GM-APD laser radar echo signal target and space-space background noise segmentation method and system

Abstract

The invention relates to the field of an echo signal classification method of a laser imaging radar based on an area array / single-point Geiger mode avalanche photodiode (Gm-APD), and provides a method and a system for segmenting a GM-APD laser radar echo signal target and space-space background noise. And efficient classification and identification of long-distance weak target echo signals and sky background noise echo signals are realized. The method comprises the following steps of: generating simulation data by Monte Carlo; generating a variable frame number statistical echo histogram; generating three-channel data on the basis of the variable-frame-number statistical echo histogram data; and carrying out deep learning training by adopting an HGSNet network, and filtering out space background noise pixels in the three-channel data by adopting the trained HGSNet network to obtain target echo signal data. The method is suitable for realizing efficient classification and identification of long-distance weak target echo signals and space background noise echo signals in a low signal-to-noise ratio environment of the space background.

Description

Technical Field

[0001] The present invention relates to the technical field of a method for classifying echo signals of a planar array / single-point Geiger-mode avalanche photodiode (Gm-APD) lidar, and particularly aims to achieve efficient classification and recognition of long-distance weak target echo signals and aerospace background noise echo signals in a low signal-to-noise ratio environment under the aerospace background. Background Art

[0002] As an advanced single-photon detection technology, GM-APD (Geiger-Mode Avalanche Photodiode) technology has significant advantages in the field of detecting weak small targets at long distances. Its single-photon level sensitivity can effectively obtain target depth information, and is particularly suitable for kilometer-level long-distance detection and low-scattering echo scenarios (such as night environments). However, the core challenge faced by this technology is that the target echo signal and the aerospace background noise are highly similar in time-domain statistical characteristics, resulting in the target signal being easily submerged by noise. Therefore, the key to improving the detection performance lies in effectively suppressing the aerospace background noise and accurately extracting the target echo information.

[0003] Traditional methods adopt a two-stage processing strategy: first, reconstruct the depth and intensity images of the target through echo data, and then perform noise filtering and target recognition based on image texture features. This method has inherent limitations: the reconstruction quality directly affects the noise filtering effect, and key signal features may be lost during the reconstruction process. To overcome these limitations, researchers have proposed a new idea of directly extracting features from echo signals for signal classification, and realizing noise filtering by analyzing multi-dimensional features such as the pulse width, curl, and gradient of echo signals. However, traditional manual feature extraction methods are limited by model robustness and feature completeness, and it is difficult to comprehensively capture the key information of signals.

[0004] In recent years, data-driven end-to-end deep learning networks have provided new solutions to this problem. Such methods not only simplify the feature extraction process by automatically learning feature representations, but also significantly improve the accuracy and efficiency of signal classification. However, the in-depth learning research on GM-APD aerospace background noise segmentation is still in its infancy. Based on this, this patent focuses on researching the GM-APD aerospace background noise segmentation method based on deep learning, aiming to break through the limitations of traditional methods and achieve a significant improvement in the detection performance of weak small targets at long distances. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for segmenting targets and aerospace background noise of GM-APD lidar echo signals, and to achieve efficient classification and recognition of long-distance weak target echo signals and aerospace background noise echo signals.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention proposes a method for segmenting target echo signals and aerospace background noise of GM-APD lidar, and the segmentation method includes the following steps:

[0008] Step S1: Generate simulation data using Monte Carlo;

[0009] Step S2: Generate a statistical echo histogram with variable number of frames based on the simulation data;

[0010] Step S3: Generate three-channel data based on the statistical echo histogram data with variable number of frames;

[0011] Step S4: Perform deep learning training using the HGSNet network, and use the trained HGSNet network to filter out the aerospace background noise pixels in the three-channel data to obtain the target echo signal data.

[0012] Further, the above step S1 is specifically:

[0013] After generating the simulation data, it is divided into a training set and a test set according to a ratio of 9:1, and simulation parameter settings are performed, specifically: the number of noise photons [0.01, 1.011], the number of laser photons [0.01, 10.01], the target position [1, 900], and the laser pulse width covers 25 bins, where [.] represents the specific value of the closed interval.

[0014] Further, the above statistical echo histogram with variable number of frames is generated by changing the average non-zero density of the data and the number of times the target echo is detected while counting the number of frames.

[0015] Further, three-channel data is generated through multi-channel data augmentation, where one-channel data is obtained by filtering background noise from the original data, two-channel data is the target echo photons calculated by inverting each bin, and three-channel data is the histogram compensation channel data.

[0016] Further, the histogram compensation channel data is established by the following formula:

[0017]

[0018] where h m ’(i) is the number of photons actually solved for each bin reaching the detector, hm(n) is the number of photons reaching the nth bin of the detector, exp() represents the natural exponential function, and P i ’ is the statistical frequency of each bin.

[0019] Further, the above HGSNet network includes a feature extraction module and a classification module;

[0020] The feature extraction module includes a channel data smoothing module, a UNet network, and a Transformer module;

[0021] The classification module is constructed based on the ResNet backbone network and the CBAM module is introduced.

[0022] Furthermore, the above-mentioned step S4 is specifically as follows:

[0023] Step S41: The three-channel data first passes through the channel data smoothing module to enhance the signal quality. Subsequently, the UNet network performs multi-scale feature extraction, and the extracted features pass through the Transformer module to complete multi-dimensional feature mining;

[0024] Step S42: The classification module realizes the effective segmentation of the target echo signal and the aerospace background noise pixels from the multi-dimensional features through pixel-by-pixel classification.

[0025] The method for segmenting the target and aerospace background noise of the GM-APD lidar echo signal according to the present invention can be entirely implemented by computer software. Therefore, correspondingly, the present invention also provides a system for segmenting the target and aerospace background noise of the GM-APD lidar echo signal, and the system includes:

[0026] A storage device for generating simulation data using Monte Carlo;

[0027] A storage device for generating a statistical echo histogram with variable frame numbers based on the simulation data;

[0028] A storage device for generating three-channel data based on the statistical echo histogram data with variable frame numbers;

[0029] A storage device for performing deep learning training using the HGSNet network, and filtering out the aerospace background noise pixels in the three-channel data using the trained HGSNet network to obtain the target echo signal data.

[0030] The beneficial effects of the present invention are as follows:

[0031] 1. Aiming at the interference problem of environmental background noise to the detection of weak targets, the present invention proposes a semantic segmentation method based on HGSNet, that is, by combining the multi-scale feature extraction ability of the Unet network and the global attention mechanism of the Transformer, the efficient distinction between the target and the background noise echo signal is realized. Specifically, HGSNet deeply excavates the tiny differences between the target and the noise echo in the high-dimensional feature space, significantly improving the accuracy of semantic segmentation, and the segmentation accuracy rate exceeds 91%. At the same time, the semantic image generated by this method has the characteristics of low false alarm rate and low missed alarm rate, providing high-quality prior information for subsequent target detection and recognition.

[0032] 2. To further improve the detection performance of weak targets, the present invention proposes a signal optimization method based on spatio-temporal domain data distribution modeling and three-channel data enhancement. By modeling the spatio-temporal characteristics of the echo signal and combining with the three-channel data enhancement strategy, this method significantly enhances the expression ability of the target signal while suppressing the interference of background noise. This method not only improves the ability of HGSNet to distinguish between targets and background noise, but also provides richer and more representative training data for deep learning models, further enhancing the robustness and generalization performance of the models.

[0033] Furthermore, compared with the prior art, the present invention has the following advantages:

[0034] (1) High-precision semantic segmentation: By combining the multi-scale feature extraction of Unet and the global attention mechanism of Transformer, HGSNet can effectively capture the subtle differences between targets and background noise, and the semantic segmentation accuracy is higher than 91%, significantly superior to traditional methods.

[0035] (2) Low false alarm rate and low miss alarm rate: The generated semantic images have the characteristics of low false alarm rate and low miss alarm rate, providing high-quality prior information for subsequent target recognition and tracking.

[0036] (3) Signal optimization and enhancement: The spatio-temporal domain data distribution modeling and three-channel data enhancement strategy significantly improve the expression ability of the target signal while suppressing the interference of background noise, enhancing the robustness and generalization performance of the model.

[0037] (4) Detection ability for distant weak targets: The present invention provides reliable technical support for the detection of distant weak targets in complex aerospace backgrounds and has broad application prospects.

[0038] The present invention is applicable to efficiently classifying and identifying the echo signals of distant weak targets and the echo signals of aerospace background noise in a low signal-to-noise ratio environment of the aerospace background. Through the present invention, the detection ability of distant weak target echoes can be significantly improved, providing reliable technical support for target recognition in complex backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0040] Figure 1It is the flowchart of the method for semantic segmentation of small and weak targets and aerospace background signals based on HGSNet proposed by the present invention;

[0041] Figure 2 It is the multi-channel enhanced histogram proposed by the present invention;

[0042] Figure 3 It is the network structure diagram of HGSNet proposed by the present invention;

[0043] Figure 4 It is the semantic segmentation effect diagram of different networks described in the present invention.

[0044] Among them, channel attention represents the channel attention mechanism, and spatial attention represents the spatial attention mechanism. Specific implementation manners

[0045] The following further details the specific implementation manners of the present invention with reference to the accompanying drawings. The following implementation manners will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several changes and improvements can still be made, and these all belong to the protection scope of the present invention.

[0046] Embodiment 1. The purpose of this embodiment is to design an integrated GM-APD semantic segmentation network for target and background classification of GM-APD lidar echo signals, and to achieve efficient classification and recognition of long-distance weak target echo signals and aerospace background noise echo signals through this semantic segmentation network. Therefore, this embodiment proposes a method for segmenting target and aerospace background noise of GM-APD lidar echo signals, and the segmentation method includes the following steps:

[0047] Step S1: Generate simulation data using Monte Carlo;

[0048] Step S2: Generate a statistical echo histogram with variable frame numbers based on the simulation data;

[0049] Step S3: Generate three-channel data based on the statistical echo histogram data with variable frame numbers;

[0050] Step S4: Perform deep learning training using the HGSNet network, and use the trained HGSNet network to filter out the aerospace background noise pixels in the three-channel data to obtain the target echo signal data.

[0051] In this embodiment, by modeling the spatio-temporal domain characteristics of the echo signal and combining the three-channel data enhancement strategy, the expression ability of the target signal is significantly enhanced, and at the same time, the interference of background noise is suppressed. It not only improves the ability of HGSNet to distinguish between the target and background noise, but also provides richer and more representative training data for the deep learning model, further enhancing the robustness and generalization performance of the model. At the same time, this embodiment combines the multi-scale feature extraction ability of the Unet network and the global attention mechanism of the Transformer to achieve efficient discrimination of the target and background noise echo signals. Specifically, HGSNet deeply explores the tiny differences between the target and noise echoes in the high-dimensional feature space, significantly improving the accuracy of semantic segmentation, and the segmentation accuracy rate exceeds 91%. At the same time, the semantic images generated by this method have the characteristics of low false alarm rate and low missed alarm rate, providing high-quality prior information for subsequent target detection and recognition.

[0052] Embodiment 2. Refer to Figures 1 to 3 This embodiment will illustrate this embodiment, which specifically illustrates a method for segmenting the target and aerospace background noise of the echo signal of a GM-APD lidar described in Embodiment 1;

[0053] Step S1: Use Monte Carlo to generate simulation data;

[0054] Specifically:

[0055] This embodiment combines the GM-APD triggering principle and uses the Monte Carlo simulation method to simulate the target echo signals under different environmental conditions. Based on the following formula 1, this embodiment uses the Monte Carlo method to generate 967,045 target echoes and 9,300 noise echo pixel points. To improve the training efficiency, this embodiment randomly selects 8,000 target echoes and 8,000 noise echo data from them to construct a dataset for binary classification. And this dataset is divided into a training set and a test set according to a ratio of 9:1. The simulation parameter settings are as follows: the number of noise photons is [0.01, 1.011], the number of laser photons is [0.01, 10.01], the target position is [1, 900], and the laser pulse width covers 25 bins, where [.] represents the specific numerical value of the closed interval.

[0056]

[0057] Step S2: Generate a statistical echo histogram with variable frame numbers according to the simulation data;

[0058] Specifically:

[0059] To reduce the risk of overfitting during model training, this embodiment proposes a histogram statistical method based on variable frame rate, that is: generating a statistical echo histogram with variable number of frames. Table 1 shows the results after statistically analyzing the test data with different statistical frame numbers. The change in the statistical frame number simultaneously changes the average density of non-zero elements (ADNZE) of the data and the number of times the target echo is detected. Therefore, the variable frame rate operation can expand the distribution of training data and reduce the risk of overfitting. In this patent, for the generated simulation data, random sampling is performed to generate histogram statistical echo signals with different statistical frame numbers.

[0060] Table 1

[0061]

[0062] Step S3: Generate three-channel data based on the statistical echo histogram data with variable number of frames;

[0063] Specifically:

[0064] This embodiment proposes physical model-guided data multi-channel enhancement (DMCE), and expands a 3-channel data model on the basis of the original data, as Figure 2 shown. Channel 1 is obtained after filtering the background noise from the original data; Channel 2 is the target echo photons calculated backward for each bin; Since the GM-APD Lidar detection imaging performs independent repeated sampling on the probability distribution, there will be timestamp distortion due to the pile-up effect during the statistical superposition process. Therefore, a histogram compensation channel 3 is established based on formula (3).

[0065]

[0066] Among them, h m (i) is the photon reaching the nth bin of the detector, and P i is the trigger probability of the detector in the nth bin.

[0067] Based on the above formula (1), after collecting several frames of data to form a statistical histogram, the statistical frequency P i ’ of each bin is obtained, and then the number of echo photons reaching each bin can be solved. The formula is as follows:

[0068]

[0069] Among them, h m ’(i) is the number of photons actually solved to reach each bin of the detector, and h m ’(0) = 0.

[0070]

[0071] Among them, C(i) is the compensation vector of the histogram, N is the total number of statistics, and h c (i) is the channel data after histogram compensation.

[0072] Step S4: Perform deep learning training using the HGSNet network, and use the trained HGSNet network to filter out the aerospace background noise pixels in the three-channel data to obtain the target echo signal data

[0073] Specifically:

[0074] In this embodiment, the HGSNet network is used for deep learning training. In the weak target detection scenario, the similarity between the low SNR target echo and the background noise is relatively high, resulting in significant interference of the background noise pixels on the target detection. Therefore, this embodiment proposes the HGSNet network, a deep learning network combined with a data augmentation strategy. This network first deeply explores the feature differences between the background noise and the low SNR echo through a feature extraction unit, and then uses a classification unit to accurately identify and filter out the background noise pixels. The target semantic prior information generated by HGSNet can provide guidance for subsequent processing, enabling subsequent algorithms to only focus on the target pixel echoes, reducing the computational complexity, and effectively improving the recovery accuracy of the target depth information. The specific architecture of this network is as Figure 3 shown.

[0075] The HGSNet network includes a feature extraction module (Feature extraction module) and a classification module (Classifiaction Module, CF-Module).

[0076] Among them, the feature extraction module (Feature extraction module) consists of three parts: a channel data smoothing module, a UNet network, and a Transformer structure. The input 3-channel data first passes through a convolutional smoothing module to enhance the signal quality. Subsequently, the UNet network is used for multi-scale feature extraction to fully capture the local and global information of the target and the background. At the same time, a Transformer module is introduced to achieve multi-dimensional feature mining through its powerful self-attention mechanism, thereby effectively separating the target and the background noise echo and screening out 32 differential features.

[0077] The Classification Module (CF-Module) is constructed based on the ResNet backbone network and incorporates the Convolutional Block Attention Module (CBAM). The CBAM module enhances the expressive power of key features by adaptively allocating channel and spatial weights, thereby significantly improving the segmentation accuracy. Based on multi-dimensional features, the CF-Module effectively segments the target from the background through pixel-by-pixel classification.

[0078] In addition, this embodiment also designs step S5 and step S6.

[0079] Among them, step S5 is specifically as follows:

[0080] Evaluate the training effect of the model. After training is completed, the network performance is evaluated through simulation data to verify the classification accuracy and robustness of the deep learning network in variable target scenarios. The evaluation metrics adopted in this embodiment include accuracy (Acc), and the specific formula is as follows:

[0081] Acc = (T P + T N ) / (T P + T N + F P + F N ) (6)

[0082] The meanings of the symbols involved in the above formula are shown in Table 2:

[0083] Table 2

[0084] Positive (target) Negative (background) Predicted positive <![CDATA[T P > <![CDATA[F P > Predicted negative <![CDATA[F N > <![CDATA[T N >

[0085] Step S6 is specifically as follows:

[0086] Conduct on-site tests by building a GM-APD lidar system to experimentally verify and quantitatively analyze the effectiveness of the method.

[0087] In summary, aiming at the interference problem of environmental background noise on the detection of weak targets, this embodiment proposes a semantic segmentation method based on HGSNet. This method combines the multi-scale feature extraction ability of the Unet network and the global attention mechanism of the Transformer to achieve efficient discrimination between target and background noise echo signals. Specifically, HGSNet deeply explores the subtle differences between target and noise echoes in the high-dimensional feature space, significantly improving the accuracy of semantic segmentation, and the segmentation accuracy exceeds 91%. At the same time, the semantic images generated by this method have the characteristics of low false alarm rate and low missed alarm rate, providing high-quality prior information for subsequent target detection and recognition.

[0088] Furthermore, to improve the detection performance of small and weak targets, this embodiment also proposes a signal optimization method based on spatio-temporal domain data distribution modeling and three-channel data enhancement. That is, by modeling the spatio-temporal domain characteristics of the echo signal and combining with the three-channel data enhancement strategy, the expression ability of the target signal is significantly enhanced, and at the same time, the interference of background noise is suppressed. This method not only improves the ability of HGSNet to distinguish between targets and background noise, but also provides richer and more representative training data for the deep learning model, further enhancing the robustness and generalization performance of the model.

[0089] Embodiment 3. The method for segmenting the target and aerospace background noise of the GM-APD lidar echo signal proposed in the above embodiment can be entirely implemented by computer software. Correspondingly, this embodiment proposes a system for segmenting the target and aerospace background noise of the GM-APD lidar echo signal, and the system includes:

[0090] A storage device for generating simulation data using Monte Carlo;

[0091] A storage device for generating a statistical echo histogram with variable frame numbers based on the simulation data;

[0092] A storage device for generating three-channel data based on the statistical echo histogram data with variable frame numbers;

[0093] A storage device for performing deep learning training using the HGSNet network, and filtering out the aerospace background noise pixels in the three-channel data using the trained HGSNet network to obtain the target echo signal data.

[0094] Embodiment 4. This embodiment proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the method for segmenting the target and aerospace background noise of the GM-APD lidar echo signal described in any one of the above embodiments.

[0095] Embodiment 5. This embodiment proposes a computer device, which includes a memory and a processor. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method for segmenting the target and aerospace background noise of the GM-APD lidar echo signal described in any one of the above embodiments.

[0096] A computer device provided by this embodiment. The hardware device in this part is a general model and is not shown in the form of a diagram. The system includes a processor and a memory. The processor and the memory can be connected through a bus or other means. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, as well as corresponding program instructions / modules. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions, and modules stored in the memory, so as to implement the methods and steps for segmenting the target and the aerospace background noise of the GM-APD lidar echo signal in the above method embodiments.

[0097] Embodiment 6: To comprehensively evaluate the performance of the model proposed by the present invention, a series of algorithms are selected for comparison in this embodiment. These algorithms include UNet, UNet++, ADAQ_UNet++, CF-Module, and HGSNet. In the subsequent analysis, this embodiment will focus on analyzing the accuracy of the model and verifying the effect of the algorithm in combination with real data, so as to prove its advantages in real scenarios.

[0098] To comprehensively verify the performance of HGSNet, Table 3 compares the test effects of different network architectures (UNet, UNet++, ADAQ_UNet++, CF-Module, and HGSNet) on simulation data. The results show that HGSNet has the highest classification accuracy and is significantly better than other networks. Specifically, compared with UNet, the accuracy of UNet++ has increased by about 4.8%, ADAQ_UNet++ has further increased by about 5.2%, and CF-Module has increased by about 5.0%. And the performance of HGSNet is the most prominent. Compared with UNet, the accuracy has increased by more than 5.5%, fully demonstrating its powerful ability in the target-background segmentation task.

[0099] Table 3

[0100] UNet UNet++ ADAQ_UNet++ CF-Module Proposed 0.8607 0.9022 0.9052 0.9039 0.9105

[0101] Figure 4 Shows the performance of the semantic segmentation task in two small high-voltage line target scenarios in the self-built GM-APD-Lidar system. To adapt to the target-background segmentation problem, the above-mentioned network architectures for comparison have all been appropriately modified and are used for feature extraction and segmentation of the target depth image after retraining. From Figure 4 it can be seen that the performance of HGSNet on real data is also very excellent. Compared with the Ground-truth, the segmentation result of HGSNet is more accurate and can clearly identify the boundary between the target and the background, while there are different degrees of blur or errors in the segmentation effects of other networks.

[0102] In the above description, specific details such as a specific system structure and technology are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, the detailed description of the well-known deep learning network training is omitted to avoid unnecessary details from interfering with the description of the present application.

[0103] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0104] The above are only the embodiments of the present invention and do not limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for segmenting a GM-APD laser radar echo signal target and space background noise, characterized in that: The method is: S1: Generate simulation data using Monte Carlo; S2: Generate a statistical echo histogram with a variable number of frames according to the simulation data; S3: Generate three-channel data based on the statistical echo histogram data of the variable frame number; S4: The HGSNet network is used for deep learning training. The trained HGSNet network is used to filter out the sky and space background noise pixels in the three-channel data to obtain the target echo signal data.

2. The method for segmenting a GM-APD laser radar echo signal target and sky-space background noise according to claim 1, characterized in that: S1 is specifically: After the simulation data is generated, it is divided into training set and test set in a ratio of 9:1, and the simulation parameters are set as follows: number of noise photons [0.01, 1.011], number of laser photons [0.01, 10.01], target position [1,900], laser pulse width covering 25 bins, where [.] represents the specific value of the closed interval.

3. The method for segmenting a GM-APD laser radar echo signal target and sky-space background noise according to claim 1, characterized in that: A statistical echo histogram with a variable frame number is generated by changing the average non-zero density of the data and the number of times the target echo is detected while counting the frame number.

4. The method for segmenting a GM-APD laser radar echo signal target and sky-space background noise according to claim 1, characterized in that: Three-channel data is generated through multi-channel data enhancement, in which the first channel data is obtained after filtering the background noise from the original data, the second channel data is the target echo photons calculated inversely for each bin, and the third channel is the histogram compensation channel data.

5. The method for segmenting a GM-APD laser radar echo signal target and sky-space background noise according to claim 4, characterized in that: The histogram compensation channel data is established by the following formula: Among them, h m ’ (i) is the actual number of photons reaching each bin of the detector, hm(n) is the number of photons reaching the nth bin of the detector, exp() represents the natural exponential function, P i ’ is the statistical frequency of each bin.

6. The method for segmenting a GM-APD laser radar echo signal target and sky-space background noise according to claim 1, characterized in that: The HGSNet network includes a feature extraction module and a classification module; The feature extraction module includes a channel data smoothing module, a UNet network, and a Transformer module; The classification module is built based on the ResNet backbone network and introduces the CBAM module.

7. The method for segmenting a GM-APD laser radar echo signal target and sky-space background noise according to claim 6, characterized in that: S4 is specifically: S41: The three-channel data first passes through the channel data smoothing module to enhance the signal quality, and then the UNet network performs multi-scale feature extraction. The extracted features pass through the Transformer module to complete multi-dimensional feature mining; S42: The classification module realizes effective segmentation of target echo signal and sky background noise pixels from multi-dimensional features through pixel-by-pixel classification.

8. Used in GM-APD laser radar echo signal target and air-space background noise segmentation system, characterized in that: The system comprises a storage device for executing the method and steps described in claim 1.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the method for segmenting a GM-APD laser radar echo signal target and sky-space background noise as described in any one of claims 1 to 7.

10. A computer device, characterized in that: The device includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method for segmenting GM-APD laser radar echo signal targets and sky-space background noise as described in any one of claims 1 to 7.

Images