Encoding methods and systems based on deep learning and compressed sensing

Abstract

This invention discloses an encoding method and system based on deep learning and compressed sensing, belonging to the field of optical transmission technology. The method includes: encoding and mapping bit data; designing a codebook using the AlexNet model; training each codebook through five convolutional layers to obtain the bit error rate (BER) under the current environment; mapping data to the codebook with the lowest BER; performing compressed sensing on the data signal mapped to the codebook; the compressed sensing signal is first modulated by OFDM, then transmitted, demodulated, reconstructed, and demapped back to the codebook signal; finally, the codebook signal is decoded back to the original signal using a decoder. By combining SCMA modulation with convolutional neural network technology and compressed sensing technology, an SCMA encoder and decoder are established, minimizing the BER. The use of compressed sensing theory to process the signal significantly reduces the amount of data to be acquired, transmitted, and processed, achieving data transmission with fewer subcarriers.

Description

Technical Field

[0001] This invention belongs to the field of optical transmission technology, specifically relating to a coding method and system based on deep learning and compressed sensing. Background Technology

[0002] 5G access is still based on Orthogonal Frequency Division Multiplexing (OFDM) waveforms, but it also features non-orthogonal waveforms and multiple access capabilities. Non-orthogonal multiple access methods effectively support higher capacity with greater flexibility, robustness, and adaptability to a large number of connections. These properties help provide a better user experience for enhanced mobile broadband (eMBB) services and can even be used in combination with other service types and other use cases defined for 5G. One non-orthogonal multiple access scheme is Sparse Code Multiple Access (SCMA). With the development of non-orthogonal multiple access technologies, SCMA has seen increasing application. To meet the needs of 5G communication, it innovatively employs codebook technology at the transmitting end, combining modulation and spread spectrum techniques from traditional communication, increasing user flexibility and resource utilization. The performance of SCMA largely depends on the codebook design. SCMA's codebook mapping is similar to the layer mapping in Long Term Evolution (LET), where one or more SCMA layers are assigned to the user / data stream. However, at each SCMA layer, a mapping from information bits to codewords is also performed; that is, the SCMA modulator maps input bits to complex, multidimensional codewords selected from the SCMA codebook of a specific layer. SCMA codewords are sparse, meaning only a few of their entries are non-zero, with the rest being zero. Information bits are assigned to multidimensional complex codewords using a predefined codebook, and these codewords are not orthogonal to each other. The received signal can be decoded using maximum likelihood (ML) decoding or via message passing algorithm (MPA). However, designing the codebook manually has drawbacks because the codewords in the codebook are not orthogonal to each other and consist of multidimensional eigenvalues. Different codebooks need to be constructed for different environments, leading to excessively high requirements for algorithm efficiency. Furthermore, decoding using ML or iterative MPA, both of which have high computational overhead, limits the real-time performance of SCMA. Therefore, optimizing the codebook design is of great importance. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the purpose of this invention is to provide an encoding method and system based on deep learning and compressed sensing, which solves the problems in the prior art.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] The deep learning- and compressed sensing-assisted coding method includes the following steps:

[0006] The bit data is encoded and mapped, and the codebook is designed using the AlexNet model. Each codebook is trained through five convolutional layers to obtain the bit error rate in the current environment.

[0007] Data mapping is performed on the codebook with the lowest bit error rate;

[0008] Compressed sensing is performed on the data signals mapped onto the codebook;

[0009] The compressed sensing signal is first modulated by OFDM, then transmitted, demodulated, reconstructed and demapped into a codebook signal, and finally decoded back into the original signal by a decoder.

[0010] Furthermore, the codebook design using the AlexNet model relies on CNN training to enable the encoder f kj (r;θ f ), decoder g(y; θ) g The original data r is correctly reconstructed; the end-to-end loss function of the CNN, considering both the encoder and decoder, is:

[0011]

[0012] In the formula, h is the channel vector, r is the target value, f is the encoder, and g is the decoder. The data reconstructed for the CNN decoder, i.e.

[0013] Furthermore, the AlexNet model in CNNs uses a dual-GPU-based convolution operation method and replaces the T activation function with the ReLU activation function. The ReLU activation function is expressed as:

[0014] f(x) = max(0,x).

[0015] Furthermore, the data mapping involves transmitting six user data points using four subcarriers. The mapping process is as follows: each of the six users transmits two symbols, with each symbol occupying one subcarrier. The four subcarriers are represented by b1, b2, b3, and b4, respectively. The first user data is mapped to b1 and b2, the second user data to b1 and b3, the third user data to b1 and b4, the fourth user data to b2 and b3, the fifth user data to b2 and b4, and the last user data to b3 and b4.

[0016] Furthermore, compressed sensing of the data signal mapped onto the codebook involves sampling and compressing the signal, and the steps include:

[0017] S31, Suppose x is a one-dimensional discrete signal of length N, then use a set of orthogonal bases. Represented by a linear combination, forming an N×N basis matrix as column vectors, i.e., Ψ = {Ψ1, Ψ2, …, Ψ N}, then the signal x can be represented by these basis vectors as:

[0018]

[0019] where θ = [θ1, θ2, …… θ N T is the sparse representation coefficient of the signal x under the basis Ψ, and only a small part of the values in θ are large, while most of the values are zero or close to zero; if there are S non-zero elements among them,

[0020] the signal x is said to be S-sparse;

[0021] When S < N, design a measurement matrix Φ of size M×N (M << N), which is uncorrelated with the transformation basis Ψ; the measurement of the signal x through the measurement matrix Φ can obtain a measurement vector of M×1:

[0022] y = Φx = ΦΨθ = Αθ

[0023] The high-dimensional signal Ψθ is mapped to a low-dimensional space under the measurement of the observation matrix Φ, that is, the compressed signal y is obtained, completing the sampling compression of the signal.

[0024] Furthermore, the steps to decode the codebook signal into the original signal are as follows:

[0025] S41, for each resource, calculate the probability of each possible codeword on this resource according to the received signal and channel gain, and transmit these probabilities to the corresponding user nodes;

[0026] S42, each user node updates the probability of the codeword it selects according to the probability information received from other resource nodes, and transmits these probabilities back to the resource nodes;

[0027] S43, repeat S41 - S42 until the predetermined number of iterations is reached;

[0028] S44, after the iteration ends, calculate the likelihood ratio of each bit according to the final probability distribution of the user nodes, and perform bit decision according to the likelihood ratio to recover the original signal.

[0029] The coding system based on deep learning and compressive sensing assistance includes:

[0030] Codebook design module: perform coding mapping on bit data, use the AlexNet model for codebook design, and each codebook obtains the bit error rate in the current environment after being trained through five convolutional layers;

[0031] ​Data mapping module: performs data mapping on the codebook with the lowest bit error rate;

[0032] Compressed sensing module: Performs compressed sensing on the data signals mapped onto the codebook;

[0033] And, the reconstruction and decoding module: the compressed sensing signal is first modulated by OFDM, then transmitted, then demodulated, then reconstructed and demapped into a codebook signal, and finally decoded into the original signal by the decoder.

[0034] A computer storage medium storing a readable program that, when the program is run, can execute the aforementioned deep learning and compressed sensing-assisted coding method.

[0035] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0036] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the deep learning and compressed sensing-assisted encoding method described above.

[0037] A computer program product includes computer instructions that instruct a computing device to perform operations corresponding to the aforementioned deep learning and compressed sensing-assisted encoding method.

[0038] The beneficial effects of this invention are:

[0039] This invention combines SCMA modulation with convolutional neural network technology and compressed sensing technology to establish an SCMA encoder and decoder. They are composed of basic AlexNet units consisting of multiple convolutional and fully connected layers, with AlexNet used for codebook design and signal encoding / decoding. The proposed neural network learns autonomously based on the environment, thereby minimizing the bit error rate. Simultaneously, this system uses compressed sensing theory to process signals, significantly reducing the amount of data that needs to be acquired, transmitted, and processed, achieving data transmission with fewer subcarriers. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the coding method based on deep learning and compressed sensing assisted by the present invention;

[0042] Figure 2 This is a diagram of the AlexNet network structure of the present invention;

[0043] Figure 3 This is a comparison diagram of the codebook mapping process based on compressed sensing in this invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] like Figure 1 As shown, the encoding method based on deep learning and compressed sensing includes the following steps:

[0047] S1, the bit data is encoded and mapped, and the codebook is designed using the AlexNet model. Each codebook is trained through five convolutional layers to obtain the bit error rate in the current environment.

[0048] First, relevant data from the SCMA system needs to be systematically collected and preprocessed, including key information such as factor graph matrices, user data, and resource block allocation. Then, a targeted feature extraction strategy is designed to extract valuable features for codebook design from the complex SCMA signals. Based on this, an AlexNet deep learning model is constructed. This model, with its multi-layer convolution, pooling, and fully connected design, is adapted to the characteristics of the SCMA system. It will be used to learn the inherent patterns of the SCMA system and the complex relationships between features, thereby guiding the selection of codebook parameters, such as constellation graph configuration, codebook dimensionality, and sparsity. After sufficient model training, the performance of the designed codebook is predicted and evaluated, focusing on key indicators such as bit error rate, throughput, and resource efficiency. Based on the evaluation results, the codebook design, as well as the parameters and structure of the deep learning model, are continuously adjusted and optimized, forming an iterative optimization closed loop.

[0049] The network structure of the AlexNet model is as follows: Figure 2 As shown, the codebook design using the AlexNet model relies on training with a convolutional neural network (CNN). The proposed CNN structure must be trained to enable the encoder f. kj (r;θ f ), decoder g(y; θ) gIt can correctly reconstruct the original data r; therefore, the end-to-end loss function of a CNN that considers both the encoder and decoder can be written as:

[0050]

[0051] In the formula, h is the channel vector, r is the target value, f is the encoder, and g is the decoder. The data reconstructed for the CNN decoder, i.e.

[0052] In this scheme, input data symbols r are randomly generated during training and fed into the encoder. Then, h and n are randomly generated and added to the encoder output to obtain hf(r; θ). f )+n, after which, hf(r; θ f +n) simulates the receiver input under actual channel conditions, which is then input into the decoder to reconstruct r. Note that the randomly generated r, h, and n constitute the training dataset for this scheme. The impact of noise n is represented by the damage level η, where η represents the average damage (noise) power divided by the average transmitted signal power (E[|x|). 2 ]), that is, η=σ 2 / E[|x| 2 To establish the optimal encoder and decoder, the appropriate corruption level must be selected to achieve the correct bit error rate through multiple training runs on different codebooks. Since each user has a completely different and compatible codebook, a large number of codebooks need to be trained, necessitating a faster training speed to meet the requirements.

[0053] The AlexNet model in convolutional neural networks uses a dual-GPU-based convolution operation method and replaces the T activation function with the ReLU activation function, greatly improving training speed and solving problems encountered during codebook training. The ReLU activation function is expressed as follows:

[0054] f(x) = max(0,x)

[0055] After training through five convolutional layers, the user codebook addresses the noise inherent in the user data within fully connected layers. By using dropout, overfitting of the neural network can be effectively prevented. Overfitting can be viewed as an overly complex model that easily exploits seemingly valid but actually useless relationships, often leading to misclassification. Therefore, with each input of the user codebook, the neural network tries a different structure, with weights shared across all these structures. For a given layer of neurons, some neurons are randomly removed using defined probabilities, while maintaining the same number of neurons in the input and output layers. Then, parameters are updated according to the neural network's learning method. In the next iteration, some neurons are randomly removed again until training ends, ensuring the accuracy of the final bit error rate.

[0056] Applying the AlexNet model to this invention can better construct high-dimensional user codebooks. Furthermore, combining the high-dimensional user codebook with the AlexNet model improves system robustness while ensuring lower decoding complexity, thus resulting in superior overall system performance.

[0057] S2, perform data mapping on the codebook with the lowest bit error rate obtained;

[0058] This data mapping can be understood as transmitting six user data entries using four subcarriers. The specific mapping process is as follows: each of the six users transmits two symbols, with each symbol occupying one subcarrier. The four subcarriers are represented by b1, b2, b3, and b4. The first user's data is mapped to b1 and b2, the second user to b1 and b3, the third user to b1 and b4, the fourth user to b2 and b3, the fifth user to b2 and b4, and the last user to b3 and b4. This ensures the uniqueness of each user's codebook, reduces subcarrier occupancy, and achieves efficient spectrum utilization.

[0059] S3 performs compressed sensing on the data signal mapped to the codebook, further optimizing the data, reducing the use of subcarriers, and reducing complexity.

[0060] Compressed sensing of data signals mapped onto a codebook involves sampling and compressing the signal. Unlike traditional signal acquisition, sampling and compression occur simultaneously in compressed sensing theory. This process discards redundant information in the current signal sample and directly obtains compressed samples from the continuous-time signal transformation; for example... Figure 3 As shown, the steps for sampling and compressing the data signal mapped onto the codebook include:

[0061] S31, assuming x is a one-dimensional discrete signal of length N, then it can be represented by a set of orthogonal bases. is represented by a linear combination, and used as column vectors to form an N×N dimensional basis matrix, i.e., Ψ = {Ψ1, Ψ2, …, Ψ N}, then the signal x can be represented by these basis vectors as:

[0062]

[0063] where θ = [θ1, θ2, …… θ N T is the sparse representation coefficient of the signal x in the basis Ψ, and only a small part of the values in θ are large, while most of the values are zero or close to zero; if there are S non-zero elements among them, the signal x is called S-sparse;

[0064] For S << N, design a measurement matrix Φ of size M×N (M << N), which is uncorrelated with the transform basis Ψ; the measurement of the signal x through the measurement matrix Φ can obtain a measurement vector of size M×1:

[0065] y = Φx = ΦΨθ = Αθ

[0066] The high-dimensional signal Ψθ is mapped to a low-dimensional space under the measurement of the observation matrix Φ, that is, the compressed signal y is obtained, and the sampling compression of the signal is completed.

[0067] S4, the signal after compressive sensing first undergoes OFDM modulation, then signal transmission, followed by signal demodulation, and then the signal is reconstructed and demapped into a codebook signal, and finally the codebook signal is decoded into the original signal by the decoder;

[0068] During transmission, the codeword of each layer is selected based on the input bit sequence. After compressive sensing, the codewords from different layers are combined before the OFDM modulator and transmitted through a smaller number of subcarriers. The signal after being modulated by the OFDM modulator is transmitted through the channel. The problem of reconstructing the signal at the receiving end is actually to solve the above formula, that is, to solve a system of linear equations, and the original data can be obtained after demapping and passing through the decoder.

[0069] The process of signal OFDM modulation is as follows: convert the serial data into multiple parallel data streams, the data streams undergo codebook mapping to form multiple complex sub-symbols. Each complex sub-symbol corresponds to one path of data, the signal is sampled and observed by compressive sensing, and the signal is mapped to a low-dimensional space through the observation matrix, and then the data is allocated to different subcarriers. Perform an inverse fast Fourier transform (IFFT) on the complex sub-symbols on the codebook. The IFFT transforms the complex sub-symbols in the frequency domain into real-valued samples in the time domain, add a cyclic prefix before the real symbols, and finally perform digital-to-analog conversion and enter the channel for transmission.

[0070] ​The process of signal demodulation, reconstruction, and demapping is as follows: remove the cyclic prefix from the received signal, perform FFT processing on the symbol block after removing the cyclic prefix to obtain the frequency domain signal, perform frequency domain equalization processing on the signal, reconstruct the signal through the observation matrix after equalization, demapping the reconstructed signal, and restore the obtained signal to the bit stream of the transmitted data through the decoder.

[0071] The steps to decode the codebook signal into the original signal are as follows:

[0072] S41, For each resource (or subcarrier), calculate the probability of each possible codeword on that resource based on the received signal and channel gain, and pass these probabilities to the corresponding user node;

[0073] S42, each user node updates the probability of its selected codeword based on the probability information received from other resource nodes, and sends these probabilities back to the resource nodes;

[0074] S43, repeat S41-S42 until the predetermined number of iterations is reached;

[0075] S44. After the iteration is completed, the likelihood ratio of each bit is calculated according to the final user node probability distribution. Bit decision is made based on the likelihood ratio to recover the original signal.

[0076] Based on a similar inventive concept, embodiments of the present invention also provide a computer storage medium storing a readable program that, when the program is run, can execute the above-described encoding method based on deep learning and compressed sensing.

[0077] Based on a similar inventive concept, this invention provides an electronic device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0078] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the deep learning and compressed sensing-assisted encoding method described above.

[0079] Based on a similar inventive concept, embodiments of the present invention also provide a computer program product, including computer instructions, which instruct a computing device to perform the operations corresponding to the above-described deep learning and compressed sensing-assisted encoding method.

[0080] Example 2

[0081] Deep learning and compressed sensing-assisted coding systems specifically include:

[0082] Codebook design module: Encodes and maps bit data, uses the AlexNet model for codebook design, and obtains the bit error rate in the current environment after training through five convolutional layers for each codebook.

[0083] Data mapping module: performs data mapping on the codebook with the lowest bit error rate;

[0084] Compressed sensing module: Performs compressed sensing on the data signals mapped onto the codebook;

[0085] Reconstruction and Decoding Module: The compressed sensing signal is first modulated by OFDM, then transmitted, demodulated, reconstructed and demapped into a codebook signal, and finally decoded into the original signal by a decoder.

[0086] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A coding method based on deep learning and compressed sensing, characterized in that, It includes the following steps: Perform coding mapping on bit data, use the AlexNet model for codebook design, and each codebook obtains the bit error rate in the current environment after being trained through five convolutional layers; Perform data mapping on the codebook with the lowest bit error rate; Perform compressive sensing on the data signal mapped to the codebook; The signal after compressive sensing first undergoes OFDM modulation, then signal transmission, followed by signal demodulation. After that, the signal is reconstructed and demapped into a codebook signal, and finally the codebook signal is decoded into the original signal through a decoder; The codebook design using the AlexNet model relies on CNN training to enable encoder f kj (r;θ f ), decoder g(y; θ) g The original data r is correctly reconstructed; the end-to-end loss function of the CNN, considering both the encoder and decoder, is: In the formula, h is the channel vector, r is the target value, f is the encoder, and g is the decoder. The data reconstructed for the CNN decoder, i.e. Performing compressive sensing on the data signal mapped to the codebook is to sample and compress the signal, and the steps include: S31, Suppose x is a one-dimensional discrete signal of length N, then use a set of orthogonal bases. The linear combination representation of Ψ is used to form an N×N dimensional basis matrix, i.e., Ψ={Ψ1,Ψ2,…,Ψ N }, then the signal x can be represented by this basis vector as: Among them, θ = [θ1, θ2,…θ N ] T θ represents the sparse representation coefficients of signal x under basis Ψ, and only a small portion of θ has large values, while most of the values ​​are zero or close to zero; if there are S non-zero elements, the signal x is said to be S sparse. S32, design a measurement matrix Φ of size M×N (M<<N), which is uncorrelated with the transform basis Ψ; the signal x can obtain a measurement vector of M×1 through the measurement of the measurement matrix Φ: y = Φx = ΦΨθ = Αθ The high-dimensional signal Ψθ is mapped to a low-dimensional space under the measurement of the observation matrix Φ, that is, the compressed signal y is obtained, and the sampling and compression of the signal are completed.

2. The coding method based on deep learning and compressed sensing assistance according to claim 1, characterized in that, The AlexNet model in CNN uses a convolutional operation method based on dual GPUs, and at the same time uses the Relu activation function to replace the T activation function. The Relu activation function is expressed as: f(x) = max(0, x).

3. The coding method based on deep learning and compressed sensing assistance according to claim 1, characterized in that, The said data mapping is to transmit 6 user data with 4 subcarriers. The mapping process is as follows: 6 users all transmit two symbols, each symbol occupies one subcarrier, and the 4 subcarriers are respectively represented by b1, b2, b3, and b4; map the first user data to b1, b2, the second user to b1, b3, the third user to b1, b4, the fourth user to b2, b3, the fifth user to b2, b4, and the last one to b3, b4.

4. The coding method based on deep learning and compressed sensing assistance according to claim 1, characterized in that, The steps to decode the codebook signal into the original signal are: S41, for each resource, calculate the probability of each possible codeword on this resource according to the received signal and channel gain, and pass these probabilities to the corresponding user nodes; S42, each user node updates the probability of the codeword it selects according to the probability information received from other resource nodes, and passes these probabilities back to the resource nodes; S43, repeat S41 - S42 until the predetermined number of iterations is reached; S44, after the iteration ends, calculate the likelihood ratio of each bit according to the final probability distribution of the user nodes, perform bit decision according to the likelihood ratio, and recover the original signal.

5. A coding system based on deep learning and compressed sensing, characterized in that, It includes: Codebook design module: Perform coding mapping on bit data, use the AlexNet model for codebook design, and each codebook obtains the bit error rate in the current environment after being trained through five convolutional layers; Data mapping module: Perform data mapping on the codebook with the lowest bit error rate; Compressive sensing module: Perform compressive sensing on the data signal mapped to the codebook; And, reconstruction and decoding module: The signal after compressive sensing first undergoes OFDM modulation, then signal transmission, followed by signal demodulation. After that, the signal is reconstructed and demapped into a codebook signal, and finally the codebook signal is decoded into the original signal through a decoder; The codebook design using the AlexNet model relies on CNN training to enable encoder f kj (r;θ f ), decoder g(y; θ) g The original data r is correctly reconstructed; the end-to-end loss function of the CNN, considering both the encoder and decoder, is: In the formula, h is the channel vector, r is the target value, f is the encoder, and g is the decoder. The data reconstructed for the CNN decoder, i.e. Performing compressive sensing on the data signal mapped to the codebook is to sample and compress the signal, and the steps include: S31, Suppose x is a one-dimensional discrete signal of length N, then use a set of orthogonal bases. The linear combination representation of Ψ is used to form an N×N dimensional basis matrix, i.e., Ψ={Ψ1,Ψ2,…,Ψ N }, then the signal x can be represented by this basis vector as: Among them, θ = [θ1, θ2,…θ N ] T θ represents the sparse representation coefficients of signal x under basis Ψ, and only a small portion of θ has large values, while most of the values ​​are zero or close to zero; if there are S non-zero elements, the signal x is said to be S sparse. S32. Design a measurement matrix Φ with a size of M×N (M << N), which is uncorrelated with the transform basis Ψ. The signal x can obtain a measurement vector of M×1 through the measurement of the measurement matrix Φ: y = Φx = ΦΨθ = Αθ The high-dimensional signal Ψθ is mapped to a low-dimensional space under the measurement of the observation matrix Φ, that is, the compressed signal y is obtained, and the sampling and compression of the signal are completed.

6. A computer storage medium storing a readable program, characterized in that, When the program runs, it can execute the encoding method based on deep learning and compressive sensing assistance described in any one of claims 1-4.

7. An electronic device, characterized in that, Including: A processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface complete mutual communication through the communication bus; The memory is used to store at least one executable instruction, and the executable instruction causes the processor to execute the operations corresponding to the encoding method based on deep learning and compressive sensing assistance described in any one of claims 1-4.

8. A computer program product comprising computer instructions, characterized in that, The computer instruction instructs the computing device to execute the operations corresponding to the encoding method based on deep learning and compressive sensing assistance described in any one of claims 1-4.

Images