to improve the generalization ability. In terms of the network structure, the attention mechanism can help the model focus on the noise area, especially when dealing with non-uniformly distributed noise. Residual learning can separate noise and signals, especially when the noise and signals are non-linearly coupled.

[0011] The present invention uses Wiener filtering, VGG16 and DnCNN for denoising performance comparison experiments. The results of multiple experiments prove that the U²Net network model has a good denoising effect on the lateral lidar signal, the signal-to-noise ratio of the echo signal is improved, and the denoised image is closest to the original image.

[0012] The denoising method proposed by the present invention can better protect the echo signal. After extracting the signal features through multiple convolutional layers, the attention mechanism is used to adjust the signal feature weights, and finally the neural network is trained by the residual learning method, which can effectively remove the noise in the image and better protect the echo signal. For the echo information in the lidar signal, the residual connection can ensure that the key information of the echo signal will not be completely eliminated or distorted during the denoising operation, thus protecting the effectiveness of the echo signal. During the denoising process of the lidar simulation signal diagram, the attention mechanism can highlight the key areas of the echo signal and at the same time suppress the influence of noise. In this way, the model can maintain the accuracy and structure of the echo signal during the denoising process.

[0013] The present invention meets the characteristics of good denoising effect and less distortion, and has a good denoising effect on lidar images; it is of great significance for the accurate inversion of atmospheric optical properties based on lateral lidar signals in the future.

[0014] The advantages of the additional aspects of the present invention will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present invention. Description of the Drawings

[0015] The specification drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0016] Figure 1It is a flowchart of the lidar signal denoising method based on the double residual neural network described in Embodiment 1 of the present invention.

[0017] Figure 2 It is a schematic structural diagram of the CCD imaging lidar described in Embodiment 1 of the present invention.

[0018] Figure 3 It is a schematic diagram of the laser beam image described in Embodiment 1 of the present invention.

[0019] Figure 4 It is a lidar signal simulation image described in Embodiment 1 of the present invention.

[0020] Figure 5 It is a schematic diagram of the U²Net network structure model based on the double residual attention mechanism described in Embodiment 1 of the present invention.

[0021] Figure 6 It is a schematic structural diagram of the RSUA-L module described in Embodiment 1 of the present invention.

[0022] Figure 7 It is a schematic structural diagram of the RSUA-4D module described in Embodiment 1 of the present invention.

[0023] Figure 8 It is a schematic structural diagram of the RAM attention module described in Embodiment 1 of the present invention.

[0024] Figure 9 It is a comparison diagram of the images after denoising in the simulation experiment described in Embodiment 1 of the present invention.

[0025] Figure 10 It is a comparison diagram of the real image and the simulated denoised image of the number of photons in the simulation experiment described in Embodiment 1 of the present invention. Detailed implementation mode

[0026] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0027] It should be noted that the terms used herein are only for describing specific implementation modes and are not intended to limit the exemplary implementation modes according to the present invention.

[0028] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0029] Embodiment 1 In one or more implementation modes, a lidar signal denoising method based on a double residual neural network is disclosed, as Figure 1As shown in the figure, it includes the following steps: Step 1: As Figure 2-3 shown in the figure, the CCD imaging lidar collects the laser beam image, and records the number of signal photons on the beam in different position pixels to form a two-dimensional image.

[0030] Step 1-1: According to the actual use environment, configure the radar parameters, mainly system settings and calibration. After configuring the radar, set up the radar system on a clear night, turn on the device and send the laser beam image captured by the CCD camera to the PC side. The laser beam image contains the scattered echo signals of atmospheric molecules and aerosols. The beam spans multiple pixels, and then use matlab to convert the gray values corresponding to these pixels into the number of scattered light photons of the laser at this place.

[0031] In this embodiment, the image data within 2 km is obtained, and the lidar photon number is calculated. The specific process is as follows: According to formula (1), the aerosol lidar ratio in the lidar equation represents the aerosol backscattering coefficient, and according to formula (2), the molecular lidar ratio represents the backscattering coefficient formula of atmospheric molecules: (1) Among them, is the backscattering coefficient of the aerosol, is the detection distance, is the laser wavelength.

[0032] (2) Among them, is the backscattering coefficient of atmospheric molecules.

[0033] The molecular lidar ratio is 8π / 3 sr, and the aerosol lidar ratio is 50 sr. Substitute the above parameters into the lidar equation to obtain the lidar echo signal photon number: (3) Among them, is the number of atmospheric backscattering echo signal photons at the backscattering lidar receiving distance z, with a scattering angle of θ and an angular width of dθ; is the number of photons of the laser radar emitted light; is the lidar system constant; A is the effective optical receiving area; is the vertical distance from the CDD camera to the beam; 、 are the backscattering coefficients of the aerosol and atmospheric molecules at the distance z in the direction of the scattering angle θ, respectively; 、 are the extinction coefficients of the aerosol and atmospheric molecules at the distance z, respectively.

[0034] Step 1-2: The signal photons at each point on the laser beam are recorded in pixels at different positions, and a two-dimensional image is formed by combining multiple pixel points.

[0035] Step 2: Obtain a simulation image based on the two-dimensional image. Specifically: Expand the two-dimensional image along the Poisson distribution in height and the Gaussian distribution in width, and convert the height into the number of pixels to obtain the simulation image.

[0036] Step 3: Construct a dual residual neural network model, as Figure 5 shown. Specifically, it is a U²Net network structure model based on the dual residual attention mechanism to denoise the simulation image of the lidar signal; The U²Net network structure model based on the dual residual attention mechanism consists of two modules, namely the residual convolution module and the first convolution module; Among them, the residual convolution module is the RSUA-L module, which includes the first residual convolution module, the second residual convolution module, the third residual convolution module, the fourth residual convolution module, the fifth residual convolution module, and the sixth residual convolution module. The structures of the first residual module and the sixth residual module are the same, both including sampling with a depth of 7, that is, the first residual module and the sixth residual module are RSUA-7; The structures of the second residual module and the fifth residual module are the same, both including sampling with a depth of 6, that is, the second residual module and the fifth residual module are RSUA-6; The structures of the third residual module and the fourth residual module are the same, both including sampling with a depth of 5, that is, the third residual module and the fourth residual module are RSUA-5; Each RSUA-5, RSUA-6, and RSUA-7 contains a RAM attention module. In the RAM attention module, the first-layer residual is used, and the backbone of the RSUA-L module uses the second-layer residual. The combination of the dual residual network can better balance the detail retention and noise suppression in image denoising. It enhances the feature extraction ability through residual connections and improves the accuracy of denoising through the attention mechanism. Compared with traditional models, this structure can often show better denoising effects in complex noise environments. The first convolution module is the RSUA-4D module that expands the feature receptive field through the dilation coefficient.

[0037] Step 3-1: Construct a RAM attention module; The RAM attention module, as Figure 8 shown, is composed of 6 residual units, 1 channel attention layer, and 1 convolution layer connected in sequence. The channel attention mechanism can assign different weights to the output feature maps of the intermediate layers, automatically obtain the importance of each feature channel, then enhance the useful features according to the importance and suppress the features that are not very useful for the current task, and guide the network to continuously reduce the dimension of the feature maps.

[0038] Among them, the residual unit includes five layers connected in sequence: a convolutional layer, a batch normalization layer, an activation function layer, a convolutional layer, and a batch normalization layer. The input of the residual unit is added to the output after passing through the five layers, and then passes through an activation function layer to obtain the output of the residual unit.

[0039] Step 3-2: Construct the RSUA-L module; The RSUA-L module is specifically as follows. First, through the residual convolution stage, including a convolutional layer, an activation function layer, and 4 convolutional blocks for convolutional layer processing to extract image features. Among them, each convolutional block includes a convolutional layer Conv composed of 64 3×3 convolutional kernels, a normalization layer BN, and an activation function Leaky ReLU. Combining the residual learning method, the input of each convolutional block is connected to the output of the second convolutional block for addition operation to form a residual structure. The mapping function is expressed as: (4) Among them, represents the feature map of the input of the second convolutional block, Fres(s) represents the mapping function of the second convolutional block, ReLU represents the non-linear activation unit, and f(s) represents the result of the intermediate layer output.

[0040] The residuals between convolutional blocks are realized through simple skip connections, which will not add extra model parameters. It is equivalent to the network only learning the residual features between the input and the output, reducing the difficulty of network training. At the same time, the introduction of the residual module can better solve problems such as gradient dispersion, gradient explosion, and gradient degradation. This is the second layer of residuals in the double-residual model, which improves the network training speed and accelerates the convergence process.

[0041] Then enter the sampling stage, including L (L = 5, 6, 7) times of downsampling modules and L (L = 5, 6, 7) times of upsampling modules. Among them, the downsampling module includes 2×2 max pooling, two 3×3 convolutions, normalization BN, and an activation function Leaky ReLU; the upsampling module includes an upsampling layer, two 3×3 convolutions, normalization BN, and an activation function LeakyReLU.

[0042] The features after sampling are output after passing through 4 convolutional blocks and a convolutional layer in the residual convolution stage.

[0043] Specifically, the RSUA-5 module includes a convolutional layer, an activation function layer, 4 convolutional blocks, 5 downsampling modules, 5 upsampling modules, 4 convolutional blocks, and a convolutional layer connected in sequence; the RSUA-6 module includes a convolutional layer, an activation function layer, 4 convolutional blocks, 6 downsampling modules, 6 upsampling modules, 4 convolutional blocks, and a convolutional layer connected in sequence; the RSUA-7 module includes a convolutional layer, an activation function layer, 4 convolutional blocks, 7 downsampling modules, 7 upsampling modules, 4 convolutional blocks, and a convolutional layer connected in sequence; the difference in the RSUA-L module when L takes different values lies in the different number of sampling times.

[0044] This module forms a left-right symmetric U-shaped structure. The residual convolution stage before downsampling and the L downsampling modules are the left branch of the U-shaped structure, and the upsampling and the residual convolution stage after upsampling are the right branch of the U-shaped structure. The symmetric structure of the left and right branches uses an Attention Gate attention gate for skip connection, and the Attention Gate attention gate is the constructed RAM attention module. Specifically: the output of the first convolutional block in the left branch is fused to the input of the fourth convolutional block in the right branch through the RAM attention module, the output of the second convolutional block in the left branch is fused to the input of the third convolutional block in the right branch through the RAM attention module, the output of the third convolutional block in the left branch is fused to the input of the second convolutional block in the right branch through the RAM attention module, the output of the fourth convolutional block in the left branch is fused to the input of the first convolutional block in the right branch through the RAM attention module, the output of the first layer of downsampling in the left branch is fused to the input of the L-th upsampling in the right branch through the RAM attention module, the output of the second layer of downsampling in the left branch is fused to the input of the (L-1)-th upsampling in the right branch through the RAM attention module, and so on. The result of the previous layer in the upsampling stage and the corresponding depth downsampling result are processed through attention, and then the processed result is concatenated with the result of the previous layer. In this way, the low-level and high-level feature information of the feature map is effectively fused, and at the same time, the features of irrelevant parts are ignored, and the key feature information is retained.

[0045] Step 3-3: Construct the RSUA-4D module; The RSUA-4D module, as Figure 7As shown, it consists of a sequentially connected convolutional layer, activation function layer, 8 convolutional blocks, and a convolutional layer. Each convolutional block consists of a convolutional layer, activation function, and batch normalization layer. Among them, the first convolutional block and the eighth convolutional block are connected by a RAM attention module, the second convolutional block and the seventh convolutional block are connected by a RAM attention module, the third convolutional block and the sixth convolutional block are connected by a RAM attention module, and the fourth convolutional block and the fifth convolutional block are connected by a RAM attention module. The convolutional layer in the convolutional block sets dilation coefficients of Dilation Rate = 1, 2, and 4. In this module, due to multiple downsamplings, the resolution of the image presents a relatively low level. Continuing downsampling will cause the loss of some feature information. Therefore, dilation coefficients are set in the convolutional layer to expand the receptive field, so as to achieve the purpose of preserving context information as much as possible.

[0046] In the RSUA-4D module, the dilation coefficient Dilation Rate "slides" the structural element over the image and checks the overlapping part of the structural element and the image at each position. If the structural element overlaps with the pixels in the image, the dilation operation will expand the central pixel to this area, which can make the foreground part in the image larger and help fill holes or eliminate small noises.

[0047] Step 3-4: Use the RSUA-L module and the RSUA-4D module to construct the dual residual neural network model U²Net network. The dual residual neural network model U²Net network consists of 2 RUSA-5 modules, 2 RUSA-6 modules, 2 RUSA-7 modules, 1 RUSA-4D module, and 3 RAM attention modules; the simulation image first passes through RSUA-7 to obtain the first layer of feature maps, the first layer of feature maps are input into RSUA-6 to obtain the second layer of feature maps, the second layer of feature maps are input into RSUA-5 to obtain the third layer of feature maps, the third layer of feature maps are input into RSUA-4D to obtain the third layer of feature maps with an expanded receptive field. The third layer of feature maps with an expanded receptive field and the third layer of feature maps passing through the RAM attention module are fused and then input into RSUA-5 to obtain the fourth layer of feature maps. The fourth layer of feature maps and the second layer of feature maps passing through the RAM attention module are fused and then input into RSUA-6 to obtain the fifth layer of feature maps. The fifth layer of feature maps and the second layer of feature maps passing through the RAM attention module are fused and then input into RSUA-7 to obtain the denoised image.

[0048] The overall model forms a left - right symmetric U - shaped structure. It sequentially passes through the RUSA - 7 module, RUSA - 6 module, and RUSA - 5 module as the encoding stage. Through gradually deepening downsampling, the receptive field is gradually expanded to extract high - level semantic features at low resolution, such as the global distribution pattern of aerosols. Then it sequentially passes through the RUSA - 5 module, RUSA - 6 module, and RUSA - 7 module as the decoding stage. Through symmetric upsampling, the spatial resolution is restored, and at the same time, shallow - layer detail features (such as local noise patterns) are fused with deep - layer semantic features to achieve multi - scale information complementarity. Deep networks (such as RSUA - 7) cover a larger range of the receptive field (such as the entire lidar image) through multiple downsamplings, and can capture the vertical distribution trend of aerosols or large - scale noise patterns. At heights z > 1 km, deep features can identify the overall extinction coefficient change of the aerosol layer, while shallow features (RSUA - 5) pay more attention to local noise near the ground.

[0049] The attention module can effectively enhance the learning of key information by focusing on important parts of the input image and suppressing unimportant parts. Through the skip connection of the Attention Gate attention module, low - level features are combined with high - level features to ensure that both low - level details and high - level semantic information can be effectively transmitted during the denoising process. Using a double - residual structure can improve the information flow, alleviate the vanishing gradient, restore more details, and remove noise.

[0050] Step 4: Use a double - residual neural network model (U²Net network) to denoise the simulated image of the lidar signal to obtain the denoised image.

[0051] Preferably, in this embodiment, the double - residual neural network model is also trained. To improve the training effect of the model, Gaussian white noise with a mean of 0 and a variance of 0.03 is added to the simulated image to obtain a noisy image, and the constructed model is trained using the noisy image. The simulated image is as Figure 4 shown, where Figure 4 (a) is the original image, Figure 4 (b) is the noisy image.

[0052] The addition of Gaussian white noise will cause the image quality to decline, such as more noise points and the image becoming blurred. Denoising the image after adding noise will have a more obvious effect. A mean of 0 is to ensure that the noise does not cause a systematic shift in brightness; a variance of 0.03 is a value selected through experiments, simulating a noise of moderate intensity. The advantage of generating a noisy image to train the model compared to directly using the simulated image is that it can enhance the robustness of the model, improve its generalization ability, and better simulate and adapt to common noises and interferences in the real world. It provides more training data for the practical application of the model and can effectively improve the actual performance of the image - processing algorithm, especially in real environments with noise pollution and poor image quality.

[0053] For convenient observation, the noisy image and the denoised image of the result are stretched into three-dimensional mesh graphs.

[0054] In this embodiment, a simulation experiment is carried out on the proposed denoising method. The U²Net network proposed by the present invention is used to compare the denoising effects with Wiener filtering, VGG16, and DnCNN, and PSNR and SSIM are used for comparison.

[0055] Use Wiener filtering, VGG16, DnCNN, and the U²Net of the present invention to process the noise model with a mean of 0 and a variance of 0.03 respectively. The images after denoising by various methods are compared as Figure 9 shown.

[0056] In order to detect the denoising effect, the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) are used to evaluate the denoising effects corresponding to different methods; Table 1 Comparison of Image PSNR

[0057] Table 2 Comparison of Image SSIM

[0058] Compare the number of photons after denoising with the true number of photons, as Figure 10 shown.

[0059] Furthermore, in this embodiment, an ablation experiment is carried out to verify the effectiveness of each module: Ablation experiment: In order to verify the effectiveness and generality of the double-residual structure adopted in the RAM module and the RSUA-L module proposed by the present invention, the modules proposed in this application are added to the DnCNN network for comparative experiments; Table 6 Comparison of Ablation Experiment Results

[0060] The PSNR of the present invention is relatively high, which means that the U model can retain more image details during the denoising process, reduce the influence of noise, and the quality of the denoised image is close to the original image. Compared with other models, it can perform better in noise suppression and detail preservation. By fusing the RAM and the residual convolution module, the trade-off between noise and details can be effectively avoided, the quality of the denoised image is improved, and thus a higher PSNR value is obtained.

[0061] The SSIM of the present invention is relatively high, indicating that while denoising, it can better preserve the structural information of the original image, avoiding obvious distortion or structural damage to the image. Compared with other methods, the model proposed in the present invention has obvious advantages in maintaining the structure and details of the image. The model proposed in the present invention improves the overall structural preservation of the image by integrating the RAM and residual convolution modules. Especially during the denoising process, it can better protect the texture and structural information of the image and maintain a higher SSIM value.

[0062] Brisque is a no-reference image quality assessment metric that mainly evaluates image quality through the natural scene statistical characteristics of the image. The Brisque value of the U model is relatively low, indicating that the image quality after denoising is good, and there will be no obvious distortion or artifacts when viewed by the human eye. A low Brisque value indicates that the denoising effect is natural and the visual perception is relatively clear.

[0063] The RAM and residual convolution modules of the present invention give full play to the advantages of both. The RAM attention mechanism module helps the model more effectively focus on processing the noise area, while the residual structure of the residual convolution module enhances the model's ability to retain details. The combination of the two enables the denoising model to reduce structural loss and visual degradation while maintaining a high-quality denoising effect.

[0064] In this embodiment, the laser is part of the entire CCD imaging lidar system. The laser emits a laser beam, and the CCD captures an image of the laser beam. The laser is a continuous laser with a wavelength of 532 nm and an output power of approximately 3 W. The CCD cameras are all MVCH089-10GM of Hikrobot, 8.9 million pixel Gigabit Ethernet area array cameras. The CCD camera has 4096 pixels in the horizontal direction and 2160 pixels in the vertical direction; the detector lens is Nikon70-300.

[0065] In this embodiment, the data processing process is carried out on Matlab.

[0066] Embodiment 2 In one or more embodiments, a lidar signal denoising system based on a dual-residual neural network is disclosed, specifically including: An acquisition unit configured to acquire lidar signals; A preprocessing unit configured to preprocess lidar signals to obtain a simulation image; A denoising unit configured to denoise the simulation image using a dual-residual neural network model; Specifically, the simulation image is input into the double residual neural network model, and the first layer of feature map is obtained through the first residual convolution module. The first layer of feature map is input into the second residual convolution module to obtain the second layer of feature map. The second layer of feature map is input into the third residual convolution module to obtain the third layer of feature map. The third layer of feature map is input into the first convolution module to obtain the third layer of feature map with an enlarged receptive field. After the third layer of feature map with the enlarged receptive field is fused with the third layer of feature map passing through the first attention module, it is input into the fourth residual convolution module to obtain the fourth layer of feature map. After the fourth layer of feature map is fused with the second layer of feature map passing through the first attention module, it is input into the fifth residual convolution module to obtain the fifth layer of feature map. After the fifth layer of feature map is fused with the second layer of feature map passing through the first attention module, it is input into the sixth residual convolution module to obtain the denoised image.

[0067] Embodiment III This embodiment provides an electronic device, including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps of the above-mentioned lidar signal denoising method based on the double residual neural network are completed.

[0068] Embodiment IV This embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by the processor, the steps of the above-mentioned lidar signal denoising method based on the double residual neural network are completed.

[0069] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0070] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device realizes the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0071] These computer program instructions can also be loaded onto a computer or other programmable data processing device, where they execute a series of operational steps to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one process or multiple processes and / or blocks. Figure 1 One process or multiple processes and / or blocks Figure 1 Steps for implementing the functions specified in one block or multiple blocks.

[0072] In the above embodiments, the descriptions of each embodiment have their own focuses. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0073] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A laser radar signal denoising method based on double residual neural network, characterized in that: include: Collecting LiDAR signals; Preprocessing the laser radar signal to obtain a simulated image; The double residual neural network model is used to denoise the simulation image; Specifically, the simulated image is input into the dual residual neural network model, and the first layer feature map is obtained by passing through the first residual convolution module, the first layer feature map is input into the second residual convolution module to obtain the second layer feature map, the second layer feature map is input into the third residual convolution module to obtain the third layer feature map, the third layer feature map is input into the first convolution module to obtain the third layer feature map with an enlarged receptive field, the third layer feature map with an enlarged receptive field is fused with the third layer feature map passing through the first attention module and then input into the fourth residual convolution module to obtain the fourth layer feature map, the fourth layer feature map is fused with the second layer feature map passing through the first attention module and then input into the fifth residual convolution module to obtain the fifth layer feature map, the fifth layer feature map is fused with the second layer feature map passing through the first attention module and then input into the sixth residual convolution module to obtain a denoised image.

2. The laser radar signal denoising method based on a double residual neural network according to claim 1, characterized in that: The training process of the dual residual neural network model includes: Preprocess the historical lidar signal to obtain multiple simulated images, add Gaussian white noise to each simulated image to obtain multiple noisy images to form a training set; The double residual neural network model is trained using the training set.

3. The laser radar signal denoising method based on double residual neural network according to claim 1, characterized in that: The first attention module includes a residual unit, a channel attention layer and a convolutional layer; wherein the residual unit includes a five-layer structure of a convolutional layer, a batch normalization layer, an activation function layer, a convolutional layer, a batch normalization layer and an activation function layer connected in sequence, and the input of the residual unit is added to the output after passing through the five-layer structure and then passed through an activation function layer to obtain the output of the residual unit.

4. The laser radar signal denoising method based on a double residual neural network according to claim 1, characterized in that: The first residual convolution module includes a residual convolution stage and a sampling stage, specifically: First, it passes through several convolution blocks in the residual convolution stage, then goes through downsampling and upsampling in the sampling stage, and finally passes through several convolution blocks in the residual convolution stage. The downsampling and several convolution blocks before downsampling constitute the encoder, and the upsampling and several convolution blocks after upsampling constitute the decoder. The symmetric layer between the encoder and the decoder is connected by the first attention module.

5. The laser radar signal denoising method based on double residual neural network according to claim 4, characterized in that: The convolution block consists of a convolution layer, an activation function and a batch normalization layer. The input of each convolution block and the output of the convolution block are added to form a residual structure as the input of the next layer.

6. The laser radar signal denoising method based on double residual neural network according to claim 1, characterized in that: The residual convolution stages of the first residual convolution module, the second residual convolution module, the third residual convolution module, the fourth residual convolution module, the fifth residual convolution module and the sixth residual convolution module have the same structure; the number of samplings in the sampling stage of the second residual convolution module is smaller than that of the first residual convolution module, and the fifth residual convolution module and the second residual convolution module have the same structure; the number of samplings in the sampling stage of the third residual convolution module is smaller than that of the second residual convolution module, and the fifth residual convolution module and the second residual convolution module have the same structure; the sixth residual convolution module and the first residual convolution module have the same structure.

7. The laser radar signal denoising method based on double residual neural network according to claim 1, characterized in that: The first convolution module consists of a convolution layer, an activation function layer, 8 convolution blocks and a convolution layer connected in sequence; among them, the first convolution block and the eighth convolution block are connected by the first attention module, the second convolution block and the seventh convolution block are connected by the first attention module, the third convolution block and the sixth convolution block are connected by the first attention module, and the fourth convolution block and the fifth convolution block are connected by the first attention module.

8. A laser radar signal denoising system based on a double residual neural network, characterized in that: include: A collection unit, which is configured to: collect laser radar signals; A preprocessing unit, which is configured to: preprocess the laser radar signal to obtain a simulated image; A denoising unit is configured to: denoise the simulation image using a double residual neural network model; Specifically, the simulated image is input into the dual residual neural network model, and the first layer feature map is obtained by passing through the first residual convolution module, the first layer feature map is input into the second residual convolution module to obtain the second layer feature map, the second layer feature map is input into the third residual convolution module to obtain the third layer feature map, the third layer feature map is input into the first convolution module to obtain the third layer feature map with an enlarged receptive field, the third layer feature map with an enlarged receptive field is fused with the third layer feature map passing through the first attention module and then input into the fourth residual convolution module to obtain the fourth layer feature map, the fourth layer feature map is fused with the second layer feature map passing through the first attention module and then input into the fifth residual convolution module to obtain the fifth layer feature map, the fifth layer feature map is fused with the second layer feature map passing through the first attention module and then input into the sixth residual convolution module to obtain a denoised image.

9. An electronic device, characterized in that: It includes a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the laser radar signal denoising method based on a double residual neural network as described in any one of claims 1 to 7 is completed.

10. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the laser radar signal denoising method based on a double residual neural network as described in any one of claims 1 to 7.

Images