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

Abstract

The application discloses a semantic communication method and system based on multi-description coding and a computer readable storage medium, and solves the problems that the existing semantic communication scheme cannot reconstruct source data or extract semantic information when channel packets are lost, and lacks multi-channel description joint semantic extraction capability. The method encodes source data sequences by first and second encoders to generate first and second description code words, and respectively decodes to obtain reconstructed sequences. A joint semantic decoder receives two description code words, jointly decodes, and extracts hidden semantic information. The code rate of the two code words satisfies the mutual information reachable rate distortion constraint, and the distortion of the three reconstructed sequences satisfies the corresponding constraint. The application considers both system robustness and semantic extraction capability.

Description

Technical Field

[0001] This application relates to the fields of semantic communication and source coding technology, and in particular to a semantic communication method, system and computer-readable storage medium based on multi-description coding. Background Technology

[0002] With the rapid development of communication technology, channel capacity is gradually approaching the Shannon limit, while emerging intelligent applications such as the Internet of Vehicles and smart grids are placing higher demands on the latency and quality of service of communication systems. Semantic communication, as a task-oriented communication paradigm, can achieve efficient communication with limited channel resources by transmitting only key semantic information useful to the receiving end's task, and has received widespread attention in recent years.

[0003] Existing deep learning-based semantic communication schemes, such as the joint source-channel coding method based on autoencoders (see Chinese patent application CN202410280036.3) and the information bottleneck-based semantic communication method (see Chinese patent application CN202310128208.0), all employ an end-to-end architecture with a single encoder and a single decoder, compressing the source data into a single description before transmission through the channel. When packet loss occurs in the channel, the receiving end is completely unable to reconstruct the source data or extract semantic information, resulting in insufficient system robustness. Furthermore, existing schemes do not consider jointly extracting hidden semantic information from the source data when receiving multiple descriptions, lacking both theoretical support and engineering implementation for semantic extraction capabilities in multi-description coding scenarios. Therefore, how to simultaneously ensure the quality of source data reconstruction and the extraction of hidden semantic information within a multi-description coding framework is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a semantic communication method and system based on multi-description coding to solve the following technical problems existing in the prior art: Existing semantic communication schemes all adopt an end-to-end architecture of a single encoder and a single decoder, which compresses the source data into a single description and transmits it through the channel. When the channel experiences packet loss, the receiving end will be completely unable to reconstruct the source data or extract semantic information, resulting in insufficient system robustness. Furthermore, existing schemes lack the ability to jointly extract hidden semantic information in the source data when receiving multiple descriptions, and the semantic reachability distortion region in the multi-description coding scenario also lacks theoretical support.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a semantic communication method based on multiple description coding, comprising the following steps: acquiring a source data sequence; encoding the source data sequence using a first encoder to generate a first description codeword; encoding the source data sequence using a second encoder to generate a second description codeword; decoding the first description codeword using a first decoder to obtain a first reconstructed sequence; decoding the second description codeword using a second decoder to obtain a second reconstructed sequence; and simultaneously receiving the first description codeword and the second description codeword using a joint semantic decoder, performing joint decoding on both to extract hidden semantic information contained in the source data sequence, thereby obtaining a semantic reconstructed sequence. The code rates of the first and second description codewords satisfy the following constraints: the first description code rate is greater than the mutual information between the source data sequence and the first reconstructed sequence; the second description code rate is greater than the mutual information between the source data sequence and the second reconstructed sequence; and the sum of the first and second description code rates is greater than the sum of the joint mutual information between the source data sequence and the first, second, and semantic reconstructed sequences, and the mutual information between the first and second reconstructed sequences. Simultaneously, the distortion of the first reconstructed sequence does not exceed a first distortion constraint, the distortion of the second reconstructed sequence does not exceed a second distortion constraint, and the distortion of the semantic reconstructed sequence does not exceed a semantic distortion constraint. Through the above architecture of parallel encoding by dual encoders and division of labor decoding by three decoders, the system can independently reconstruct source data even when only a single description is received. When both descriptions are available, it can further jointly extract hidden semantic information, achieving a balance between robustness and semantic extraction capability.

[0006] Furthermore, both the first encoder and the second encoder are implemented using an autoencoder structure. The autoencoder structure includes an encoding subnetwork and a decoding subnetwork. The encoding subnetwork compresses the source data sequence into a hidden layer feature representation. The decoding subnetwork is signal-connected to the encoding subnetwork and is used to reconstruct the corresponding reconstructed sequence based on the hidden layer feature representation. Employing an autoencoder structure allows the encoder to adaptively learn the optimal compressed representation in an end-to-end trainable manner, eliminating the need for manually designed feature extraction rules.

[0007] Furthermore, the method further includes: receiving the source data sequence and the first reconstructed sequence through a first discriminator, and estimating the mutual information between the source data sequence and the first reconstructed sequence; receiving the source data sequence and the second reconstructed sequence through a second discriminator, and estimating the mutual information between the source data sequence and the second reconstructed sequence; and feeding back the estimation results of the first discriminator and the second discriminator to the first encoder and the second encoder, respectively, to perform adversarial training updates on the network parameters of the first encoder and the second encoder. By introducing a discriminator to perform variational estimation of mutual information, it is possible to achieve efficient data-driven estimation of mutual information without explicitly obtaining the probability distribution function.

[0008] Specifically, both the first and second discriminators are implemented using deep neural networks. The optimal discriminator function is obtained by optimizing the log-likelihood ratio of the joint distribution samples and the product samples of the marginal distributions, thereby estimating the corresponding mutual information. This variational estimation method transforms the mutual information estimation problem into a binary classification problem, which can be efficiently solved using stochastic gradient descent, reducing the difficulty of model training.

[0009] Furthermore, when jointly training the first encoder, the second encoder, and the joint semantic decoder, a Lagrange functional is used as the training objective function. This Lagrange functional includes: a mutual information term corresponding to the first description code rate weighted by a first compression ratio weight coefficient; a mutual information term corresponding to the second description code rate weighted by a second compression ratio weight coefficient; a distortion term of the first reconstructed sequence weighted by a first distortion weight coefficient; a distortion term of the second reconstructed sequence weighted by a second distortion weight coefficient; and a distortion term of the semantic reconstructed sequence weighted by a semantic distortion weight coefficient. By adjusting the first compression ratio weight coefficient, the second compression ratio weight coefficient, and the distortion weight coefficient, the trade-off between compression ratio and reconstruction distortion is controlled. This Lagrange functional design directly corresponds to the reachability distortion region theory of semantic multi-description coding, ensuring the theoretical interpretability of the model training objective and the controllability of the trade-offs.

[0010] In some embodiments, the distortion of the first reconstructed sequence is measured by mean squared error, the distortion of the second reconstructed sequence is measured by mean squared error, and the distortion of the semantic reconstructed sequence is measured by cross-entropy loss. Mean squared error is suitable for evaluating the reconstruction quality of continuous source data, while cross-entropy loss is suitable for evaluating the extraction accuracy of discrete semantic labels. The combined use of these two metrics enables the system to simultaneously optimize both source data reconstruction and semantic information extraction objectives.

[0011] This invention also provides a semantic communication system based on multi-description coding, including an encoding device and a decoding device. The encoding device includes: a first encoder for encoding an input source data sequence to generate a first description codeword; and a second encoder, configured in parallel with the first encoder, for encoding the source data sequence to generate a second description codeword. The first encoder and the second encoder are respectively connected to a channel signal for transmitting the first description codeword and the second description codeword to the decoding device via the channel. The decoding device includes: a first decoder connected to the channel signal for receiving the first description codeword and decoding it to obtain a first reconstructed sequence; a second decoder connected to the channel signal for receiving the second description codeword and decoding it to obtain a second reconstructed sequence; and a joint semantic decoder connected to the channel signal for simultaneously receiving the first description codeword and the second description codeword, performing joint decoding on both, and extracting hidden semantic information contained in the source data sequence to obtain a semantic reconstructed sequence.

[0012] Furthermore, both the first encoder and the second encoder are implemented using an autoencoder network structure, which includes an encoding subnetwork and a decoding subnetwork. The encoding subnetwork is used to compress the source data sequence into a hidden layer feature representation, and the decoding subnetwork is signal-connected to the encoding subnetwork and is used to reconstruct the corresponding reconstructed sequence based on the hidden layer feature representation.

[0013] Furthermore, the system further includes: a first discriminator, signal-connected to the first encoder, for receiving the source data sequence and the first reconstructed sequence, estimating the mutual information between them through a deep neural network, and feeding back the estimation result to the first encoder; and a second discriminator, signal-connected to the second encoder, for receiving the source data sequence and the second reconstructed sequence, estimating the mutual information between them through a deep neural network, and feeding back the estimation result to the second encoder.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the semantic communication method based on multiple description coding described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing dual encoders to encode the source data sequence in parallel, generating independent first and second description codewords respectively, even if a single packet is lost in the channel, the receiver can still independently reconstruct the source data using the other description codeword. This fundamentally overcomes the defect of the existing single encoder architecture that the receiver cannot recover the data at all when a single transmission fails, and significantly improves the robustness of the system under unreliable channel conditions.

[0016] 2. By setting up a joint semantic decoder to simultaneously receive the first description codeword and the second description codeword for joint decoding, the hidden semantic information contained in the source data sequence can be jointly extracted when both descriptions are successfully received, and a semantic reconstruction sequence is obtained. This realizes the effective mining of semantic information under the multi-description coding framework and makes up for the shortcomings of existing schemes that cannot perform joint semantic extraction when multiple descriptions are available at the same time.

[0017] 3. From the perspective of information theory, the reachability distortion region of semantic multi-description coding is derived, and the theoretical relationship between the first description code rate, the second description code rate and their sum and the three-way distortion constraint is clearly given. This provides a rigorous theoretical basis for the design of deep learning models and ensures the controllability and interpretability of the trade-off between compression rate and three-way reconstruction distortion.

[0018] 4. By introducing a first discriminator and a second discriminator to perform variational estimation of mutual information, the mutual information estimation problem is transformed into an optimization problem of the log-likelihood ratio of the product of the joint distribution samples and the marginal distribution samples. Without explicitly obtaining the probability distribution function, data-driven efficient estimation of mutual information is achieved, reducing the difficulty of model training and enabling the system to complete end-to-end training under the condition of only having data samples. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of a semantic communication system based on multi-description coding.

[0020] Figure 2 This is a schematic diagram of the achievable code rate for semantic multi-description coding under a given distortion constraint.

[0021] Figure 3 This is a schematic diagram of a semantic multi-description encoding network structure based on deep learning.

[0022] Figure 4 This is a schematic diagram showing the experimental results of mutual information and classification accuracy under different compression ratio weight coefficients and distortion weight coefficients. Detailed Implementation

[0023] The invention will now be described in further detail with reference to the accompanying drawings. This invention provides a semantic communication method and system based on multi-description coding. Its core improvement lies in the introduction of a dual-encoder parallel coding mechanism and a three-decoder collaborative decoding mechanism on the basis of the traditional single-encoder-single-decoder semantic communication architecture. The first encoder and the second encoder respectively generate independent first and second description codewords for the source data sequence. A joint semantic decoder, when both descriptions are available, jointly extracts the hidden semantic information contained in the source data. This solves the technical problem in existing single-channel transmission schemes where the receiver is completely unable to reconstruct the source data or extract semantic information when packet loss occurs in the channel.

[0024] Example 1 This embodiment provides a semantic communication system based on multiple description coding, combined with Figure 1 and Figure 3 Please provide an explanation.

[0025] like Figure 1 As shown, the system includes an encoding device and a decoding device. The encoding device includes a first encoder 101 and a second encoder 102, which are arranged in parallel and both receive the same source data sequence X as input. The decoding device includes a first decoder 103, a second decoder 104, and a joint semantic decoder 105. The first encoder 101 and the second encoder 102 transmit their encoded outputs to the decoding device via a channel.

[0026] The first encoder 101 encodes the input source data sequence X to generate a first description codeword M1, which is then transmitted to the decoding device via a first channel path. The second encoder 102 encodes the same source data sequence X to generate a second description codeword M2, which is then transmitted to the decoding device via a second channel path. The two description codewords are independent of each other, and the loss of either one does not affect the normal reception and decoding of the other.

[0027] The first decoder 103 receives only the first description codeword M1, decodes it, and obtains the first reconstructed sequence. 1; The second decoder 104 receives only the second description codeword M2, decodes it, and obtains the second reconstructed sequence. 2. When a single packet is lost in the channel, the receiver can still independently reconstruct the source data using another descriptor codeword.

[0028] The joint semantic decoder 105 simultaneously receives the first description codeword M1 and the second description codeword M2, performs joint decoding on them, extracts the hidden semantic information contained in the source data sequence, and obtains the semantic reconstruction sequence. The hidden semantic information S is associated with the source data X through a joint probability distribution p(s,x). For example, when X is a digital image, S is the digit category information contained in the image.

[0029] like Figure 3 As shown, both the first encoder 101 and the second encoder 102 are implemented using an autoencoder network structure. Specifically, the first encoder 101 corresponds to the autoencoder network E. π The parameter is π; the second encoder 102 corresponds to the autoencoder network G. φ The parameter is φ. Each autoencoder network includes an encoder subnetwork and a decoder subnetwork. The encoder subnetwork compresses the source data sequence X into a hidden layer feature representation, and the decoder subnetwork reconstructs the corresponding reconstructed sequence based on the hidden layer feature representation. The joint semantic decoder 105 corresponds to the decoder network K. μ The parameter is μ, and after receiving the first description codeword M1 and the second description codeword M2, it outputs a semantic reconstruction sequence. .

[0030] Furthermore, this embodiment also includes a first discriminator 106 and a second discriminator 107, used to perform variational estimation of mutual information during the training phase. The first discriminator 106 is implemented using a deep neural network with parameters ν1, and receives the source data sequence X and the first reconstructed sequence. 1. By analyzing the joint distribution sample P(X, 1) and the product of the marginal distribution samples P(X)P( 1) The log-likelihood ratio is optimized to obtain the optimal discriminator function F1*, in order to estimate the mutual information I(X; 1).

[0031] Specifically, the first discriminator 106 obtains the optimal discriminator function by solving the following optimization problem: Equation (1): Where σ(·) is the sigmoid function. The second discriminator 107 has the same structure as the first discriminator 106, with parameter ν2, and is used to estimate the mutual information I(X; 2). The estimation results of the two discriminators are fed back to the first encoder 101 and the second encoder 102 respectively, and the network parameters of the two are updated through adversarial training.

[0032] When jointly training the first encoder 101, the second encoder 102, and the joint semantic decoder 105, a Lagrange functional is used as the training objective function, specifically in the form of: Equation (2): in, Given a source data sequence X, the first reconstructed sequence is... Second reconstruction sequence and semantic reconstruction sequence The joint conditional probability distribution; This is the first compression ratio weighting coefficient. This is the second compression ratio weighting coefficient. , , These are the first distortion weight coefficient, the second distortion weight coefficient, and the semantic distortion weight coefficient, respectively. By adjusting these weight coefficients, the trade-off between compression ratio and reconstruction distortion can be controlled.

[0033] First Reconstruction Sequence The distortion of sequence 1 is measured by mean square error, and the second reconstructed sequence... The distortion of 2 is measured by mean squared error, and the semantic reconstruction sequence The distortion is measured using cross-entropy loss. Mean squared error is suitable for assessing the reconstruction quality of continuous source data, while cross-entropy loss is suitable for assessing the extraction accuracy of discrete semantic labels.

[0034] Through the above architecture of parallel encoding by dual encoders and division of labor decoding by three decoders, the system can independently reconstruct source data when only a single description is received, and can further jointly extract hidden semantic information when both descriptions are available, thus achieving a balance between robustness and semantic extraction capability.

[0035] Example 2 This embodiment describes the reachability distortion region of semantic multi-description coding, combined with... Figure 2 Describe it.

[0036] Let S and X be a pair of memoryless random variables, linked by a joint probability distribution p(s,x), forming a semantic information source. For an independent and identically distributed sequence S of length n... n and X n The first encoder 101 and the second encoder 102 respectively transmit X n The mapping is described by codewords M1 and M2, with code rates R1 and R2 respectively.

[0037] Given the distortion constraint D=(D0,D1,D2), the achievable code rate of semantic multi-description coding for regions is as follows: Figure 2 As shown. According to Theorem 1, the reachable code rate pair (R1, R2) satisfies the following constraints: The first description code rate R1 is greater than the source data sequence X and the first reconstructed sequence. Mutual information I(X;) between 1 1) The second description code rate R2 is greater than the source data sequence X and the second reconstructed sequence. Mutual information I(X;) between 2 2); The sum of the first description code rate and the second description code rate, R1+R2, is greater than the joint mutual information I(X); 1, 2, )and 1 and Mutual information I between 2 1; 2) and.

[0038] Meanwhile, the above bitrate pair must satisfy three-way distortion constraints: the first reconstructed sequence The distortion of sequence 1 does not exceed the first distortion constraint D1, and the second reconstructed sequence... The distortion of 2 does not exceed the second distortion constraint D2, and the semantic reconstruction sequence The distortion does not exceed the semantic distortion constraint D0.

[0039] like Figure 2 As shown, Figure 2 (a) represents the region of reachable code rate pairs under the original closure. Figure 2 (b) To simplify the reachable bitrate pair region under the closure. The simplified closure replaces the first encoder and its corresponding decoding subnetwork, and the second encoder and its corresponding decoding subnetwork, respectively, so that the boundary of the corresponding rate-distortion region is simplified from the diagonal corner of the original closure to the right-angle boundary of the simplified closure, which facilitates the implementation of deep learning methods and the design of training objectives.

[0040] The derivation of the reachability distortion region mentioned above provides a rigorous theoretical basis for the design of deep learning models, ensuring the controllability and interpretability of the trade-off between compression ratio and three-way reconstruction distortion.

[0041] Example 3 This embodiment experimentally verifies the system described in Embodiment 1 on the MNIST dataset, combining... Figure 4 Please provide an explanation.

[0042] The experimental setup is as follows: Source data X consists of MNIST handwritten digit images, and hidden semantic information S represents the digit category (0 to 9) in the images. Mean squared error is used for both source data reconstruction distortion metrics d1 and d2, while cross-entropy loss is used for semantic extraction distortion metric d0. Classification accuracy is used to explicitly evaluate semantic information extraction performance. In the experiment, β1 = β2 = β3 = β, meaning the three distortion weight coefficients are equal.

[0043] like Figure 4 (a) and Figure 4 As shown in (b), when β is small (e.g., β < 10000), the mutual information I(X; 1) and I(X; 2) It decreases as α1 and α2 increase. This is because the larger the weights corresponding to mutual information in the objective function, the more the model tends to compress the description, and the mutual information naturally decreases. As β increases, mutual information increases significantly because the model tends to reduce the distortion between the compressed image and the original image; the smaller the distortion, the greater the mutual information.

[0044] When β is very large, the difference in mutual information under different values ​​of α1 and α2 is not significant. This is because the influence of β on the model is much greater than that of α1 and α2, resulting in very low distortion between the compressed image and the original image under different values ​​of α1 and α2, and the mutual information tends to be consistent.

[0045] like Figure 4 As shown in (c), when β is small, the larger α1 and α2 are, the lower the classification accuracy. This is because the heavier the image compression by the autoencoder, the smaller the mutual information between the compressed image and the original image, resulting in greater image distortion and thus a decrease in classification accuracy. Figure 4 (a) and Figure 4 The results in (b) are consistent. The larger the β, the greater the mutual information, the more original information is preserved in the compressed image, the easier it is to distinguish different digits, and the higher the classification accuracy.

[0046] like Figure 4 As shown in (d), a comparative experiment was conducted by fixing α2 = 70 and changing the value of α1. The dashed line represents I(X; The value of 1) is represented by the solid line, where I(X; 2) The value of I(X; As β increases, 1) and I(X; 2) Both increase. Under the same β, I(X; 1) Greater than I(X; 2) As expected: Due to the large α2, the model tends to reduce the compression ratio of autoencoder 2, resulting in X and The mutual information between 2 is low.

[0047] The above experimental results verify the effectiveness of the semantic multi-description coding reachability distortion region theory proposed in this invention, and prove that by adjusting the compression ratio weight coefficients α1, α2 and the distortion weight coefficient β, the trade-off between the compression ratio and reconstruction distortion of the two descriptions can be flexibly controlled to achieve the expected semantic extraction performance.

[0048] It is understood that the autoencoder network structure in the above embodiments is not limited to a fully connected neural network. Depending on the actual application requirements, other deep learning network structures such as convolutional neural networks, recurrent neural networks, or transformer networks can also be used.

[0049] It is understood that the source data sequence in the above embodiments is not limited to image data, but can also be text, voice or other types of continuous or discrete data, and the distortion measurement method can be adjusted accordingly.

[0050] Obviously, the above mutual information estimation methods are not limited to variational discriminator methods based on log-likelihood ratio optimization. Other variational mutual information estimation methods can also be used, such as the KL divergence estimation method based on kernel Hilbert space and other commonly used mutual information estimation methods in this field.

[0051] It is understandable that the distortion measure in the Lagrange functional mentioned above is not limited to the combination of mean squared error and cross-entropy loss. Other distortion measures, such as structural similarity index and perceptual loss, can also be selected according to the specific task requirements.

Claims

1. A semantic communication method based on multi-description encoding, characterized in that, Includes the following steps: Obtain the source data sequence; The source data sequence is encoded by a first encoder to generate a first description codeword; The source data sequence is encoded by a second encoder to generate a second description codeword; The first descriptive codeword is decoded by the first decoder to obtain the first reconstructed sequence; The second descriptive codeword is decoded by the second decoder to obtain the second reconstructed sequence; By simultaneously receiving the first description codeword and the second description codeword through a joint semantic decoder, performing joint decoding on the two, and extracting the hidden semantic information contained in the source data sequence, a semantic reconstruction sequence is obtained. The code rate pair of the first description codeword and the second description codeword satisfies the following constraints: the first description code rate is greater than the mutual information between the source data sequence and the first reconstructed sequence; the second description code rate is greater than the mutual information between the source data sequence and the second reconstructed sequence; the sum of the first description code rate and the second description code rate is greater than the sum of the joint mutual information between the source data sequence and the first reconstructed sequence, the second reconstructed sequence, and the semantic reconstructed sequence and the mutual information between the first reconstructed sequence and the second reconstructed sequence; simultaneously, the distortion of the first reconstructed sequence does not exceed the first distortion constraint, the distortion of the second reconstructed sequence does not exceed the second distortion constraint, and the distortion of the semantic reconstructed sequence does not exceed the semantic distortion constraint.

2. The semantic communication method based on multiple description coding according to claim 1, characterized in that, Both the first encoder and the second encoder are implemented using an autoencoder structure. The autoencoder structure includes an encoding subnetwork and a decoding subnetwork. The encoding subnetwork is used to compress the source data sequence into a hidden layer feature representation. The decoding subnetwork is signal-connected to the encoding subnetwork and is used to reconstruct the corresponding reconstructed sequence based on the hidden layer feature representation.

3. The semantic communication method based on multiple description coding according to claim 2, characterized in that, It also includes the following steps: The source data sequence and the first reconstructed sequence are received by the first discriminator, and the mutual information between the source data sequence and the first reconstructed sequence is estimated. The source data sequence and the second reconstructed sequence are received by the second discriminator, and the mutual information between the source data sequence and the second reconstructed sequence is estimated. The estimation results of the first discriminator and the second discriminator are fed back to the first encoder and the second encoder, respectively, to perform adversarial training and update the network parameters of the first encoder and the second encoder.

4. The semantic communication method based on multiple description coding according to claim 3, characterized in that, Both the first and second discriminators are implemented using deep neural networks. The optimal discriminator function is obtained by optimizing the log-likelihood ratio of the joint distribution samples and the product samples of the marginal distributions, so as to estimate the corresponding mutual information.

5. The semantic communication method based on multiple description coding according to claim 1, characterized in that, When jointly training the first encoder, the second encoder, and the joint semantic decoder, a Lagrange functional is used as the training objective function. The Lagrange functional includes: a mutual information term corresponding to the first description code rate weighted by a first compression ratio weight coefficient, a mutual information term corresponding to the second description code rate weighted by a second compression ratio weight coefficient, a distortion term of the first reconstructed sequence weighted by a first distortion weight coefficient, a distortion term of the second reconstructed sequence weighted by a second distortion weight coefficient, and a distortion term of the semantic reconstructed sequence weighted by a semantic distortion weight coefficient. By adjusting the first compression ratio weight coefficient, the second compression ratio weight coefficient, and the distortion weight coefficient, the trade-off between compression ratio and reconstruction distortion is controlled.

6. The semantic communication method based on multiple description coding according to claim 1, characterized in that, The distortion of the first reconstructed sequence is measured by mean squared error, the distortion of the second reconstructed sequence is measured by mean squared error, and the distortion of the semantic reconstructed sequence is measured by cross-entropy loss.

7. A semantic communication system based on multi-description coding, comprising an encoding device and a decoding device, characterized in that: The encoding device includes: The first encoder is used to encode the input source data sequence to generate the first description codeword; The second encoder, arranged in parallel with the first encoder, is used to encode the source data sequence to generate a second description codeword; the first encoder and the second encoder are respectively connected to the channel signal and are used to transmit the first description codeword and the second description codeword to the decoding device via the channel respectively. The decoding device includes: A first decoder, connected to the channel signal, is used to receive the first description codeword, decode the first description codeword, and obtain a first reconstructed sequence. The second decoder is connected to the channel signal and is used to receive the second description codeword, decode the second description codeword, and obtain the second reconstructed sequence. A joint semantic decoder, connected to the channel signal, is used to simultaneously receive the first description codeword and the second description codeword, perform joint decoding on the two, extract the hidden semantic information contained in the source data sequence, and obtain a semantic reconstruction sequence.

8. The semantic communication system based on multiple description coding according to claim 7, characterized in that, Both the first encoder and the second encoder are implemented using an autoencoder network structure, which includes an encoding subnetwork and a decoding subnetwork. The encoding subnetwork is used to compress the source data sequence into a hidden layer feature representation, and the decoding subnetwork is signal-connected to the encoding subnetwork and is used to reconstruct the corresponding reconstructed sequence based on the hidden layer feature representation.

9. The semantic communication system based on multiple description coding according to claim 7, characterized in that, Also includes: The first discriminator is connected to the first encoder signal and is used to receive the source data sequence and the first reconstructed sequence, estimate the mutual information between the two through a deep neural network, and feed the estimation result back to the first encoder. The second discriminator, connected to the second encoder signal, is used to receive the source data sequence and the second reconstructed sequence, estimate the mutual information between them through a deep neural network, and feed the estimation result back to the second encoder.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the following steps: Obtain the source data sequence; The source data sequence is encoded by a first encoder to generate a first description codeword; The source data sequence is encoded by a second encoder to generate a second description codeword; The first descriptive codeword is decoded by the first decoder to obtain the first reconstructed sequence; The second descriptive codeword is decoded by the second decoder to obtain the second reconstructed sequence; By simultaneously receiving the first description codeword and the second description codeword through a joint semantic decoder, performing joint decoding on the two, and extracting the hidden semantic information contained in the source data sequence, a semantic reconstruction sequence is obtained. The code rate pair of the first description codeword and the second description codeword satisfies the following constraints: the first description code rate is greater than the mutual information between the source data sequence and the first reconstructed sequence; the second description code rate is greater than the mutual information between the source data sequence and the second reconstructed sequence; the sum of the first description code rate and the second description code rate is greater than the sum of the joint mutual information between the source data sequence and the first reconstructed sequence, the second reconstructed sequence, and the semantic reconstructed sequence and the mutual information between the first reconstructed sequence and the second reconstructed sequence; simultaneously, the distortion of the first reconstructed sequence does not exceed the first distortion constraint, the distortion of the second reconstructed sequence does not exceed the second distortion constraint, and the distortion of the semantic reconstructed sequence does not exceed the semantic distortion constraint.

Images