A lightweight pulse neural network single photon radar back-end signal processing method

Abstract

The present invention provides a lightweight pulse neural network single photon radar back-end signal processing method, including: simulating the optical and electrical processes of a SPAD sensor to collect images, calculating the photon trigger probability of the SPAD sensor; using the photon trigger probability and the grayscale image data set to generate pulse sequence data that obeys the Poisson distribution, and then training the constructed pulse neural network to obtain a trained pulse neural network; quantizing the network parameters of the trained pulse neural network to obtain a lightweight pulse neural network, and then deploying it to a single photon radar signal processing module, and predicting the pulse data collected by the SPAD sensor to obtain classification or imaging results. The present invention can minimize the precision loss of quantization by quantizing the network parameters of the pulse neural network, and converting the network parameters into integers, which can reduce the overhead of computing and hardware resources, and realize high-accuracy, high-efficiency, and high-dynamic range processing of on-chip pulse signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser radar signal processing, and in particular relates to a lightweight pulse neural network single-photon radar back-end signal processing method. Background Art

[0002] Single photon radar uses a single photon avalanche diode (SPAD) working in Geiger mode. It has the advantages of high gain, high sensitivity and low power consumption, and is widely used in fluorescence detection, medical imaging, and autonomous driving. The sensitive response of single photon radar to photons enables it to perform well in low-light conditions.

[0003] In recent years, the spiking neural network (SNN), as the third generation of neural networks, has gained wide attention in the fields of artificial intelligence and neuroscience. Compared with the traditional artificial neural network (ANN), SNN pays more attention to simulating discrete pulse signals in biological nervous systems. It can not only accurately and efficiently process time series data, but also conform to the real characteristics of biology. The SNN network structure is currently the most accurate network to simulate the signal transmission mechanism of the human brain, and has more efficient computing characteristics. It is especially suitable for low-power embedded systems related to timing. It has become the main solution for deploying neural network computing in embedded devices. In recent years, it has been widely used in the application of lidar sensors. With the rapid development of artificial intelligence and hardware technology, the development of the collaboration between embedded devices and neural networks has undoubtedly accelerated. As a result, the pulse signal processing method based on SNN has emerged as a branch, which has better performance than other ANNs in processing time series data on embedded devices.

[0004] As a typical problem of laser radar signal analysis, single photon radar signal processing is widely used in commercial, industrial, military and other fields. Traditional laser radar ranging and imaging algorithms require complex circuit design and occupy a lot of storage resources. Signal processing based on neural networks only requires multiplication and addition operations, which can efficiently extract radar data features, have a large dynamic range, strong noise suppression ability, simple process, easy integration, and can perform real-time data analysis and processing at the hardware level. Although radar data processing based on neural networks has great application potential, the hardware implementation of neural networks based on single photon radar still faces some challenges. First, there is a lack of single photon radar signal data sets based on time series. The existing method is to directly use grayscale images to train CNN (Convolutional Neural Networks) networks, and then convert them into spiking neural networks (SNNs) by adjusting the activation function. Another method is to copy the grayscale image in the time dimension T, so that it has time series information, and directly use the SNN network for training. Although these two methods can train SNN networks with good accuracy, the data type cannot be coupled with the direct output data of the single photon radar, and further processing is required. Secondly, the parameters such as the weights and biases of the trained neural network are 32-bit floating-point numbers, which are unacceptable for the deployment of integrated hardware. Moreover, floating-point operations take up more resources and consume more power, which is far from meeting the requirements of chip integration and low power consumption. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a lightweight pulse neural network single photon radar back-end signal processing method. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] A lightweight pulse neural network single photon radar back-end signal processing method includes:

[0007] S100, simulating the optical and electrical process of the SPAD sensor to collect images and calculating the photon triggering probability of the SPAD sensor;

[0008] S200, using the photon trigger probability and the pre-acquired grayscale image data set to generate pulse sequence data that obeys Poisson distribution, and use the data set as a training data set;

[0009] S300, training the constructed spiking neural network using the training data set to obtain a trained spiking neural network;

[0010] S400, quantizing the network parameters of the trained spiking neural network to obtain a lightweight spiking neural network;

[0011] S500, deploying the lightweight pulse neural network on the single-photon radar signal processing module, and predicting the pulse data collected by the SPAD sensor to obtain classification or imaging results.

[0012] Beneficial effects:

[0013] The present invention provides a lightweight pulse neural network single photon radar back-end signal processing method, comprising:

[0014] Simulate the optical and electrical processes of the SPAD sensor to collect images, and calculate the photon trigger probability of the SPAD sensor; use the photon trigger probability and the pre-acquired grayscale image data set to generate pulse sequence data that obeys the Poisson distribution, and use it as a training data set; use the training data set to train the constructed pulse neural network to obtain a trained pulse neural network; quantize the network parameters of the trained pulse neural network to obtain a lightweight pulse neural network; deploy the lightweight pulse neural network on an embedded device to predict the pulse data collected by the SPAD sensor to obtain classification or imaging results. The present invention can minimize the precision loss of quantization by quantizing the network parameters of the pulse neural network, and at the same time convert the network parameters into integers, which is very friendly to hardware storage and calculation, thereby reducing the overhead of calculation and hardware resources, and realizing high-accuracy, high-efficiency, and high-dynamic range processing of on-chip pulse signals.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a lightweight pulse neural network single photon radar back-end signal processing method provided by the present invention;

[0017] Figure 2 It is a schematic diagram of the process of the lightweight pulse neural network single photon radar back-end signal processing method provided by the present invention;

[0018] Figure 3 Schematic diagram of two light uniforming methods provided by the present invention;

[0019] Figure 4 It is a schematic diagram of the process of generating a training data set provided by the present invention;

[0020] Figure 5 is a schematic diagram of a spiking neural network provided by the present invention;

[0021] Figure 6 It is a flow chart of different states of spiking neurons in a lightweight spiking neural network provided by the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0023] like Figure 1 As shown, the present invention provides a lightweight pulse neural network single photon radar back-end signal processing method comprising:

[0024] S100, simulating the optical and electrical process of the SPAD sensor to collect images and calculating the photon triggering probability of the SPAD sensor;

[0025] The present invention simulates image data collected by a SPAD sensor in a real environment. The simulation is divided into two stages: one is the process from emission to reception of laser signals, and the other is the process of the SPAD sensor capturing photons and converting them into electrical signals.

[0026] Combination Figure 1 and Figure 2 In a specific implementation of the present invention, S100 includes:

[0027] S110, simulating the process of laser signal from emission to reception by SPAD sensor, and calculating the energy density of target reflected light irradiated to SPAD sensor;

[0028] In the LiDAR system, the active imaging scheme uses the laser pulses emitted by the laser to perform imaging based on the intensity and flight time of the received reflected light, while the passive imaging scheme can directly perform imaging based on the natural light reflected by the target, such as sunlight, moonlight, etc. Therefore, the active imaging model can be regarded as considering the emission of laser pulses on the passive imaging model.

[0029] refer to Figure 2 In a specific implementation of the present invention, S110 includes:

[0030] S111, using an active imaging-based uniform illumination method, calculate the energy density of the field of view plane generated by the transmitting laser of the single-photon radar through rectangular or circular uniform illumination;

[0031] Active imaging is based on uniform light illumination. There are two common methods: Figure 3 , one is the rectangular homogenization scheme, the energy density of the field of view plane generated by the emitted laser through the rectangular homogenization is expressed by the formula:

[0032] (1);

[0033] In the formula, is the energy density of the field of view plane produced by the rectangular homogenization, is the laser emission power, and the envelope of the laser pulse conforms to the Gaussian distribution; is the distance from the homogenizer to the field of view plane; and They are the field angles of rectangular homogenization in the vertical and horizontal directions respectively;

[0034] For the circular light homogenization scheme, the horizontal and vertical fields of view are the same, using The energy density of the field of view plane generated by the emitted laser through circular homogenization is expressed as:

[0035] (2);

[0036] In the formula, is the energy density in the field of view plane produced by the circular uniform light, It is the field angle of circular uniform light in the vertical or horizontal direction;

[0037] S112, obtaining a grayscale image dataset, and calculating the pixel grayscale value of each grayscale image contained therein;

[0038] S113, normalizing the pixel grayscale value and using it as reflectivity;

[0039] S114, using the energy density of the field of view plane and the reflectivity, calculate the energy density of the target reflected light irradiating the SPAD sensor.

[0040] When the emitted laser irradiates the target, all targets are diffuse reflectors, and the energy density of the target reflected light irradiating the SPAD sensor is expressed by the formula:

[0041] (3);

[0042] In the formula, is the solid angle formed by the reflected light and the receiving lens of the SPAD sensor. is the reflectivity, is the diameter of the receiving lens, It is the energy density of the field of view plane generated by the emitted laser through circular or rectangular uniform light. and are all integrating variables.

[0043] S120, calculating the photon triggering probability of the SPAD sensor by utilizing the energy density and the process of the SPAD sensor capturing photons and converting them into electrical signals.

[0044] refer to Figure 2 In a specific implementation of the present invention, S120 includes:

[0045] S121, simulating the receiving process of the SPAD sensor to determine the number of photons of the echo received by the SPAD sensor and the number of photons of the background light;

[0046] The SPAD sensor receives reflected light and converts the optical signal into an electrical signal. Due to certain limitations of the SPAD device itself, it cannot receive all reflected light. Considering the influence of the device itself, the number of photons received by the SPAD is expressed as:

[0047] (4);

[0048] In the formula, Indicates the number of echo photons received by the SPAD sensor. represents the transmittance of the receiving lens, represents the fill rate, represents the total area of ​​the pixel, represents the time step, represents the lens coefficient, represents the photon energy, Indicates the number of pixels;

[0049] In addition to the optical signal of the laser transmitter echo, there is also a background light signal. The background light signal is mainly the light signal reflected by the target object irradiated by natural light, external light, etc. The number of background photons received by SPAD is expressed by the formula:

[0050] (5);

[0051] In the formula, is the power of the background light, is the lens bandwidth;

[0052] S122, using the number of photons in the echo and the number of photons in the background light, correspondingly calculating the average number of photon triggering per unit time step based on active imaging and passive imaging;

[0053] After the SPAD receives a photon, the photodiode in Geiger mode will be excited with a certain probability, namely PDE. However, since the hardware circuit itself also has a certain trigger probability when no photons arrive, the count rate of non-light response is called dark count rate. Therefore, the average number of photon triggers per unit time step based on active imaging is expressed as:

[0054] (6);

[0055] In the formula, is the photon detection efficiency;

[0056] When passive imaging is performed, laser echo is not considered. The average number of photon triggers per unit time step based on passive imaging is expressed as:

[0057] (7);

[0058] S123, using the reflectivity and the average number of photon triggers per unit time step based on active imaging and passive imaging, calculate the photon trigger probability of the SPAD sensor for each pixel; the photon trigger probability includes the photon trigger probability based on active imaging and the photon trigger probability based on passive imaging.

[0059] Considering that the number of photon triggers of the SPAD sensor follows the Poisson distribution, it can be concluded that the photon trigger probability can be expressed as:

[0060] (8);

[0061] In the formula, is the photon trigger probability based on active imaging, It is the photon triggering probability based on passive imaging.

[0062] S200, using the photon trigger probability and the pre-acquired grayscale image data set to generate pulse sequence data that obeys Poisson distribution, and use the data set as a training data set;

[0063] refer to Figure 2 In a specific implementation of the present invention, S200 includes:

[0064] S210, for each pixel, inputting the photon triggering probability of the pixel into a Poisson encoder, so that the Poisson encoder generates a random number between 0 and 1, and outputs a pulse of 1 or 0 according to whether the random number is less than the photon triggering probability corresponding to the pixel;

[0065] If the random number is less than the photon triggering probability corresponding to the pixel point, 1 is output, otherwise 0 is output;

[0066] S220, pulses outputting 1 or 0 are organized into pulse sequence data according to a time series, and used as a training data set.

[0067] refer to Figure 4 ,This step uses Poisson coding to generate pulse sequence data that obeys Poisson distribution according to the photon triggering probability, which is used for SNN network training.

[0068] S300, training the constructed spiking neural network using the training data set to obtain a trained spiking neural network;

[0069] refer to Figure 5, the spiking neural network includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer is composed of multiple convolutional layers, pooling layers, and spiking neuron layers, and the output layer is composed of a fully connected layer and a spiking neuron layer; the input layer is used to input spiking data, the convolutional layer is used to extract features using convolution kernels, the pooling layer is used for data pooling, the spiking neuron layer is used for nonlinear mapping, and the fully connected layer is used to integrate and output prediction results; the spiking neurons in the spiking neuron layer are divided into charging, discharging, and resetting states.

[0070] When a pulse arrives, the spiking neuron enters the charging state:

[0071] (9);

[0072] in, is the voltage at the current moment, is the voltage at the previous moment, It is the pulse data at the current moment.

[0073] When the voltage exceeds the threshold When the reset state is entered, there are two reset modes, namely soft reset and hard reset. The voltage after soft reset and hard reset is expressed as:

[0074] (10);

[0075] in, is the reset voltage of the setting.

[0076] When there is no pulse firing, the spiking neuron will leak electricity over time, either proportionally or with a fixed value:

[0077] (11);

[0078] in, is the leakage voltage, is the leakage coefficient, when When , the voltage is reset to zero. In order to reduce the calculation, pulse neurons are used, and the leakage coefficient is 1. Indicates less than or equal to.

[0079] The convolutional layer of the present invention utilizes The convolution kernel is used for feature extraction with a step size of 1, and padding is performed to prevent convolution data loss. Pooling layers of different sizes are adaptively selected according to the pixel size of the input image. For large-scale pixels, The pooling layer quickly reduces the amount of data, and small-scale pixels are used Finally, the fully connected layer is used to integrate features to improve the fitting expression ability of the model.

[0080] Combination Figure 2 and Figure 5 In this step, the training data set can be divided into N blocks, and then the constructed spiking neural network is trained using a synchronous training method. Convolution, spiking neuron activation, and pooling layer pooling complete the convolution part. If the convolution part is not completed, it returns Convolution. If the convolution part is not finished, the image is flattened in one dimension, the fully connected layer, the spike neuron activation, the time dimension data summation, the loss function calculation and the reverse gradient propagation are performed until the training is completed. If the training is not finished, it returns to the convolution part.

[0081] S400, quantizing the network parameters of the trained spiking neural network to obtain a lightweight spiking neural network;

[0082] In order to effectively reduce hardware parameter storage, optimize hardware computing speed, and reduce computing resources, the present invention designs a parameter quantization method for pulse neural networks. The weight parameters, bias parameters, output data of the convolution layer, and input and output data of the pooling layer after general neural network training are 32-bit floating point numbers. The high bit width will not only occupy a large amount of storage resources, but also the calculation cost of decimals will be even larger. In order to better realize the hardware deployment of the prediction network, a parameter quantization method for the SNN network is designed. The network mainly performs weight quantization on the weights. , Bias , output data and the threshold of the spiking neuron Quantify.

[0083] In a specific implementation of the present invention, S400 includes:

[0084] S410, selecting some training samples in the training data set as input data, and inputting all the input data into the trained spiking neural network to obtain output data of the convolutional layer and the fully connected layer;

[0085] S420, using the weights, biases, and output data in the convolutional layer and the fully connected layer as network parameters, and counting the maximum and minimum values ​​of the network parameters;

[0086] Among them, the maximum values ​​of weight, bias and output data are expressed as , , , the minimum value corresponds to .

[0087] S430, setting the number of quantization bits, and calculating the quantization interval based on the number of quantization bits; the quantization interval is expressed as ,in, , where is the number of quantization bits;

[0088] S440, using the quantization interval, determining whether the numerical interval where the network parameter is located includes a true zero point, and if not, supplementing the numerical interval to include the true zero point, thereby obtaining a final numerical interval where the network parameter is located;

[0089] It is worth noting that the weights, biases, and output data are all 32-bit floating point numbers. This step requires quantizing the 32-bit floating point numbers into Since the numerical intervals of weights, biases, and output data may not contain the true zero point, direct quantization will cause the quantization zero point to not correspond to the true zero point. Therefore, this step needs to determine whether the numerical interval contains the true zero point, and adjust the numerical interval to include the true zero point, so as to ensure the calibration of the quantization zero point and the true zero point.

[0090] S450, using the final numerical interval and the quantization formula, calculating the quantization scale and quantization zero point of the network parameter; wherein the quantization formula is expressed as:

[0091] (12);

[0092] In the formula, It is the quantitative scale, is the quantized zero point, is the minimum value of the unquantized network parameter, is the maximum value of the unquantized network parameter, are unquantized network parameters, is the quantized network parameter;

[0093] This step uses the quantization formula to calculate the quantization scale of weights, biases, and output data. , , and quantized zero point , , .

[0094] S460, quantizing the network parameter by using the quantization scale and the quantization zero point of the network parameter to obtain a quantized network parameter;

[0095] This step uses the quantization scale and quantization zero point of the network parameters to quantize the network parameters to obtain the quantized network parameters and the quantized weights. and bias .

[0096] S470, calculating a quantization threshold of a spiking neuron using a quantization scale of a network parameter;

[0097] This step uses the quantization scale of the output data to calculate the quantization threshold of the corresponding spiking neuron. The quantization threshold is expressed as:

[0098] ;

[0099] In the formula, is the unquantized threshold of the spiking neuron, is the quantization scale of the output data of the fully connected layer or convolutional layer, is the quantization threshold.

[0100] S480, integrating the quantized network parameters, and updating the integrated network parameters and quantization thresholds to the trained spiking neural network to obtain a lightweight spiking neural network.

[0101] The quantization principle of the lightweight pulse neural network is:

[0102] (14);

[0103] In the formula, represents the quantization scale of the output data, Represents the quantization zero point of the output data. Represents the quantized weight. Indicates the quantized scale of the input data. represents the quantized input data, represents the quantization zero point of the input data, represents the quantized bias, Represents the number of weights of the convolutional layer or the fully connected layer.

[0104] Since the input is pulse data, there is no need to process the pulse data To quantify, that is:

[0105] (15);

[0106] The output data of the convolutional layer or fully connected layer of the lightweight spiking neural network is expressed as:

[0107] (16);

[0108] In the formula, Represents unquantized input data.

[0109] The integrated network parameters include integrated weights and integrated biases, where the integrated weights are expressed as , the integrated bias is expressed as: .

[0110] S500, deploying the lightweight pulse neural network on the single-photon radar signal processing module, and predicting the pulse data collected by the SPAD sensor to obtain classification or imaging results.

[0111] refer to Figure 6 As shown in FIG, when a lightweight spiking neural network is deployed on a single-photon radar signal processing module to predict the collected pulse data, the spiking neurons are charged, discharged, or reset according to the quantization threshold. Figure 6 The pulse data to be processed in the input is expressed as , the convolutional layer and the fully connected layer output integer data are represented as , After data center processing, it is expressed as , the voltage after the pulse neuron is charged is expressed as , If the voltage after charging is greater than the quantization threshold , then reset its voltage, the pulse neuron outputs 1, if the charged voltage is less than the quantization threshold , then it is discharged, and the voltage after discharge is expressed as , at this time the pulse neuron output is 0.

[0112] The present invention uses a lightweight spiking neural network to predict spiking data. Since the storage and calculation of floating-point numbers are generated by the convolutional layer and the fully connected layer, it is only necessary to quantize the output data, weights, biases of the convolutional layer and the fully connected layer, and the threshold of the spiking neurons.

[0113] The present invention can minimize the precision loss of quantization by asymmetric quantization of the network parameters of the pulse neural network, and at the same time convert the weight value into an integer, which is very friendly to the storage and calculation of the hardware. For the pulse neuron activation layer, negative voltage inhibits the activation of the pulse neuron, and positive voltage promotes the activation of the pulse neuron. Therefore, centralization should be performed before inputting the pulse neuron so that the pulse neuron can work normally.

[0114] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0115] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A lightweight pulse neural network single photon radar back-end signal processing method, characterized in that: include: S100, simulating the optical and electrical process of the SPAD sensor to collect images and calculating the photon triggering probability of the SPAD sensor; S200, using the photon trigger probability and the pre-acquired grayscale image data set to generate pulse sequence data that obeys Poisson distribution, and use the data set as a training data set; S300, training the constructed spiking neural network using the training data set to obtain a trained spiking neural network; S400, quantizing the network parameters of the trained spiking neural network to obtain a lightweight spiking neural network; S500, deploying the lightweight pulse neural network on a single-photon radar signal processing module, and predicting the pulse data collected by the SPAD sensor to obtain a classification or imaging result; S100 includes: S110, simulating the process of laser signal from emission to reception by SPAD sensor, and calculating the energy density of target reflected light irradiated to SPAD sensor; S120, calculating the photon triggering probability of the SPAD sensor by using the energy density and the process of the SPAD sensor capturing photons and converting them into electrical signals; S110 includes: S111, using an active imaging-based uniform illumination method, calculate the energy density of the field of view plane generated by the transmitting laser of the single-photon radar through rectangular or circular uniform illumination; S112, obtaining a grayscale image dataset, and calculating the pixel grayscale value of each grayscale image contained therein; S113, normalizing the pixel grayscale value and using it as reflectivity; S114, using the energy density of the field of view plane and the reflectivity, calculate the energy density of the target reflected light irradiating the SPAD sensor.

2. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 1 is characterized in that: The energy density of the field of view plane generated by the emitted laser in S111 through rectangular homogenization is expressed by the formula: Where P rect is the energy density of the viewing plane produced by the rectangular homogenization, P emit is the laser emission power, and the envelope of the laser pulse conforms to the Gaussian distribution; s is the distance from the homogenizer to the field of view plane; θ ⊥ and θ || They are the field angles of rectangular homogenization in the vertical and horizontal directions respectively; The energy density of the field of view plane generated by the circular homogenization of the emitted laser in S111 is expressed by the formula: Where P circ is the energy density in the field of view plane produced by circular uniform light, θ circ It is the field angle of circular uniform light in the vertical or horizontal direction; The energy density of the target reflected light in S114 irradiating the SPAD sensor is expressed by the formula: Where α is the solid angle between the reflected light and the receiving lens of the SPAD sensor, ρ is the reflectivity, and D lens is the diameter of the receiving lens, P field is the energy density of the field of view plane generated by the emitted laser through circular or rectangular homogenization, θ and are all integrating variables.

3. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 2 is characterized in that: S120 includes: S121, simulating the receiving process of the SPAD sensor to determine the number of photons of the echo received by the SPAD sensor and the number of photons of the background light; S122, using the number of photons in the echo and the number of photons in the background light, correspondingly calculating the average number of photon triggering per unit time step based on active imaging and passive imaging; S123, using the reflectivity and the average number of photon triggers per unit time step based on active imaging and passive imaging, calculate the photon trigger probability of the SPAD sensor for each pixel; the photon trigger probability includes the photon trigger probability based on active imaging and the photon trigger probability based on passive imaging.

4. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 3 is characterized in that: The number of photons in the echo of S121 is expressed by the formula: Where N signal represents the number of photons received by the SPAD sensor, τ represents the transmittance of the receiving lens, ff represents the filling rate, and A pix Represents the total area of ​​the pixel, T bin represents the time step, F represents the lens coefficient, E represents the photon energy, N pixel Indicates the number of pixels; The number of photons of background light in S121 is expressed by the formula: Where P backg is the power of the background light, W band is the lens bandwidth; The average number of photon triggers per unit time step based on active imaging in S122 is expressed as: N acti =(N signal +N backg )·PDE+R dark ; In the formula, R dark is the dark count rate, PDE is the photon detection efficiency; The average number of photon triggers per unit time step based on passive imaging in S122 is expressed as: N passi =N backg ·PDE+R dark ; The photon trigger probability in S123 is expressed by the formula: Where P acti is the photon trigger probability based on active imaging, P passi It is the photon triggering probability based on passive imaging.

5. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 1 is characterized in that: S200 includes: S210, for each pixel, inputting the photon triggering probability of the pixel into a Poisson encoder, so that the Poisson encoder generates a random number between 0 and 1, and outputs a pulse of 1 or 0 according to whether the random number is less than the photon triggering probability corresponding to the pixel; If the random number is less than the photon triggering probability corresponding to the pixel point, 1 is output, otherwise 0 is output; S220, pulses outputting 1 or 0 are organized into pulse sequence data according to a time series, and used as a training data set.

6. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 1 is characterized in that: The pulse neural network includes an input layer, a hidden layer, and an output layer connected in sequence; the hidden layer is composed of multiple convolutional layers, a pooling layer, and a pulse neuron layer, and the output layer is composed of a fully connected layer and a pulse neuron layer; the input layer is used to input pulse data, the convolutional layer is used to extract features using convolution kernels, the pooling layer is used for data pooling, the pulse neuron layer is used for nonlinear mapping, and the fully connected layer is used to integrate and output prediction results; the pulse neurons in the pulse neuron layer are divided into charging, discharging, and resetting states.

7. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 6 is characterized in that: S400 includes: S410, selecting some training samples in the training data set as input data, and inputting all the input data into the trained spiking neural network to obtain output data of the convolutional layer and the fully connected layer; S420, using the weights, biases, and output data in the convolutional layer and the fully connected layer as network parameters, and counting the maximum and minimum values ​​of the network parameters; S430, setting a quantization bit number, and calculating a quantization interval based on the quantization bit number; S440, using the quantization interval, determining whether the numerical interval where the network parameter is located includes a true zero point, and if not, supplementing the numerical interval to include the true zero point, thereby obtaining a final numerical interval where the network parameter is located; S450, using the final numerical interval and the quantization formula, calculating the quantization scale and quantization zero point of the network parameter; S460, quantizing the network parameter by using the quantization scale and the quantization zero point of the network parameter to obtain a quantized network parameter; S470, calculating a quantization threshold of a spiking neuron using a quantization scale of a network parameter; S480, integrating the quantized network parameters, and updating the integrated network parameters and quantization thresholds to the trained spiking neural network to obtain a lightweight spiking neural network.

8. The lightweight pulse neural network single photon radar back-end signal processing method according to claim 7 is characterized in that: The quantization interval in S430 is expressed as [Q min , Q max ],in, Where N bit is the number of quantization bits; The quantization formula in S450 is expressed as: In the formula, S is the quantization scale, Z is the quantization zero point, and D min is the minimum value of the unquantized network parameter, D max is the maximum value of the unquantized network parameter, D is the unquantized network parameter, Q D is the quantized network parameter; The quantization threshold in S460 is expressed as: Where V th is the unquantized threshold of the spiking neuron, S Output is the quantization scale of the output data of the fully connected layer or convolutional layer, Q th is the quantization threshold.

Images