A communication method based on deep learning and joint source-channel coding optimization

By using a deep learning-based joint source-channel coding optimization method, image features are extracted and combined with digital channel coding. This solves the compatibility and sensitivity issues of existing technologies, achieving more efficient image restoration and communication reliability, and avoiding the 'cliff effect'.

CN117939150BActive Publication Date: 2024-11-12THE CHINESE UNIV OF HONG KONG (SHENZHEN) +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410280036.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-12
Publication Date
2024-11-12
Estimated Expiration
2044-03-12

AI Technical Summary

Technical Problem

Existing deep learning joint source-channel coding techniques are incompatible with modern communication hardware and are sensitive to changes in channel statistics and machine tasks, failing to effectively link deep source coding with channel statistics.

Method used

A joint source-channel coding optimization method based on deep learning is adopted. Image data features are extracted by deep source encoder and combined with digital channel coding. Entropy coding and distribution learning modules are used to optimize bit stream transmission. A joint source-channel optimization algorithm is designed to minimize end-to-end distortion.

Benefits of technology

Achieving more efficient image restoration in modern communication hardware avoids the 'cliff effect' in traditional source-channel separation systems, thus improving the reliability and efficiency of communication.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117939150B_ABST
    Figure CN117939150B_ABST
Patent Text Reader

Abstract

The present invention relates to a communication method based on deep learning and joint source channel coding optimization, comprising: at the transmitting end, a deep source encoder encodes input image data x into a bit rate b, and a digital channel encoder divides the bit rate b into multiple data packets of equal length of N bits for transmission; at the receiving end, a digital channel decoder decodes the symbols in the received multiple data packets into a bit stream #imgabs0# and sends the bit stream #imgabs1# to the deep source decoder for data recovery to obtain the restored image data #imgabs2#, and in the overall data transmission between the transmitting end and the receiving end, the average end-to-end distortion is optimized to the minimum through a joint source channel optimization algorithm. The present invention utilizes a source encoder based on a deep neural network to extract and encode image semantic features, and combines digital channel coding technology to protect the coded bits from the influence of channel errors, and has a better image recovery effect while being compatible with modern communication hardware.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular to a communication method based on deep learning and joint source channel coding optimization. Background Art

[0002] With limited spectrum resources, the sixth generation of mobile networks has more stringent performance requirements than the fifth generation of mobile communications. Semantic communication has initiated a new paradigm that significantly reduces communication overhead by effectively extracting and transmitting semantic features of raw data; however, the complexity of the channel environment and the limitations of existing data hardware have brought challenges to the development of artificial intelligence transceivers for semantic communication.

[0003] A representative technology of semantic communication at present is the use of deep neural networks for joint source channel coding (DJSCC). This technology has shown higher transmission efficiency than traditional source channel separation coding. However, the current DJSCC technology is incompatible with modern communication hardware, is sensitive to changes in channel statistics and machine tasks, or does not consider the correlation between deep source coding and channel statistical information; therefore, how to solve this problem is currently needed to be considered. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a communication method based on deep learning and joint source-channel coding optimization to solve the deficiencies of the prior art.

[0005] The object of the present invention is achieved by the following technical solution: a communication method based on deep learning and joint source-channel coding optimization, the communication method comprising:

[0006] At the transmitting end, the deep source encoder encodes the input image data x into a bit rate b, and the digital channel encoder divides the bit rate b into multiple packets of equal length N bits for transmission;

[0007] At the receiving end, the digital channel decoder decodes the symbols in the received multiple data packets into a bit stream and the bitstream Send it to the deep source decoder for data recovery to get the restored image data

[0008] And in the overall data transmission between the transmitter and the receiver, the average end-to-end distortion is optimized to the minimum through a joint source channel optimization algorithm.

[0009] The deep source encoder encodes the input data x into a bit rate b, specifically including the following:

[0010] The M-dimensional image data Through the function F φExtract features, map them into K-dimensional continuous vector y, and then divide them into two directions for operation, where φ represents the function F φ Neural network parameters;

[0011] In the first direction, the feature vector y is quantized and the quantized vector is output

[0012] The second direction is to pass the feature vector y through a distribution learning module to output a bit sequence and the probability distribution function B z is the quantized feature vector The length of the bit sequence obtained by entropy coding;

[0013] Through the probability distribution function Vector Encode into a bit sequence using an entropy encoder B y is the quantized feature vector The length of the bit sequence obtained by entropy coding;

[0014] b y and b z Concatenate into a bit stream b∈{0,1} B Transmission is performed, B = B y +B z , B is the length of the bit sequence obtained after concatenation.

[0015] The bit stream is output through a distribution learning module and the probability distribution function Specifically include the following:

[0016] The input feature vector y is passed through a parameter θ 1 Neural Network Then output the feature vector z and obtain it through quantization

[0017] Will Enter a parameter θ 2 Neural Network The output vector u=[u 1 ,u 2 ,...,u K ],σ=[σ 1 ,σ 2 ,...,σ K ], and obtain the probability distribution function k i Representation vector The value of the i-th element in , represents all integer field vectors of dimension K, The mean value is u i , the variance is Gaussian probability density function of ;

[0018] At the same time, the entropy encoder according to Encoded into bit stream b z .

[0019] In digital channel coding, the channel rate R is used for the i-th data packet. c The universal (N,L) block code encodes the data bits into a symbol sequence of length L. The transmission symbol meets the unit power constraint, the total number of transmitted packets is T, and the bandwidth ratio is set to To measure the average channel usage of an element in the transmitted image data x, the input-output relationship of the point-to-point channel is represents the codebook of the transmitted symbol, M is the dimension of the image data x, s i =[s i,1 ,s i,2 ,…,s i,L ] represents the symbol in packet i, r i =[r i,1 ,r i,2 ,…,r i,L ] indicates the received symbol, p i Indicates the transmit power, n i =[n i,1 ,n i,2 ,…,n i,L ] has a mean of zero and a variance of δ 2 I is an independent and identically distributed circularly symmetric complex Gaussian noise, δ represents the variance of the noise signal caused by each symbol transmission, I represents the identity matrix, is the Rayleigh channel coefficient.

[0020] The bitstream Send it to the deep source decoder for data recovery to get the restored image data Specifically include the following:

[0021] The deep source decoder receives the bitstream Divided into and Two parts, Enter the entropy decoder according to Decoded as Then Input to the neural network get

[0022] At the same time, the entropy decoder uses and For the feature vector Decode and convert Input to the recovery function G with parameter ψ ψ The restored image data is obtained

[0023] The joint source channel optimization algorithm specifically includes the following contents:

[0024] The average end-to-end distortion is reduced by the joint source channel optimization algorithm under the constraint of the average channel usage times. Minimization, the optimization problem is expressed as:

[0025]

[0026]

[0027] Where N = {n i},n i =[n i,1 ,n i,2 ,…,n i,L ] has a mean of zero and a variance of δ 2 I is independent and identically distributed circularly symmetric complex Gaussian noise, represents the mathematical expectation of variables x and N, d represents the average number of times a channel is used to transmit a picture, which is used to characterize the average delay of transmitting pictures, L is the data bit encoding length, R c is the channel rate, B is the length of the bit sequence obtained after concatenation, Ф is the neural network parameter of the deep source encoder, and ψ is the recovery function G ψ Parameters;

[0028] Using inequality And set the average length of the deep source encoder output bit sequence to It means to find the mathematical expectation for x, and transform the optimization problem into:

[0029]

[0030]

[0031] and The estimation and optimization problem of

[0032] Said The estimates specifically include the following:

[0033] Assume that the recovery function G in the deep source decoder is ψ exist Upper Lipschitz continuity, that is, there exists C ψ >0 and in is the set of all possible K-dimensional data y;

[0034] Then there is satisfy in represents the recovered data under error-free transmission, ρ is the packet error probability of the polar code, and D s represents the distortion caused by deep source coding, It means that under the condition of known channel error, when the error probability of polar code group is ρ and the parameter of deep source encoder is Φ, The maximum variance of all elements in ;

[0035] It represents the distortion introduced by channel transmission and is recorded as set up in Then we get

[0036] When the neural network parameters {Φ,ψ} in the deep source decoder are fixed, we get represents a constant related to the neural network parameter ψ of the deep source decoder, which is calculated as When the neural network parameters {Φ,ψ} are fixed, C ψ and are constants. For simplicity, we can express the part related to these two constants as one term, so we introduce

[0037] The solution to the optimization problem specifically includes the following contents:

[0038] The optimization problem is transformed into:

[0039]

[0040]

[0041] The optimization problem is solved by making a lookup table and jointly model selection and rate control.

[0042] The preparation of the lookup table specifically includes the following contents:

[0043] A set of neural network parameters {Φ, ψ} are trained under error-free transmission to balance the source rate R by minimizing the rate-distortion functions and source distortion D s ,Right now Among them, λ>0 represents a hyperparameter;

[0044] By changing λ, a set of neural network models related to different source rates and distortion levels are obtained. The validation dataset is used to estimate the R s and D s , to construct different source rates, source distortion levels, hyperparameters λ and A lookup table of P models.

[0045] The joint model selection and rate control specifically include the following contents:

[0046] When the parameters {Φ,ψ} are fixed, the optimization problem Simplified to

[0047] When ρ is small When Simplified to

[0048] When the channel rate is greater than the channel capacity, When , due to the high error probability, it is impossible to construct a block code to achieve communication. Therefore, the optimal source coding neural network model must ensure according to The polar code with the packet error probability given by The optimal solution is Among them, the parameter β 1 and β 2 It depends on the signal-to-noise ratio γ and the data bit coding length L;

[0049] Therefore, when the parameters of the source coding neural network model are fixed, the minimum distortion is solved For different source coding neural network models, an exhaustive search algorithm is used to find The optimal source coding neural network model with minimal end-to-end distortion.

[0050] The present invention has the following advantages: a communication method based on deep learning and joint source-channel coding optimization, using a deep source encoder based on a deep neural network to extract and encode image semantic features, and combining digital channel coding technology to protect the coded bits from channel errors, and having better image restoration effects while being compatible with modern communication hardware. In addition, in a digital communication system with traditional source-channel separation, when digital channel decoding goes wrong due to poor channel conditions, serious errors will occur in source decoding, a phenomenon known as the "cliff effect". Because the present invention jointly optimizes source coding and channel coding settings according to channel conditions, this method can avoid the "cliff effect" compared to traditional digital communication technology with source-channel separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic diagram of the present invention;

[0052] Figure 2 Schematic diagram of a deep source encoder;

[0053] Figure 3 Schematic diagram of the deep source decoder. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0055] like Figure 1 As shown, the present invention specifically relates to a communication method based on deep learning and joint source channel coding optimization, which combines the image low-dimensional semantic feature extraction technology based on deep neural network and the digital channel coding technology, and designs a joint optimization algorithm of the two to minimize the end-to-end distortion of image transmission; it includes two parts: a semantic communication system based on deep neural network and an optimization algorithm for joint source channel coding.

[0056] (1) Semantic communication system based on deep neural network;

[0057] like Figure 2As shown, at the transmitter, the input image data x is encoded into a bit stream b by the deep source encoder. Then, b is divided into multiple data packets of equal length N bits for transmission. For the i-th data packet, the channel rate is R c The universal (N,L) block code encodes the data bits into a symbol sequence of length L. The transmission symbol satisfies the unit power constraint, that is, The total number of packets transmitted is in represents the rounding up operation. The bandwidth ratio is defined as It is used to measure the average channel usage of transmitting an element in x.

[0058] The input-output relationship of a point-to-point channel can be expressed as where s i =[s i,1 ,s i,2 ,…,s i,L ] T represents the symbol in packet i, r i =[r i,1 ,r i,2 ,…,r i,L ] T represents the received symbol, p i Indicates the transmit power, n i =[n i,1 ,n i,2 ,…,n i,L ] T is zero and has a variance of δ 2 I is the independent and identically distributed circularly symmetric complex Gaussian (CSCG) noise, is the Rayleigh channel coefficient. It is assumed that all channel coefficients remain unchanged in each transmission block, but may change in different transmission blocks. It is assumed that the channel coefficients and noise variance are completely known at the transmitter and receiver. In order to overcome fading, channel inversion power control is adopted, that is, Thus, the block fading channel can be converted into an AWGN channel, where the received signal-to-noise ratio is Each block is identical.

[0059] At the receiving end, the digital channel decoder decodes the symbols in the received T packets into a bit stream Then, the bitstream Send it to the deep source decoder for data recovery Then the average end-to-end distortion of the system can be defined as:

[0060]

[0061] Where N = {n i},ni =[n i,1 ,n i,2 ,…,n i,L ] has a mean of zero, It means to find the mathematical expectation for variables x and N.

[0062] Furthermore, the M-dimensional image data First, through the function F φ Extract features and map them into a K-dimensional continuous vector y. Among them, the function F φ is a function composed of a neural network, and the neural network parameter is φ. After obtaining the feature vector y, there are two steps: the first step is to quantize the feature vector y and output it as a quantized vector The second step is to pass the feature vector y through a distribution learning module to output a bit sequence and the probability distribution function

[0063] The design of the distribution learning module is as follows: first, the input feature vector y is passed through a parameter θ 1 Neural Network The output is the feature vector z. Then, the feature vector z is quantized to obtain Then, The input parameter is θ 2 Neural Network The output is a vector u=[u 1 ,u 2 ,...,u K ],σ=[σ 1 ,σ 2 ,...,σ K ]. The final probability distribution function It can be calculated as:

[0064]

[0065] in, k i Representation vector The value of the i-th element in , represents all integer field vectors of dimension K, The mean value is u i , the variance is The Gaussian probability density function, whose variable is y i .

[0066] Get the feature vector and the probability distribution function back. The entropy encoder is based on its PMF Encoded into bit stream bz Similarly, using the probability distribution function Vector Encode into a bit sequence using an entropy encoder Finally, b y and b z Concatenate into a bit stream b∈{0,1} B Transmission is performed, B = B y +B z , B is the length of the bit sequence obtained after concatenation, B y is the quantized feature vector The length of the bit sequence obtained by entropy coding, B z is the quantized feature vector The length of the bit sequence obtained by entropy coding.

[0067] Among them, the deep source encoder can use the neural network parameters Φ={φ,θ 1 ,θ 2} to specify.

[0068] like Figure 3 As shown, the deep source decoder receives the bit stream Divided into and Two parts. First, Sent to the entropy decoder, according to decoding Then, Enter the neural network get Entropy decoder uses and For the feature vector Finally, Input to the parameter ψ recovery function G ψ , restore image data

[0069] Consider an arbitrary (N,L) block code, such as a BPSK-modulated polar code, with a channel encoder and a channel decoder Composition, of which L and N represent the codebook of transmitted symbols, the set of received symbols, the block length, and the data bit length, respectively. The channel rate can be expressed as In an additive white Gaussian noise (AWGN) channel with a signal-to-noise ratio of γ, for practical channel coding and modulation (e.g., polar coding and QPSK modulation), the average block error probability can be approximated as The parameter β1 and β 2 Depends on γ and L and can be estimated.

[0070] (2) Optimization algorithm for joint source-channel coding;

[0071] The present invention designs an optimization algorithm that jointly considers the source channel coding parameters to minimize the average end-to-end distortion under the constraint of the average number of channel usage. The optimization problem can be expressed as:

[0072]

[0073]

[0074] Where d represents the average number of times a channel is used to transmit a picture, which is used to characterize the average delay of transmitting pictures, and d>0. Using the inequality And the average length of the bit sequence output by the deep source encoder is recorded as R s ,Right now It means that the mathematical expectation for x is obtained, and the optimization problem can be transformed into:

[0075]

[0076]

[0077] in, The optimization problem after conversion is solved by The solution is carried out in two parts: estimation and optimization problem.

[0078] Further, The estimates specifically include the following:

[0079] By observation, it is found that the eigenvector y is sparse and its elements can be estimated as independent Gaussian random variables, that is, in Diag(v) represents a diagonal matrix with the elements in vector v as diagonal elements, where v is an arbitrary vector. Assume that the recovery function G in the deep source decoder is ψ exist Upper Lipschitz continuity, that is, there exists C ψ >0 and

[0080] It can be proved that under known channel error conditions, when the polar code group error probability is ρ and the deep source encoder parameter is Φ, The maximum variance of all elements in satisfy Indicates the recovered data under error-free transmission. The average number of packets into which the bit sequence output by the source encoder is divided.

[0081] Where D s represents the distortion caused by deep source coding, It represents the distortion introduced by channel transmission and is recorded as definition in but When the neural network parameters {Φ,ψ} in the deep source codec are fixed, the constant can be estimated

[0082] Furthermore, the solution to the optimization problem specifically includes the following contents:

[0083] Using the above The optimization problem to be solved is further transformed into:

[0084]

[0085]

[0086] The optimization problem is divided into two parts: making the lookup table and joint model selection and rate control.

[0087] Furthermore, making a lookup table specifically includes the following contents:

[0088] First, a set of neural network parameters {Φ, ψ} are transmitted without error (i.e. The core idea is to balance the source rate R by minimizing the rate distortion function. s and source distortion D s ,Right now

[0089] Where λ>0 is a hyperparameter. By changing λ, a set of neural network models associated with different source rates and distortion levels can be obtained. For each neural network model, the parameters can be estimated using the validation dataset R s and D s Combined with the above operations, construct different source rates, source distortion levels, hyperparameters λ and parameters A lookup table of P models.

[0090] Furthermore, the joint model selection and rate control specifically include the following:

[0091] Based on the previous lookup table and specific block codes, when the parameters {Φ,ψ} are fixed, can be simplified to

[0092] When ρ is very small, Convert to

[0093] It is observed that when the channel rate is greater than the channel capacity, that is, When , the error probability is very high, and it is impossible to construct a block code to achieve reliable communication. Therefore, the optimal source coding neural network model must ensure It can be proved that for The polar code with the packet error probability given by The optimal solution R c for:

[0094]

[0095] Therefore, when the parameters of the source coding neural network model are fixed, the minimum distortion can be solved. For different source coding neural network models, an exhaustive search algorithm can be used to find The optimal source coding neural network model with minimal end-to-end distortion.

[0096] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A communication method based on deep learning and joint source-channel coding optimization, Features: The communication method comprises: At the transmitting end, the deep source encoder encodes the input image data x into a bit rate b, and the digital channel encoder divides the bit rate b into multiple packets of equal length N bits for transmission; At the receiving end, the digital channel decoder decodes the symbols in the received multiple data packets into a bit stream and the bitstream Send it to the deep source decoder for data recovery to get the restored image data And in the overall data transmission between the transmitter and the receiver, the average end-to-end distortion is optimized to the minimum through the joint source channel optimization algorithm; The average end-to-end distortion is minimized by the joint source channel optimization algorithm under the constraint of the average channel usage times. Minimize, the optimization problem is expressed as: Where N = {n i },n i =[n i,1 ,n i,2 ,…,n i,L ] has a mean of zero and a variance of δ 2 I is independent and identically distributed circularly symmetric complex Gaussian noise, represents the mathematical expectation of variables x and N, d represents the average number of times a channel is used to transmit a picture, which is used to characterize the average delay of transmitting pictures, L is the data bit encoding length, R c is the channel rate, B is the length of the bit sequence obtained after concatenation, Ф is the neural network parameter of the deep source encoder, and ψ is the recovery function G of the deep source decoder. ψ Parameters, M is the dimension of image data; Using inequality And set the average length of the deep source encoder output bit sequence to It means to find the mathematical expectation for x, and transform the optimization problem into: and The estimation and optimization problem of 2. A communication method based on deep learning and joint source-channel coding optimization according to claim 1, Features: The deep source encoder encodes the input data x into a bit rate b, specifically including the following: The M-dimensional image data Through the function F φ Extract features, map them into K-dimensional continuous vector y, and then divide them into two directions for operation, where φ represents the function F φ Neural network parameters; In the first direction, the feature vector y is quantized and the quantized vector is output The second direction is to pass the feature vector y through a distribution learning module to output a bit sequence and the probability distribution function B z is the quantized feature vector The length of the bit sequence obtained by entropy coding; Through the probability distribution function Vector Encode into a bit sequence using an entropy encoder B y is the quantized feature vector The length of the bit sequence obtained by entropy coding; b y and b z Concatenate into a bit stream b∈{0,1} B Transmission is performed, B = B y +B z , B is the length of the bit sequence obtained after concatenation.

3. A communication method based on deep learning and joint source channel coding optimization according to claim 2, Features: The bit stream is output through a distribution learning module and the probability distribution function Specifically include the following: The input feature vector y is passed through a parameter θ 1 Neural Network Then output the feature vector z and obtain it through quantization Will Enter a parameter θ 2 Neural Network The output vector u=[u 1 ,u 2 ,...,u K ],σ=[σ 1 ,σ 2 ,...,σ K ], and obtain the probability distribution function k i Representation vector The value of the i-th element in , represents all integer field vectors of dimension K, The mean value is u i , the variance is The Gaussian probability density function, y i Indicates its variables; At the same time, the entropy encoder According to the probability mass function Encoded into bit stream b z .

4. The communication method based on deep learning and joint source-channel coding optimization according to claim 1, Features: In digital channel coding, the channel rate R is used for the i-th data packet. c The universal (N,L) block code encodes the data bits into a symbol sequence of length L. The transmission symbol meets the unit power constraint, the total number of transmitted packets is T, and the bandwidth ratio is set to To measure the average channel usage of an element in the transmitted image data x, the input-output relationship of the point-to-point channel is represents the codebook of the transmitted symbol, M is the dimension of the image data x, s i =[s i,1 ,s i,2 ,…,s i,L ] represents the symbol in packet i, represents the received symbol, p i Indicates the transmit power, n i =[n i,1 ,n i,2 ,…,n i,L ] has a mean of zero and a variance of δ 2 I is an independent and identically distributed circularly symmetric complex Gaussian noise, δ represents the variance of the noise signal caused by each symbol transmission, I represents the identity matrix, is the Rayleigh channel coefficient.

5. The communication method based on deep learning and joint source-channel coding optimization according to claim 1, Features: The bitstream Send it to the deep source decoder for data recovery to get the restored image data Specifically include the following: The deep source decoder receives the bitstream Divided into and Two parts, Enter the entropy decoder according to Decoded as Then Input to the neural network get At the same time, the entropy decoder uses and For the feature vector Decode and convert Input to the recovery function G with parameter ψ ψ The restored image data is obtained 6. The communication method based on deep learning and joint source-channel coding optimization according to claim 1, Features: Said The estimates specifically include the following: Assume that the recovery function G in the deep source decoder is ψ exist Upper Lipschitz continuity, that is, there exists C ψ >0 and in is the set of all possible K-dimensional data y; Then there is satisfy in represents the recovered data under error-free transmission, ρ is the packet error probability of the polar code, and D s represents the distortion caused by deep source coding, It means that under the condition of known channel error, when the error probability of polar code group is ρ and the parameter of deep source encoder is Φ, The maximum variance of all elements in ; It represents the distortion introduced by channel transmission and is recorded as set up in Then we get When the neural network parameters {Φ,ψ} in the deep source decoder are fixed, we get represents a constant related to the neural network parameters ψ of the deep source decoder.

7. The communication method based on deep learning and joint source-channel coding optimization according to claim 6, Features: The solution to the optimization problem specifically includes the following contents: The optimization problem is transformed into: The optimization problem is solved by making a lookup table and jointly model selection and rate control.

8. A communication method based on deep learning and joint source-channel coding optimization according to claim 7, Features: The preparation of the lookup table specifically includes the following contents: A set of neural network parameters {Φ, ψ} are trained under error-free transmission to balance the source rate R by minimizing the rate-distortion function s and source distortion D s ,Right now Among them, λ>0 represents a hyperparameter; By changing λ, a set of neural network models related to different source rates and distortion levels are obtained. The validation dataset is used to estimate the R s and D s , to construct different source rates, source distortion levels, hyperparameters λ and A lookup table of P models.

9. The communication method based on deep learning and joint source-channel coding optimization according to claim 7, Features: The joint model selection and rate control specifically include the following contents: When the parameters {Φ,ψ} are fixed, the optimization problem Simplified to When ρ is small When Simplified to When the channel rate is greater than the channel capacity, When , due to the high error probability, it is impossible to construct a block code to achieve communication. Therefore, the optimal source coding neural network model must ensure according to The polar code with the packet error probability given by The optimal solution is Among them, the parameter β 1 and β 2 It depends on the signal-to-noise ratio γ and the data bit coding length L; Therefore, when the parameters of the source coding neural network model are fixed, the minimum distortion is solved For different source coding neural network models, an exhaustive search algorithm is used to find The optimal source coding neural network model with minimal end-to-end distortion.

Citation Information

Cited By

  • Joint source channel coding method and system based on deep compressed sensing

    CN120434389A

  • A semantic communication method, system and computer readable storage medium based on multi-description coding

    CN122932899A