Array SNSPD readout method and device based on compressed sensing

Abstract

The present invention discloses a method and device for array SNSPD readout based on compressed sensing, the method comprising: (1) constructing a deep learning network integrating compressed sensing sampling and reconstruction; (2) using natural scene pictures as sample inputs for deep learning network training; (3) forming a compressed sensing sampling matrix S according to the deep learning network. r 、S c ; (4) Compressed sensing sampling vector S ri ,S ci Load the bias current onto the array SNSPD; (5) Combine the outputs of all pixels in the array SNSPD and obtain the total number of pixels in the array SNSPD that respond based on the output signal amplitude. i ; (6) Repeat steps (4) and (5) until the sampling matrix S r 、S c The rows of are traversed to form the sampling result vector X={x i} T (7) The sampling result vector X is input into the preliminary reconstruction sub-network in the trained deep learning network, and the signal output by the encoding and decoding reconstruction sub-network is the reconstructed signal Y. The present invention is suitable for simultaneous response of multiple pixels, has a simple circuit and a high speed.

Description

Technical Field

[0001] The present invention relates to photon detection technology, and in particular to a method and device for reading out an array SNSPD based on compressed sensing. Background Art

[0002] After nearly 20 years of development, superconducting nanowire single-photon detectors (SNSPDs) have achieved extremely high detection efficiency, extremely low dark counts, and a wide spectral response range. They have played a major role in deep space communications, quantum information, lidar, and bioimaging.

[0003] Single-photon lidar based on SNSPDs has achieved three-dimensional imaging with millimeter resolution and a kilometer range. Previously reported SNSPD-based lidar imaging systems generally use single-pixel detectors, requiring additional scanning systems to achieve large-scale imaging. The lateral resolution of the image is limited by the resolution of the scanning stepping system and is only suitable for imaging static targets. Therefore, a large array of SNSPDs is required to achieve non-scanning imaging. In addition, current single-pixel devices generally have a small photosensitive surface, are not very practical, have limited count rates, and cannot distinguish the number of photons. Therefore, the development of array SNSPDs is an inevitable trend.

[0004] The development of array SNSPDs is primarily limited by the readout circuitry. The most direct readout method is to read out each pixel individually. However, as the number of pixels increases, the heat load introduced by the coaxial line increases, and the limited cooling power of the refrigerator makes this direct readout method unsuitable for larger arrays. Pulse amplitude multiplexing, which uses a single readout line to simultaneously resolve the spatial position and time of photon arrival, also poses a challenge for large-scale scalability. The largest array SNSPD currently contains 1024 pixels and employs row-column multiplexing for readout and a 64-channel time-to-digital converter for recording. A double-layer thermal coupling readout method can avoid the current redistribution effects of row-column multiplexing, but has not yet been fully implemented in thousand-pixel array SNSPDs. In addition to row-column multiplexing, a superconducting nano-delay line based on time-division multiplexing has been applied to readout of array SNSPDs, enabling single-photon imaging with 590 effective pixels. These aforementioned methods are currently relatively efficient readout methods, but they suffer from the limitation of being able to read out multiple pixels simultaneously.

[0005] Although the frequency-division multiplexing readout method can realize the readout of multiple pixels simultaneously, its duty cycle is usually very low and the bias circuit is relatively complex. Single-flux quantum readout circuits and readout circuits based on superconducting nanowire logic devices (nTron) are both based on on-chip encoding to achieve readout, and are two very promising solutions. However, the preparation process of these two methods is relatively complex, especially when the array scale is large, so neither has been achieved on a large scale. Therefore, in order to achieve fast and high-quality three-dimensional imaging, in addition to further expanding the pixel size and area of ​​SNSPD, it is also necessary to develop a readout method that can respond to multiple pixels simultaneously, has a simple circuit, and is fast. Summary of the Invention

[0006] Purpose of the invention: To address the problems existing in the prior art, the present invention provides a method and device for array SNSPD readout based on compressed sensing, which can respond to multiple pixels simultaneously, has a simple circuit and a high speed.

[0007] Technical solution: The array SNSPD readout method based on compressed sensing of the present invention includes:

[0008] (1) Constructing a deep learning network integrating compressed sensing sampling and reconstruction, wherein the deep learning network specifically includes a compressed sampling subnetwork, a preliminary reconstruction subnetwork, and a codec reconstruction subnetwork connected in sequence;

[0009] (2) Using natural scene images as samples to input deep learning networks for training;

[0010] (3) Extract the weight matrix of the compressed sampling subnetwork from the trained deep learning network as the compressed sensing sampling matrix S r 、S c ;

[0011] (4) Compressed sensing sampling vector S ri ,S ci The bias current is loaded onto the array SNSPD, S ri ,S ci Respectively represent S r 、S c The i-th row, S ri ,S ci The number of elements of is consistent with the number of rows and columns of the array SNSPD;

[0012] (5) Combine the outputs of all pixels in the array SNSPD, and use the analog-to-digital conversion module to read the amplitude of the combined output signal, and obtain the total number of pixels x in the array SNSPD according to the amplitude. i ;

[0013] (6) Repeat steps (4) and (5) until the sampling matrix S r 、S c The rows of are traversed to form the sampling result vector X={x i} T ;

[0014] (7) The sampling result vector X is input into the preliminary reconstruction subnetwork in the trained deep learning network, and the signal output by the encoding and decoding reconstruction subnetwork is the reconstructed signal Y, which is the readout signal of the array SNSPD.

[0015] Furthermore, the compressed sampling subnetwork is specifically a first fully connected layer, which is used to perform the following calculation process:

[0016] X j =W1 b D j

[0017] W1 b (i, (u-1) × N + v) = C r (i,u)×C ci (i,v),i∈(1,q),u,v∈(1,N)

[0018] Where D j ∈R m×1 represents the jth sample of the compressed sampling subnetwork input, m represents the sample dimension, m = NxN, N represents the number of rows and columns of the array SNSPD, X j ∈R q×1 It's D j The compressed sensing measurement value obtained after compressed sampling, q is the number of measurements, W1 b ∈R q×m Represents the binary weight matrix, which is composed of the weight matrix C in the fully connected layer setting r 、C c Composition, W1 b (*, #) indicates W1 b The element in the *th row and the #th column of r (i,u) represents C r The uth element of the i-th row, C ci (i,v) represents C c The vth element of the i-th row:

[0019] The preliminary reconstruction sub-network is specifically the second fully connected layer, which is used to perform the following calculation process:

[0020] D j r =X j W2

[0021] Where W2∈Rm×q is the mapping matrix, D j r ∈R m×1 is the response signal after preliminary reconstruction;

[0022] The loss function is:

[0023]

[0024] Where L(W2) represents the loss function and M represents the number of samples;

[0025] The encoding and decoding reconstruction subnetwork includes three convolutional layers and three deconvolutional layers connected in sequence. The three convolutional layers are used to extract signal features, and the three deconvolutional layers are used to reconstruct the response signal.

[0026] Furthermore, step (2) specifically includes:

[0027] (2-1) Divide the natural scene image into several N×N image blocks, extract the brightness component of each image block and store it in a one-dimensional vector. Each one-dimensional vector is used as a sample input to the established deep learning network, where N represents the number of rows and columns of the array SNSPD;

[0028] (2-2) The established deep learning network is trained using the back-propagation method.

[0029] Furthermore, step (3) specifically includes: extracting the weight matrix C of the compressed sampling sub-network from the trained deep learning network r 、C c , as the compressed sensing sampling matrix S r ,S c ∈R q×N :S r =C r ,S c =C c .

[0030] Furthermore, step (4) specifically includes:

[0031] (4-1) The sampling matrix S ri Load it into the array SNSPD according to the following rules: ri The u-th element S ri (u)=1,u∈(1,N) then bias current is applied to the uth row in the array SNSPD, otherwise no bias current is applied;

[0032] (4-2) The sampling matrix S ci Load it into the array SNSPD according to the following rules: ci The vth element S ci(v) = 1, v∈(1, N), then a bias current is applied to the vth column in the array SNSPD, otherwise no bias current is applied.

[0033] The array SNSPD readout device based on compressed sensing of the present invention comprises:

[0034] A deep learning network that integrates compressed sensing sampling and reconstruction. Specifically, it includes a compressed sampling subnetwork, a preliminary reconstruction subnetwork, and an encoding and decoding reconstruction subnetwork connected in sequence. The network uses natural scene images as samples for training.

[0035] The compressed sensing sampling matrix generation module is used to extract the weight matrix of the compressed sampling sub-network from the trained deep learning network as the compressed sensing sampling matrix S r 、S c ;

[0036] Bias current loading module, used to compress the sampling vector S ri ,S ci The bias current is loaded onto the array SNSPD, S ri ,S ci Respectively represent S r 、S c The i-th row, S ri ,S ci The number of elements of is consistent with the number of rows and columns of the array SNSPD;

[0037] The pixel response counting module is used to combine the outputs of all pixels in the array SNSPD, and use the analog-to-digital conversion module to read the amplitude of the combined output signal, and obtain the total number of pixels responding in the array SNSPD according to the amplitude i ;

[0038] The sampling result summary module is used to repeatedly execute the bias current loading module and the pixel response counting module until the sampling matrix S r 、S c The rows of are traversed to form the sampling result vector X={x i} T ;

[0039] The signal reconstruction module is used to input the sampling result vector X into the preliminary reconstruction subnetwork in the trained deep learning network. The signal output by the encoding and decoding reconstruction subnetwork is the reconstructed signal Y, which is the readout signal of the array SNSPD.

[0040] Furthermore, the compressed sampling subnetwork is specifically a first fully connected layer, which is used to perform the following calculation process:

[0041] X j =W1 b Dj

[0042] W1 b (i, (u-1) × N + v) = C r (i,u)×C ci (i,v),i∈(1,q),u,v∈(1,N)

[0043] Where D j ∈R m×1 represents the jth sample of the compressed sampling subnetwork input, m represents the sample dimension, m = NxN, N represents the number of rows and columns of the array SNSPD, X j ∈R q×1 It's D j The compressed sensing measurement value obtained after compressed sampling, q is the number of measurements, W1 b ∈R q×m Represents the binary weight matrix, which is composed of the weight matrix C in the fully connected layer setting r 、C c Composition, W1 b (*, #) indicates W1 b The element in the *th row and the #th column of r (i,u) represents C r The uth element of the i-th row, C ci (i,v) represents C c The vth element of the i-th row;

[0044] The preliminary reconstruction sub-network is specifically the second fully connected layer, which is used to perform the following calculation process:

[0045] D j r =X j W2

[0046] Where W2∈R m×q is the mapping matrix, D j r ∈R m×1 is the response signal after preliminary reconstruction;

[0047] The loss function is:

[0048]

[0049] Where L(W2) represents the loss function and M represents the number of samples;

[0050] The encoding and decoding reconstruction subnetwork includes three convolutional layers and three deconvolutional layers connected in sequence. The three convolutional layers are used to extract signal features, and the three deconvolutional layers are used to reconstruct the response signal.

[0051] Furthermore, the training method of the deep learning network specifically includes:

[0052] The natural scene image is divided into several N×N image blocks, the brightness component of each image block is extracted and stored in a one-dimensional vector, and each one-dimensional vector is used as a sample input to the established deep learning network, where N represents the number of rows and columns of the array SNSPD;

[0053] The established deep learning network is trained using the back-propagation method.

[0054] Furthermore, the compressed sensing sampling matrix generation module is specifically used to:

[0055] Extract the weight matrix C of the compressed sampling subnetwork from the trained deep learning network r 、C c , as the compressed sensing sampling matrix S r ,S c ∈R q×N :S r =C r ,S c =C c .

[0056] Furthermore, the bias current loading module specifically includes:

[0057] The first loading unit is used to load the sampling matrix S ri Load it into the array SNSPD according to the following rules: ri The u-th element S ri (u)=1,u∈(1,N) then bias current is applied to the uth row in the array SNSPD, otherwise no bias current is applied;

[0058] The second loading unit is used to load the sampling matrix S ci Load it into the array SNSPD according to the following rules: ci The vth element S ci (v) = 1, v∈(1, N), then a bias current is applied to the vth column in the array SNSPD, otherwise no bias current is applied.

[0059] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0060] 1. Compared with existing readout solutions, the present invention greatly reduces the complexity of the readout system, shifting the readout complexity to the bias terminal. Readout only requires a single-channel ADC, and the circuit is simple.

[0061] 2. It can handle the situation where multiple pixels respond simultaneously and can read out with high fidelity, which is impossible to achieve with row and column readout and other methods;

[0062] 3. Only some pixels are working at each moment, which can reduce the thermal load of the system while ensuring accurate and efficient reading;

[0063] 4. The sampling matrix and reconstruction process are trained and optimized using deep learning, which has higher accuracy than the traditional compressed sensing reconstruction algorithm. The trained network is used for reconstruction, which improves the readout speed of the array SNSPD. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of compressed sensing readout circuit;

[0065] Figure 2 A diagram of the architecture of a deep learning network that integrates compressed sensing sampling and reconstruction.

[0066] Figure 3 Reconstruct the effect diagram for simulation;

[0067] Figure 4 This is a scanning electron microscope image of a 4×4 pixel SNSPD array;

[0068] Figure 5 A three-dimensional schematic diagram of the compressed sensing readout results obtained based on a 4×4 pixel SNSPD array. DETAILED DESCRIPTION

[0069] The compressed sensing array readout scheme is similar to the process of single-pixel imaging. For an array SNSPD detector with N×N pixels, a sampling matrix S∈R is constructed. q×m , where m = N × N, q = m × MR is the number of sampling times, and MR is the sampling rate. Single-pixel imaging usually uses a random Gaussian matrix or a Hadamard matrix for sampling, but the loading of these sampling matrices requires individual control of each pixel. To solve this problem, the present invention proposes a new solution, which uses a new sampling method based on row and column offset transformation. The principle is as follows Figure 1 As shown, the bias of rows and columns is determined by a one-dimensional random Gaussian matrix S r ,S c ∈R q×N Control and design bias circuits so that they can detect photons normally only when bias current is provided to the rows and columns corresponding to each pixel at the same time. Figure 1 The white and dark gray circles represent the presence or absence of bias in the corresponding rows or columns, respectively. Light-colored nanowires represent pixels that are functioning properly, while black nanowires represent pixels that are not functioning. This is similar to the switching of each digital micromirror (DMD) in single-pixel imaging. The sampling of the readout signal can be expressed as:

[0070] X=S×D

[0071] where D∈R m×1 is the one-dimensional representation of the response signal of each pixel of the detector. q×1 is the sampling value, corresponding to the superposition value of the signal amplitude of each pixel read out under q different biases, that is, Figure 1 The sampling matrix S is composed of the row and column bias matrix C. r ,C c ∈R q×N generate:

[0072] S(i, (u-1)×N+v)=C r (i,u)×C ci (i,v),i∈(1,q),u,v∈(1,N)

[0073] Based on the preset sampling matrix and sampling values, the detector's response signal can be reconstructed, indicating whether different pixels are responsive or not. Because commonly used row and column bias matrices are not conventional sampling matrices, commonly used algorithms such as OMP, ROMP, IHT, and TVAL3 cannot effectively reconstruct the signal. To address this problem, the present invention proposes to construct a deep learning network that integrates sampling and reconstruction, simultaneously optimizing the row and column bias matrices and the reconstruction algorithm during training.

[0074] Based on the above analysis, the present invention provides an array SNSPD readout method based on compressed sensing, which specifically includes the following steps:

[0075] (1) Construct a deep learning network that integrates compressed sensing sampling and reconstruction, such as Figure 2 As shown, the deep learning network specifically includes a compressed sampling subnetwork, a preliminary reconstruction subnetwork and a codec reconstruction subnetwork connected in sequence.

[0076] The compressed sampling subnetwork is specifically the first fully connected layer, which is used to compress the original signal and generate a compressed sensing (CS) measurement value of the original signal. The following calculation process is specifically performed:

[0077] X j =W1 b D j

[0078] W1 b (i, (u-1) × N + v) = C r (i,u)×C ci (i,v),i∈(1,q),u,v∈(1,N)

[0079] Where D j ∈R m×1represents the jth sample of the compressed sampling subnetwork input, m represents the sample dimension, m = NxN, N represents the number of rows and columns of the array SNSPD, X j ∈R q×1 It's D j The compressed sensing measurement value obtained after compressed sampling, q is the number of measurements, W1 b ∈R q×m Represents the binary weight matrix, which is composed of the weight matrix C in the fully connected layer setting r 、C c Composition, W1 b (*, #) indicates W1 b The element in the *th row and the #th column of r (i,u) represents C r The uth element of the i-th row, C ci (i,v) represents C c The vth element of the i-th row: W1 b It can be regarded as a compressed sensing measurement matrix to replace the traditional random Gaussian measurement matrix; q is the number of measurements, q = m × MR, which changes with the measurement rate MR. For an input of 1*1024, q = 256, 103, 41, and 11 correspond to measurement rates MR = 0.25, 0.10, 0.04, and 0.01, respectively.

[0080] W1 b The binarization method used is a deterministic method based on the sign function Sign, namely:

[0081]

[0082] Among them, W1 b is a binary variable, W1 is a real-valued variable, that is, the weight of the compressed sampling sub-network before binarization. Its implementation is very simple and very effective in practice.

[0083] The preliminary reconstruction subnetwork is specifically the second fully connected layer, which is used to preliminarily reconstruct the compressed measurement signal and specifically performs the following calculation process:

[0084] D j r =X j W2

[0085] Where W2∈R m×q is the mapping matrix, the mapping can be regarded as an approximate linear mapping, D j r ∈R m×1 is the response signal after preliminary reconstruction;

[0086] The loss function is:

[0087]

[0088] Where L(W2) represents the loss function and M represents the number of samples.

[0089] The encoding and decoding reconstruction subnetwork is used to further reconstruct the response signal. It includes three convolutional layers and three deconvolutional layers connected in sequence, corresponding to the encoding and decoding processes, respectively. The three convolutional layers are used to extract signal features, and the three corresponding deconvolutional layers are used to reconstruct the response signal. The convolution kernel size of all convolutional layers is 3*3, and the step size is 1*1. The first convolutional layer has 1 input channel and 64 output channels. The last deconvolution layer has 64 input channels and 1 output channel. The remaining convolutional and deconvolution layers have 64 input and output channels. The ReLU activation function is used after each convolutional layer.

[0090] (2) Use natural scene images as samples to input the deep learning network for training.

[0091] This step specifically includes:

[0092] (2-1) Divide a natural scene image into several N×N image blocks. Extract the luminance component of each image block and store it in a one-dimensional vector. Each one-dimensional vector is input into the established deep learning network as a sample, where N represents the number of rows and columns in the array SNSPD. For images whose size is not a multiple of N, resize the image and subtract the remainder from N.

[0093] (2-2) The established deep learning network is trained using the back-propagation method.

[0094] Gradient calculation method for the weight matrix of the compressed sampling sub-network:

[0095] The known binarization operation (i.e. the forward propagation process) is as follows:

[0096] q=W1 b =Sign(W1)

[0097] The derivative of the sign function Sign is zero, so it is obviously impossible to perform back propagation. Therefore, it is necessary to relax the sign function during the back propagation process, assuming that the overall gradient of the preliminary reconstruction sub-network and the deep convolution reconstruction sub-network is g q , where C is the loss function of the entire deep learning network. If the gradient of q is known, then the gradient of W1 is That is, the derivative formula of C with respect to W1 is as follows:

[0098]

[0099] Note that this preserves the information of the gradient and cancels the gradient when W1 is too large, because not canceling the gradient when W1 is too large will significantly degrade the performance. To compress the gradient of the sampling subnetwork, it can be expressed as the Htanh function, which is why the function becomes differentiable, as follows:

[0100] Htanh(W1)=max(-1,min(1,W1))

[0101] (3) Extract the weight matrix C of the compressed sampling subnetwork from the trained deep learning network r 、C c , as the compressed sensing sampling matrix S r ,S c ∈R q×N :S r =C r ,S c =C c .

[0102] (4) Compressed sensing sampling vector S ri ,S ci The bias current is loaded onto the array SNSPD, S ri ,S ci Respectively represent S r 、S c The i-th row of , i = 1,…,q.

[0103] This step specifically includes:

[0104] (4-1) The sampling matrix S ri Load it into the array SNSPD according to the following rules: ri The u-th element S ri (u)=1,u∈(1,N) then bias current is applied to the uth row in the array SNSPD, otherwise no bias current is applied;

[0105] (4-2) The sampling matrix S ci Load it into the array SNSPD according to the following rules: ci The vth element S ci (v) = 1, v∈(1, N), then a bias current is applied to the vth column in the array SNSPD, otherwise no bias current is applied.

[0106] (5) Combine the outputs of all pixels in the array SNSPD, and use the analog-to-digital conversion module to read the amplitude of the combined output signal, and obtain the total number of pixels x in the array SNSPD according to the amplitude. i .

[0107] The analog-to-digital conversion module (ADC) is used to record the output signal amplitude. The number of responding pixels can be determined from the amplitude. For example, if the amplitude of a pixel is a and the current signal amplitude is 3a, then the number of responding pixels can be determined to be 3. Therefore, the total number of responding pixels in the array SNSPD is 3.

[0108] (6) Repeat steps (4) and (5) until the sampling matrix S r 、S c The rows of S are traversed, so ri ,S ci There are q in total, so steps (4) and (5) are repeated q times to complete the traversal and form the sampling result vector X = {x i |i=1,…,q} T .

[0109] (7) The sampling result vector X is input into the preliminary reconstruction subnetwork in the trained deep learning network, and the signal output by the encoding and decoding reconstruction subnetwork is the reconstructed signal Y, which is the readout signal of the array SNSPD.

[0110] This embodiment further provides an array SNSPD readout device based on compressed sensing, comprising:

[0111] A deep learning network that integrates compressed sensing sampling and reconstruction. Specifically, it includes a compressed sampling subnetwork, a preliminary reconstruction subnetwork, and an encoding and decoding reconstruction subnetwork connected in sequence. The network uses natural scene images as samples for training.

[0112] The compressed sensing sampling matrix generation module extracts the weight matrix C of the compressed sampling sub-network from the trained deep learning network. r 、C c , as the compressed sensing sampling matrix S r ,S c ∈R q×N :S r =C r ,S c =C c ;

[0113] Bias current loading module, used to compress the sampling vector S ri ,S ci The bias current is loaded onto the array SNSPD, S ri ,S ci Respectively represent S r 、S c The i-th row of

[0114] The pixel response counting module is used to combine the outputs of all pixels in the array SNSPD, and use the analog-to-digital conversion module to read the amplitude of the combined output signal, and obtain the total number of pixels responding in the array SNSPD according to the amplitude i ;

[0115] The sampling result summary module is used to repeatedly execute the bias current loading module and the pixel response counting module until the sampling matrix S r 、S c The rows of are traversed to form the sampling result vector X={x i} T ;

[0116] The signal reconstruction module is used to input the sampling result vector X into the preliminary reconstruction subnetwork in the trained deep learning network. The signal output by the encoding and decoding reconstruction subnetwork is the reconstructed signal Y, which is the readout signal of the array SNSPD.

[0117] Furthermore, the compressed sampling subnetwork is specifically a first fully connected layer, which is used to perform the following calculation process:

[0118] X j =W1 b D j

[0119] W1 b (i, (u-1) × N + v) = C r (i,u)×C ci (i,v),i∈(1,q),u,v∈(1,N)

[0120] Where D j ∈R m×1 represents the jth sample of the compressed sampling subnetwork input, m represents the sample dimension, m = NxN, N represents the number of rows and columns of the array SNSPD, X j ∈R q×1 It's D j The compressed sensing measurement value obtained after compressed sampling, q is the number of measurements, W1 b ∈R q×m Represents the binary weight matrix, which is composed of the weight matrix C in the fully connected layer setting r 、C c Composition, W1 b (*, #) indicates W1 b The element in the *th row and the #th column of r (i,u) represents C r The uth element of the i-th row, C ci (i,v) represents C c The vth element of the i-th row;

[0121] The preliminary reconstruction sub-network is specifically the second fully connected layer, which is used to perform the following calculation process:

[0122] D j r =X j W2

[0123] Where W2∈R m×q is the mapping matrix, D j r ∈R m×1 is the response signal after preliminary reconstruction;

[0124] The loss function is:

[0125]

[0126] Where L(W2) represents the loss function and M represents the number of samples;

[0127] The encoding and decoding reconstruction subnetwork includes three convolutional layers and three deconvolutional layers connected in sequence. The three convolutional layers are used to extract signal features, and the three deconvolutional layers are used to reconstruct the response signal.

[0128] Furthermore, the training method of the deep learning network specifically includes:

[0129] The natural scene image is divided into several N×N image blocks, the brightness component of each image block is extracted and stored in a one-dimensional vector, and each one-dimensional vector is used as a sample input to the established deep learning network, where N represents the number of rows and columns of the array SNSPD;

[0130] The established deep learning network is trained using the back-propagation method.

[0131] Furthermore, the bias current loading module specifically includes:

[0132] The first loading unit is used to load the sampling matrix S ri Load it into the array SNSPD according to the following rules: ri The u-th element S ri (u)=1,u∈(1,N) then bias current is applied to the uth row in the array SNSPD, otherwise no bias current is applied;

[0133] The second loading unit is used to load the sampling matrix S ci Load it onto the array SNSPD according to the following rules: ci The vth element S ci (v) = 1, v∈(1, N), then a bias current is applied to the vth column in the array SNSPD, otherwise no bias current is applied.

[0134] This device corresponds to the above method one by one. For any incomplete details, please refer to the method and will not be repeated here.

[0135] The present invention was simulated and validated using the TensorFlow framework for training and testing. The training set used the same 91 images as the SRCNN network. Each of these 91 images was partitioned into 32x32 blocks with a stride of 14, resulting in a total of 21,760 blocks sampled from the 91 images. The test set was also the same as that used in the SRCNN network. The maximum number of iterations and learning rate were set to 1500 and 0.0001, respectively. Figure 3 is the result of the trained network reconstruction. Even at a sampling rate as low as 0.01, the algorithm can still reconstruct the image contour. Taking the 4×4 pixel SNSPD array as an example, Figure 4 A scanning electron microscope image of the device is shown. Figure 5 The imaging results of the light spot hitting the array SNSPD at different sampling rates are shown. The results obtained at the three sampling rates all restore the information of the light spot well, proving the effectiveness of the present invention.

Claims

1. A method for reading array SNSPD based on compressed sensing, characterized in that The method includes: (1) Constructing a deep learning network that integrates compressed sensing sampling and reconstruction, wherein the deep learning network specifically includes a compressed sampling subnetwork, a preliminary reconstruction subnetwork, and a codec reconstruction subnetwork connected in sequence; (2) Use natural scene images as samples to input deep learning networks for training; (3) Extract the weight matrix of the compressed sampling subnetwork from the trained deep learning network as the compressed sensing sampling matrix ; (4) Compressed sensing sampling vector Loaded onto the array SNSPD in the form of bias current, Respectively No. i OK, The number of elements of is consistent with the number of rows and columns of the array SNSPD; (5) Combine the outputs of all pixels of the array SNSPD, and use the analog-to-digital conversion module to read the amplitude of the combined output signal, and obtain the total number of pixels responding in the array SNSPD based on the amplitude x i ; (6) Repeat steps (4) and (5) until the sampling matrix The rows are traversed to form a sampling result vector ; (7) The sampling result vector X is input into the preliminary reconstruction sub-network in the trained deep learning network. The signal output by the encoding and decoding reconstruction sub-network is the reconstructed signal Y, which is the readout signal of the array SNSPD; Wherein, step (4) specifically includes: (4-1) The sampling matrix Load it onto the array SNSPD according to the following rules: No. u Elements =1, , then the first u If the bias current is applied, no bias current is applied. (4-2) The sampling matrix Load it onto the array SNSPD according to the following rules: No. v Elements =1, , then the first v The bias current is applied to the column, otherwise no bias current is applied.

2. The array SNSPD readout method based on compressed sensing according to claim 1, characterized in that: The compressed sampling subnetwork is specifically the first fully connected layer, which is used to perform the following calculation process: ， ， Where, represents the first input of the compressed sampling subnetwork j samples, m represents the sample dimension, m=NxN, Indicates the number of rows and columns of the array SNSPD, yes The compressed sensing measurement value obtained after compressed sampling, q is the number of measurements, Represents the binary weight matrix, which is composed of the weight matrix C in the fully connected layer setting r 、C c constitute, express The element in the *th row and the #th column of r (i,u) represents C r No. i The first u elements, C ci (i,v) represents C c No. i The first v elements; The preliminary reconstruction sub-network is specifically the second fully connected layer, which is used to perform the following calculation process: , Where, is the mapping matrix, is the response signal after preliminary reconstruction; The loss function is: , Where, represents the loss function, represents the number of samples; The encoding and decoding reconstruction subnetwork includes three convolutional layers and three deconvolutional layers connected in sequence. The three convolutional layers are used to extract signal features, and the three deconvolutional layers are used to reconstruct the response signal.

3. The array SNSPD readout method based on compressed sensing according to claim 1, characterized in that: Step (2) specifically includes: (2-1) Divide the natural scene image into several N×N image blocks, extract the brightness component of each image block and store it in a one-dimensional vector, and use each one-dimensional vector as a sample input to establish a deep learning network. Indicates the number of rows and columns of the array SNSPD; (2-2) The established deep learning network is trained using the back-propagation method.

4. The array SNSPD readout method based on compressed sensing according to claim 1, characterized in that: Step (3) specifically includes: extracting the weight matrix C of the compressed sampling sub-network from the trained deep learning network r 、C c , as the compressed sensing sampling matrix : .

5. A SNSPD array readout device based on compressed sensing, characterized in that include: A deep learning network that integrates compressed sensing sampling and reconstruction. Specifically, it includes a compressed sampling subnetwork, a preliminary reconstruction subnetwork, and an encoding and decoding reconstruction subnetwork connected in sequence. The network uses natural scene images as samples for training. The compressed sensing sampling matrix generation module is used to extract the weight matrix of the compressed sampling sub-network from the trained deep learning network as the compressed sensing sampling matrix ; Bias current loading module, used to compress the sensing sampling vector Loaded onto the array SNSPD in the form of bias current, Respectively No. i OK, The number of elements of is consistent with the number of rows and columns of the array SNSPD; The pixel response counting module is used to combine the outputs of all pixels of the array SNSPD, and use the analog-to-digital conversion module to read the amplitude of the combined output signal, and obtain the total number of pixels responding in the array SNSPD based on the amplitude x i ; The sampling result summary module is used to repeatedly execute the bias current loading module and the pixel response counting module until the sampling matrix The rows are traversed to form a sampling result vector ; Signal reconstruction module, used to transform the sampling result vector The initial reconstruction sub-network in the trained deep learning network is input, and the signal output by the encoding and decoding reconstruction sub-network is the reconstructed signal Y, which is the readout signal of the array SNSPD; Wherein, the bias current loading module specifically includes: The first loading unit is used to load the sampling matrix Load it onto the array SNSPD according to the following rules: No. u Elements =1, , then the first u If the bias current is applied, no bias current is applied. The second loading unit is used to load the sampling matrix Load it onto the array SNSPD according to the following rules: No. v Elements =1, , then the first v The bias current is applied to the column, otherwise no bias current is applied.

6. The array SNSPD readout device based on compressed sensing according to claim 5, characterized in that: The compressed sampling subnetwork is specifically the first fully connected layer, which is used to perform the following calculation process: , ， Where, represents the first input of the compressed sampling subnetwork j samples, m represents the sample dimension, m=NxN, Indicates the number of rows and columns of the array SNSPD, yes The compressed sensing measurement value obtained after compressed sampling, q is the number of measurements, Represents the binary weight matrix, which is composed of the weight matrix C in the fully connected layer setting r 、C c constitute, express The element in the *th row and the #th column of r (i,u) represents C r No. i The first u elements, C ci (i,v) represents C c No. i The first v elements; The preliminary reconstruction sub-network is specifically the second fully connected layer, which is used to perform the following calculation process: , Where, is the mapping matrix, is the response signal after preliminary reconstruction; The loss function is: , Where, represents the loss function, represents the number of samples; The encoding and decoding reconstruction subnetwork includes three convolutional layers and three deconvolutional layers connected in sequence. The three convolutional layers are used to extract signal features, and the three deconvolutional layers are used to reconstruct the response signal.

7. The array SNSPD readout device based on compressed sensing according to claim 5, characterized in that: The training method of the deep learning network specifically includes: Divide the natural scene image into several N×N image blocks, extract the brightness component of each image block and store it in a one-dimensional vector, and use each one-dimensional vector as a sample input to establish a deep learning network. Indicates the number of rows and columns of the array SNSPD; The established deep learning network is trained using the back-propagation method.

8. The array SNSPD readout device based on compressed sensing according to claim 5, characterized in that: The compressed sensing sampling matrix generation module is specifically used for: Extract the weight matrix C of the compressed sampling subnetwork from the trained deep learning network r 、C c , as the compressed sensing sampling matrix : .

Images