SAGAN-based streak tube laser radar stripe echo signal enhancement and restoration method

Abstract

The invention discloses an SAGAN-based streak tube laser radar fringe echo signal enhancement and restoration method. The method comprises the following steps: adding a noise signal simulation underwater signal fringe pattern into an original ground signal fringe pattern; inputting the obtained simulated underwater signal fringe pattern as a noise signal into a generator network, and inputting an output result of a generator and a ground fringe pattern into a discriminator network; training the generative adversarial network by adopting a mixed precision floating-point number method until the discriminator network cannot distinguish the output result of the generator and the original ground fringe pattern; and obtaining parameters of the trained generator network, and inputting the original underwater signal fringe pattern into the trained generator network to obtain a restored and enhanced underwater signal fringe pattern. According to the method, echo signals of an underwater target are restored and enhanced by introducing a generative adversarial network (SAGAN) with a self-attention mechanism.

Description

Technical Field

[0001] The present invention relates to a method for enhancing and restoring the fringe echo signal of a streak tube lidar, and more particularly to a method for enhancing and restoring the fringe echo signal of a streak tube lidar based on a generative adversarial network (SAGAN). Background Art

[0002] Compared with traditional detection means such as sound and magnetism, airborne lidar is a typical cross-media detection method, which has incomparable advantages in terms of mobility, safety, adaptability, etc. The advantages of using airborne lidar for shallow water small target detection are mainly reflected in the following aspects: First, excellent mobility. With the advantages of integrated integration with the flight platform and high cruising speed, it can quickly reach and cover the operation sea area; second, superior safety. It is not affected by factors such as shallow water reefs and underwater obstacles, and can carry out detection operations day and night, with strong anti-interference ability; third, high operation efficiency. Thanks to the technical characteristics of cross-media detection, it has a fast detection speed and a high proportion of effective action time (characteristics such as no deployment and recovery of sensors and flexible turning), and can quickly search and detect a large area of water.

[0003] Compared with traditional microwave radars, lidars have the characteristics of high precision, high resolution, high detection sensitivity, high confidentiality, small size, and light weight, which are convenient for airborne and shipborne use. In addition, due to different working mechanisms, the laser pulses emitted by lidars have stronger anti-interference ability and richer data information than traditional microwave radar signals, resulting in higher detection and recognition capabilities. The array lidar based on a streak tube has a high imaging frame rate, full waveform sampling, high detection sensitivity and range resolution, and has broad prospects in airborne radar mapping and military applications.

[0004] The airborne lidar system has obvious advantages in target detection, but for shallow water targets, it also faces a series of technical difficulties, such as insufficient detection resolution for small-size targets, a significant decrease in the measured depth due to the influence of water bodies on the echo signal, and the need to further improve the single flight coverage.

[0005] Currently, the detection of underwater targets by lidar is usually based on platforms carried on helicopters or small unmanned aerial vehicles. For example, the AN / AES-1 Airborne Laser Mine Detection System (ALMDS) in the United States, the SHOALS system (Scanning Hydrographic Operational Airborne Lidar Survey System) in Canada, and the CZMIL (Coastal Zone Mapping and Imaging Lidar) system designed to solve the measurement problems in turbid waters and shallow water areas. These lidar systems for underwater target detection are carried on helicopter platforms, with a flight altitude usually only 400 - 500m and a detection width usually less than 300m.

[0006] Due to the unique working mechanism of the Airborne Streak Tube Imaging Lidar (ASTIL), the echo signals it collects can be represented as two-dimensional single-channel digital images. Compared with the signals collected by traditional single-point detection LiDAR, the echo signals collected by ASTIL can contain the surface cross-sectional morphology within a detection area of hundreds of meters and the echo intensity information of each target. Moreover, the flight altitude of the radar-carrying platform is usually several kilometers above the ground. Its echo signals are rich in semantic information and have the potential to directly identify ground objects and regions based on single echo and single data source. A single flight can cover a very large range of targets.

[0007] Researchers expect to be able to use the powerful detection ability of ASTIL to simultaneously achieve the function of underwater mapping. However, when ASTIL detects underwater targets, due to the inevitable multi-faceted influence of water on the propagation of laser, the echo signal stripe patterns received by ASTIL from water areas in actual applications are usually relatively weak.

[0008] Figure 1 Typical stripe patterns for underwater and ground targets. When ASTIL detects underwater targets, due to the multiple reflections, refractions of laser in water, as well as the absorption and scattering of laser by water and impurities in water and other multi-faceted factors, the signals received by ASTIL have problems such as quality degradation and intensity attenuation. The obtained echo signal stripe patterns usually have weak signal intensity and low signal-to-noise ratio, which will make it difficult to carry out subsequent analysis and processing steps such as image processing and point cloud recognition.

[0009] Currently, in the field of image enhancement, it can be mainly divided into methods based on traditional feature calculation, methods based on computer graphics, and methods based on neural networks.

[0010] Traditional feature - based calculation methods usually require estimating the distance between the underwater target and ASTIL and the water depth of the water body to calculate the effects of reflection, refraction, scattering, etc. of the laser in the water body. However, it is usually difficult to meet the above - mentioned conditions in practical applications, so such methods are generally not adopted in practical applications.

[0011] Graphics - based image enhancement methods mainly rely on traditional image - processing techniques such as mathematical transformation, morphological operations, and filter design. Its core is to achieve contrast adjustment, noise suppression, and detail enhancement through spatial - domain or frequency - domain analysis of the image, and perform spatial - domain and frequency - domain changes on the image for image enhancement. Graphics - based methods have the advantages of strong algorithm - process interpretability and high computational efficiency. However, the ability of graphics - based methods to handle complex scenes is limited, it is easy to lose the detailed content in the image, and the processing effect usually depends on parameter selection, requiring manual intervention to achieve the expected effect, and the selection of parameters usually has a greater impact on the final processing effect.

[0012] Neural - network - based methods mainly include two methods: diffusion models and generative adversarial networks (GANs). Methods based on this type are driven by a large amount of data, have strong global optimization capabilities, and can handle more complex degradation patterns and generate high - quality results.

[0013] The core idea of the diffusion model is to achieve data generation through two main processes: the forward diffusion process and the reverse diffusion process. The forward diffusion process is a parameterized Markov chain that gradually adds noise to the original data until the data finally becomes pure noise. The reverse diffusion process is the inverse process of the forward process. It starts from pure noise and gradually removes the noise through the learning of the neural network to restore the original data. As the reverse diffusion process progresses, the noise is gradually removed, the features of the data are gradually restored, and finally a generated sample close to the original data is obtained.

[0014] GAN consists of two parts: a generator and a discriminator. The generator is responsible for generating as realistic samples as possible, and the discriminator is responsible for distinguishing between the generated samples and the real samples. The goal of the generator is to convert low - quality lidar signals into high - quality signals, while the goal of the discriminator is to determine whether the signal comes from real data or is generated by the generator. Through continuous adversarial training, the generator can learn how to generate high - quality data closer to real signals, thereby achieving signal enhancement.

[0015] Compared with the diffusion model, the generated samples of GAN have higher detail richness and better sample diversity. Secondly, during the training process, the training goal of GAN is simpler and more direct, which also makes the training efficiency of GAN higher than that of the diffusion model. In addition, the GAN model is more flexible and can be flexibly combined with other models, facilitating subsequent improvement of the model according to the actual application requirements.

[0016] However, the ordinary GAN's method for image enhancement has a weak ability to distinguish between noise and edge details. The generated images may over-suppress noise or lose detail information, resulting in untrue or inaccurate generated images, which are difficult to meet the actual requirements of signal enhancement. At the same time, GAN is unstable during the training process and is prone to situations such as gradient vanishing and gradient explosion, and the network needs to be improved to prevent the waste of computing power caused by gradient explosion and vanishing during the training process.

[0017] During the detection process based on airborne streak tube lidar, there is a large degree of loss and attenuation of the echo signal of underwater targets. During the process of point cloud restoration, the point cloud density of the water area part is significantly reduced compared with that of the ground part in the point cloud map. Summary of the Invention

[0018] In view of the above problems, the present invention provides a method for enhancing and restoring the streak echo signal of a streak tube lidar based on SAGAN. This method restores and enhances the echo signal of underwater targets by introducing an adversarial generative network with self-attention mechanism (SAGAN).

[0019] The object of the present invention is achieved by the following technical solutions:

[0020] A method for enhancing and restoring the streak echo signal of a streak tube lidar based on SAGAN, comprising the following steps:

[0021] Step 1: Save the to-be-processed ground signal streak map and the original underwater signal streak map to relevant folders respectively;

[0022] Step 2: Add noise signals to the original ground signal streak map in Step 1 to simulate the underwater signal streak map. The specific steps are as follows:

[0023] Step 2-1: Add reverberation noise and random Gaussian noise to the original ground signal streak map;

[0024] Step 2-2: Add random attenuation to the original ground signal streak map;

[0025] Step 3: Improve the original GAN. The generator network uses a U-Net architecture, and a residual module is introduced into the generator to construct an adversarial generative network with self-attention mechanism;

[0026] Step 4: Scale down the size of the simulated underwater signal streak map obtained in Step 2 from 512×1024 to 256×512 proportionally, input it as a noise signal into the generator network, and input the output result of the generator and the ground streak map in Step 1 into the discriminator network;

[0027] Step 5: Train the adversarial generative network using the method of mixed-precision floating-point numbers until the discriminator network can no longer distinguish the output result of the generator from the original ground fringe pattern, where: the method of single-precision floating-point numbers is used during the normalization layer and the process of saving the model, and the method of half-precision floating-point numbers is used in other layers and during the training process;

[0028] Step 6: Obtain the parameters of the trained generator network in Step 5, and input the original underwater signal fringe pattern into the trained generator network to obtain the restored and enhanced underwater signal fringe pattern.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] 1. The present invention enhances and restores the waveform-sampled lidar echo signal based on SAGAN. Through the organic combination of structural innovation and training strategies, it demonstrates excellent detail recovery ability and robustness in fringe pattern image enhancement, and solves the problem that the fringe pattern of the echo signal received by the waveform-sampled lidar from water areas is usually weak in signal and low in signal-to-noise ratio.

[0031] 2. During the training process, the method of reducing the image size to one-fourth of the original and using the method of mixed-precision floating-point numbers is adopted, which greatly improves the training speed of the model on the premise of maintaining the computing accuracy of the model. Description of the Drawings

[0032] Figure 1 is a typical fringe pattern. The left figure is a typical fringe pattern of an underwater target, and the right figure is a typical fringe pattern of a ground target;

[0033] Figure 2 is a schematic diagram of the overall process of the fringe tube lidar fringe echo signal enhancement and restoration method based on SAGAN;

[0034] Figure 3 is the preprocessing effect diagram of the original fringe pattern. The left figure is a typical original fringe pattern, and the right figure is the fringe pattern after preprocessing;

[0035] Figure 4 is a schematic diagram of the generator network architecture;

[0036] Figure 5 is a diagram of the residual module architecture;

[0037] Figure 6 is a diagram of the self-attention module architecture;

[0038] Figure 7 is a schematic diagram of the discriminator network architecture;

[0039] Figure 8 is the original underwater target fringe pattern;

[0040] Figure 9 It is the fringed pattern after image restoration and enhancement;

[0041] Figure 10 It is the satellite image of the detection target;

[0042] Figure 11 It is the original signal point cloud map;

[0043] Figure 12 It is the point cloud map after signal enhancement and restoration;

[0044] Figure 13 It is the underwater target point cloud map. The left figure is the original point cloud map, and the right figure is the point cloud map after enhancement and restoration;

[0045] Figure 14 It is the original signal point cloud density distribution map;

[0046] Figure 15 It is the point cloud density distribution map after signal restoration and enhancement. Specific implementation manners

[0047] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention, without departing from the spirit and scope of the technical solutions of the present invention, shall be covered by the protection scope of the present invention.

[0048] The present invention provides a method for enhancing and restoring the fringe echo signal of a streak tube lidar based on SAGAN. The method uses SAGAN to restore and enhance the echo signal of the airborne streak tube lidar for low-quality target signals. The overall process is as Figure 2 shown. SAGAN is similar to the original GAN and consists of two parts: a generator network G and a discriminator network D.

[0049] During the training process, the input of the generator network G is the signal containing noise, and the output is the false clear signal generated by the generator network G; the main function of the discriminator network D is to discriminate whether the signal input into it is a real clear signal or a false signal generated by the generator network. The ultimate training goal of the network is to make the discriminator network D unable to discriminate whether the signal input into it is a real clear signal or a false signal generated by the generator network G. That is, the generator network G can generate a real restored target signal from the signal containing noise. The generator network G obtained after training can be used for the restoration and enhancement of low-quality target signals. Taking the original target fringe signal as the input of the generator network G, after the calculation of the generator network G, the output result is the result obtained after the enhancement and restoration of the low-quality target signal image.

[0050] I. Preprocessing of training data

[0051] Before network training, it is necessary to preprocess the training data and simulate the characteristics of low-quality signals with clear ground target signals. This requires adding noise to the clear ground target signals first to simulate the echo signals of low-quality targets, which are used as the input of the generator during the training process. The original size of the fringe pattern of the echo signal of the streak tube lidar is 512×1024.

[0052] It can be seen from the literature that laser in water is mainly affected by the following two factors. On the one hand, multiple reflections and refractions occur at the dielectric interface when the laser enters and leaves the water surface. On the other hand, the absorption and scattering of the water body occur during the propagation of the laser in the water body.

[0053] Considering the effects of multiple reflections, refractions at the dielectric interface and the scattering of the water body, reverberation noise and random Gaussian noise are added to the clear ground target signals; to simulate the absorption of the laser by the water body, random attenuation is added to the signals. The comparison between the original ground target signal and the ground target signal after adding noise is as Figure 3 shown.

[0054] At the same time, since the echo signal of the streak tube lidar, that is, the fringe pattern structure, is usually relatively simple, to reduce the computational load of subsequent model training, the fringe pattern with an original size of 512×1024 is scaled down proportionally to an image with a size of 256×512. This processing will introduce an error of about one pixel distance on the original fringe pattern, that is, the point position in the subsequent generated point cloud map will have an error of less than 0.3m, which is an acceptable error range. At the same time, the total number of parameters of the generator network is reduced from about 160M to about 41.9M, and the total number of parameters of the discriminator network is reduced from about 8.5M to about 1.93M, that is, the network training parameters are reduced to about one-fourth of the original network. This operation can greatly reduce the computational load and computational time required for model training, and at the same time, the error introduced by this optimization operation is also controlled within a reasonable range.

[0055] II. SAGAN Model Network Architecture

[0056] The overall architecture of the network is mainly composed of a generator network G and a discriminator network D. Among them, the main role of the generator network G is to enhance and restore the noise signal; the main role of the discriminator network D is to evaluate the similarity between the signal fringe pattern generated by the generator network G and the fringe pattern of the original clear signal, and to guide the training process of the generator network G accordingly.

[0057] To improve the training stability of the network, SAGAN made the following modifications based on the original adversarial generative network: First, self-attention modules were introduced into the discriminator network and the generator network. The effective integration of this module enables the network to truly capture and correlate long-range spatial information, allowing the model to simultaneously capture the global semantic information in the image and enhancing the model's ability to recognize details and pattern structures in the image. Second, spectral normalization was used in the discriminator network and the generator network. This method can effectively constrain the parameter update speed of each convolutional layer during training, prevent gradient explosion or gradient disappearance during training, and improve the convergence of model training. At the same time, the TTUR method was also used, that is, different learning rates were adopted for the generator and the discriminator, which can balance the training speeds of the generator network and the discriminator network, improving the stability of model training.

[0058] To improve the training speed of the network, on the basis of reducing the image size, the method of mixed-precision floating-point numbers was adopted during training to improve the training efficiency by reasonably using numerical formats with different precisions. Since the half-precision floating-point number (FP16) reduces the memory occupancy by 50% compared with the single-precision floating-point number (FP32) and increases the calculation speed by 2 - 8 times, FP32 is used during the normalization layer and the model saving process, and FP16 is used in other layers and during training. Under this strategy, while maintaining the model performance, the training speed is significantly improved (expected to be accelerated by 30 - 50%), and the video memory consumption is reduced by about 30% at the same time. At the same time, compared with directly using 16-bit floating-point numbers, mixed-precision training does not affect the storage format of model parameters, and the saved or loaded model files are fully compatible with the original FP32 version.

[0059] 1. Generator Network Architecture

[0060] The architecture of the generator network G is as Figure 4 shown. Its main architecture is a U-Net model, and at the same time, residual blocks (Residual Block) and self-attention blocks (SelfAttention Block) are introduced. The input of the generator network G is a signal containing noise, and the output is a fake clear signal. The training objective of the generator network G is to generate a clear stripe signal from the stripe signal containing noise. After the generator network G is trained, when a real underwater target stripe signal is input, a clear underwater target stripe signal can be output.

[0061] In the generator network G, all normalization layers adopt the Instance Normalization Layer. The Instance Normalization Layer normalizes each channel of each sample during the calculation process. The specific calculation method of the Instance Normalization Layer is as follows:

[0062] (1)

[0063] Among them, is the input tensor at the th sample, the th channel, the th row, and the th column; is the mean of the th sample and the th channel; is the variance of the th sample and the th channel; is the normalization scaling factor; is the normalization offset; and are learnable parameters; is a small constant to prevent division by zero.

[0064] Compared with the commonly used Batch Normalization Layer and Layer Normalization Layer, which normalize each channel of each small batch and all channels of each sample respectively, the Instance Normalization Layer can retain the differences between samples and can effectively improve the model performance and stability in image generation tasks.

[0065] The specific architecture of the residual module in the network architecture is as Figure 5 shown. The residual module consists of two convolutional layers, two normalization layers, and one ReLU layer. The output size of the residual module is the same as the input size. The output of the residual module is composed of the original input of the module and the calculation results of each layer of the network added together. Therefore, introducing the residual module can effectively alleviate the problems of gradient disappearance and deep network degradation, and at the same time can retain the features of the previous layer, improving the performance and generalization ability of the network.

[0066] The specific architecture of the self-attention module in the network architecture is as Figure 6As shown in the figure, the introduction of the self-attention module can enhance the model's ability to capture long-range dependencies and the model's parallelism, thereby improving the model's expressive ability and highlighting important features in the image. The main part of the self-attention module adopted here is similar to the original self-attention module, but a step similar to the residual module is added to the output part of the module, directly adding the original input to the output result of the self-attention module, further enhancing the model's resistance to gradient disappearance and deep network degradation, and improving the performance and generalization ability of the network.

[0067] 2. Discriminator Network Architecture

[0068] The architecture of the discriminator network D is as Figure 7 shown. Its main architecture is a convolutional neural network model, in which a self-attention module is added, and spectral normalization is performed on the convolutional layer. The discriminator network D is used to judge whether the input picture is a real clear signal or a signal generated by the generator network G. After the training is completed, the judgment accuracy of the discriminator network D should fluctuate around 50%, indicating that at this time, the discriminator network D can no longer distinguish whether the input data is generated by the generator network G.

[0069] The self-attention module in the discriminator network D is the same as that in the generator network G, and will not be elaborated here.

[0070] LeakyReLU (Leaky Rectified Linear Unit) is adopted in the discriminator network D. This is an improved ReLU activation function, which is widely used in deep learning and performs excellently especially in generative adversarial networks (GANs). Its functional form is as follows:

[0071]

[0072] where is the negative half-axis slope, taking 0.2. Compared with the ordinary ReLU layer, it completely discards the gradient in the negative interval. LeakyReLU retains the small gradient in the negative interval, maintains the parameter update ability, can prevent the discriminator from converging prematurely and balance the training dynamics of the generator and the discriminator, and significantly improves the training stability of the deep model.

[0073] The goal of spectral normalization is to make the spectral norm (the largest singular value) of the weight matrix equal to 1. By iteratively calculating the left and right singular vectors of the weight matrix to approximate the singular vectors corresponding to the largest singular value, the spectral normalization of the weight matrix is achieved. After spectral normalization, the numerical update range of the weight matrix can be constrained to prevent gradient explosion.

[0074] III. According to the design scheme, the specific implementation steps are as follows:

[0075] Step 1: Save the to-be-processed ground signal fringe pattern and the original underwater signal fringe pattern to relevant folders respectively;

[0076] Step 2: Add the designed noise signal to the original ground signal fringe pattern in Step 1 to simulate the underwater signal fringe pattern;

[0077] Step 3: Input the simulated underwater signal fringe pattern obtained in Step 2 as the noise signal into the generator network, and input the output result of the generator and the ground fringe pattern in Step 1 into the discriminator network;

[0078] Step 4: Train the network as shown in Step 3 until the discriminator network cannot distinguish the output result of the generator from the original ground fringe pattern;

[0079] Step 5: Obtain the parameters of the trained generator network in Step 4, and input the original underwater signal fringe pattern into the trained generator network to obtain the restored and enhanced underwater signal fringe pattern.

[0080] IV. Technical effects:

[0081] Compare the training effects of the original GAN and the improved SAGAN. After both networks are trained 300 times, the evaluation results of the restored image compared with the original image are shown in Table 1. Compared with GAN, after being trained by SAGAN, the peak signal-to-noise ratio (PSNR) of the result increases by 3.857, the structural similarity index (SSIM) increases by 0.09, and the mean square error (MSE) is reduced to 30.5% of the GAN result. It shows that SAGAN has greatly improved the ability of image enhancement and restoration compared with the original GAN, and improved the quality of the restored image.

[0082] Table 1 Comparison of training results

[0083]

[0084] Taking Figure 8 the single underwater target signal fringe pattern shown as an example, after using the trained generator model G to perform image enhancement and restoration on the original signal fringe pattern, the obtained result is as Figure 9 shown. The signal-to-noise ratio of the fringe image has a large improvement.

[0085] The real satellite image of the detected target area is as Figure 10 shown. After the original fringe pattern detected and the fringe pattern obtained after being processed by the model are respectively drawn into point cloud maps, the obtained results are as Figure 11 and Figure 12 shown. As Figure 13As shown in the figure, the point cloud map of the water area is intercepted. After statistics, the echo stripe map of the original underwater target generates a total of 51,894 points, while after being processed by the model, the underwater target stripe map generates a total of 94,916 points, and the number of point clouds increases by 82.9%.

[0086] The point cloud density distributions of the original signal and the enhanced and restored image are respectively as Figure 14 and Figure 15 shown. After being processed by SAGAN, the point cloud density in the water area has a significant increase. In some areas with better effects, the point cloud density of the underwater signal after enhancement and restoration can approach that of the ground signal.

Claims

1. A method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN, characterized in that The method includes the following steps: Step 1: Save the to-be-processed ground signal fringe pattern and the original underwater signal fringe pattern to relevant folders respectively; Step 2: Add a noise signal to the original ground signal fringe pattern in Step 1 to simulate the underwater signal fringe pattern; Step 3: Improve the original GAN. The generator network uses a U-Net architecture, introduce a residual module in the generator, and construct an adversarial generation network with a self-attention mechanism; Step 4: Scale down the size of the simulated underwater signal fringe pattern obtained in Step 2 from 512×1024 to 256×512 proportionally, input it as a noise signal into the generator network, and input the output result of the generator and the ground fringe pattern in Step 1 into the discriminator network; Step 5: Train the adversarial generation network using the method of mixed-precision floating-point numbers until the discriminator network cannot distinguish the output result of the generator from the original ground fringe pattern, where: use the method of single-precision floating-point numbers during the normalization layer and the process of saving the model, and use the method of half-precision floating-point numbers in other layers and the training process; Step 6: Obtain the parameters of the trained generator network in Step 5, and input the original underwater signal fringe pattern into the trained generator network to obtain the restored and enhanced underwater signal fringe pattern.

2. The method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN according to claim 1, wherein The specific steps of Step 2 are as follows: Step 2-1: Add reverberation noise and random Gaussian noise to the original ground signal fringe pattern; Step 2-2: Add random attenuation to the original ground signal fringe pattern.

3. The method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN according to claim 1, wherein The main architecture of the generator network is a U-Net model, and at the same time introduce a residual module and a self-attention module into it. The input of the generator network is a signal containing noise, and the output is a fake clear signal. The training objective of the generator network is to generate a clear fringe signal from the fringe signal containing noise. After the generator network is trained, input the real underwater target fringe signal and the clear underwater target fringe signal will be output.

4. The method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN according to claim 3, wherein In the generator network, all normalization layers adopt instance normalization layers. During the calculation of the instance normalization layer, normalization is performed on each channel of each sample. The specific calculation method of the instance normalization layer is as follows: (1) Among them, is the input tensor at the th sample, the th channel, the th row, and the th column; is the mean of the th sample and the th channel; is the variance of the th sample and the th channel; is the normalization scaling factor; is the normalization offset; and are learnable parameters; is a constant to prevent division by zero.

5. The method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN according to claim 1 or 3, characterized in that The residual module consists of two convolutional layers, two normalization layers and one ReLU layer. The output size of the residual module is the same as the input size. The output of the residual module is composed of the original input of the module and the calculation results of each layer of the network added together.

6. The method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN according to claim 3, wherein The main architecture of the discriminator network is a convolutional neural network model, add a self-attention module to it, and perform spectral normalization on the convolutional layer.

7. The method for enhancing and restoring the stripe echo signal of a streak tube lidar based on SAGAN according to claim 6, characterized in that LeakyReLU is adopted in the discriminator network, and its function form is as follows: Among them, is the slope of the negative semi-axis.

Images