A joint source-channel coding method, device and medium based on meta-learning

Abstract

The present invention discloses a joint source channel coding method, device, and medium based on meta-learning, the method comprising: acquiring a target image; performing source channel coding on the target image using a preset target model to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model through a deep joint source channel coding model based on a number of gradients of a number of meta-learning tasks, so that the external network parameters of the neural network model are iteratively updated; and transmitting the target image according to the coding result. The present invention proposes a joint source channel coding method, device, and medium based on meta-learning, which performs source channel coding on the target image using a target model with excellent channel environment adaptability and image coding transmission capability, thereby obtaining a coding result for transmitting the target image, and can solve the problem of difficulty in performing effective image transmission under different channels and small sample conditions.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a meta-learning-based joint source-channel coding method, device, and medium. Background Art

[0002] Joint Source-Channel Coding (JSCC) is an important technology in wireless image transmission. It aims to achieve high-quality image transmission under limited wireless channel bandwidth and transmission capacity by jointly designing source and channel coding schemes. Few-shot learning is a key research direction in machine learning. To effectively utilize limited training samples, researchers have proposed numerous few-shot learning methods, aiming to learn generalized feature representations and models from limited samples to improve wireless image transmission performance under small-shot conditions. By combining few-shot learning methods, source coding techniques, channel coding techniques, and joint optimization algorithms, the problems of few-shot learning in wireless image transmission have been resolved, improving the reliability and efficiency of image transmission.

[0003] However, in practical applications, the limited number of samples required to train the network leads to a decline in network performance when faced with unseen channel conditions, thus affecting its generalization ability. Secondly, since the communication channel is incorporated into the neural network architecture, its performance depends on the accuracy of the channel model. In complex or unstable communication environments, it is difficult for the model to accurately capture the dynamic characteristics of the channel, thus affecting the reliability of the communication system. Thirdly, since the training set contains limited samples of specific channel conditions, the model network overfits to these conditions and performs poorly under other channel conditions. Summary of the Invention

[0004] The present invention provides a meta-learning-based joint source-channel coding method, device and medium to solve the problem of difficulty in performing effective image transmission under different channels and small sample conditions.

[0005] In order to solve the above problems, the present invention provides a joint source-channel coding method based on meta-learning, comprising:

[0006] Acquire the target image;

[0007] Using a preset target model to perform source channel coding on the target image to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model through a deep joint source channel coding model based on multiple gradients of multiple meta-learning tasks, so as to iteratively update external network parameters of the neural network model; the multiple meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, and the deep joint source channel coding model is obtained by performing inner and outer loop training on the preset model;

[0008] The target image is transmitted according to the encoding result.

[0009] The present invention uses a target model to perform source channel coding according to a target image, and can quickly obtain a coding result to transmit the target image. The source channel coding method is simple, fast, and highly practical. Among them, since the deep joint source channel coding model is obtained by performing inner and outer loop training on a preset model, the internal network parameters and external network parameters of the deep joint source channel coding model can be optimized and updated during the training process, so that the deep joint source channel coding model can fully learn the connection and difference between information in a limited training task, and has good image coding transmission capability and adaptability, so that the neural network model can be trained end-to-end according to a number of meta-learning tasks, so that the obtained target model can further improve the image coding transmission capability in a number of meta-learning tasks while learning the ability of the deep joint source channel coding model to quickly adapt to different channel environments.

[0010] Compared with the existing technology, the present invention performs source channel coding on the target image by using a target model with excellent channel environment adaptability and image coding and transmission capabilities, so as to obtain the coding result for transmitting the target image. This can avoid the problem that the model is difficult to accurately capture the dynamic characteristics of the channel or overfitting occurs during encoding, so it can solve the problem of difficulty in effective image transmission under different channels and small sample conditions.

[0011] As a preferred solution, the deep joint source-channel coding model is obtained by performing inner and outer loop training on a preset model, specifically:

[0012] Initializing the internal network parameters of the preset model, using a preset support set, and optimizing and updating the initialized internal network parameters according to a preset first loss function to obtain a first training model;

[0013] Using a preset query set, optimizing and updating the external network parameters of the first training model according to a preset meta-loss function to obtain a second training model;

[0014] Training the second training model using a first training task among a plurality of training tasks to obtain a third training model; wherein the plurality of training tasks are constructed based on a Rayleigh slow fading model and a preset channel signal-to-noise ratio;

[0015] Iteratively optimize and update the internal network parameters and the external network parameters of the third training model, and use the optimized and updated third training model to traverse the remaining tasks in the plurality of training tasks to obtain the deep joint source-channel coding model.

[0016] In this preferred solution, initializing the internal network parameters of the preset model allows the preset model to better generalize to new tasks and concepts. In addition, the parameters of the inner loop are optimized and updated by observing the data of the support set. These parameters focus more on the details of the specific task, so they can help the preset model learn more connections and differences between information in the training tasks, thereby enhancing the model's adaptability.

[0017] From an overall perspective, this preferred solution updates the internal network parameters and external network parameters of the preset model through alternating training, which can continuously optimize the model and improve its ability to adapt to new tasks using a small number of samples. The final deep joint source channel coding model has both good image coding transmission capabilities and adaptability, and can be used for subsequent training of the neural network model to enable the neural network model to quickly adapt to different channel environments.

[0018] As a preferred solution, a preset query set is used to optimize and update the external network parameters of the first training model according to a preset meta-loss function to obtain a second training model, specifically:

[0019] Establishing the meta-loss function using a loss function of the external network parameters of the first training model on the query set after the first training task;

[0020] With the goal of minimizing the meta-loss function, the external network parameters of the first training model are optimized and updated to obtain the second training model.

[0021] The first training model of this preferred solution uses a query set to optimize its performance. Since a query set usually contains multiple tasks or concepts and covers a wide range of data information, by calculating a meta-loss function on the query set and then using this loss to update the parameters of the first training model, the loss of the second training model obtained on the entire query set can be reduced, so that the second training model has good generalization ability.

[0022] As a preferred solution, the second training model is trained using the first training task among the plurality of training tasks to obtain a third training model, specifically:

[0023] Mapping a preset original input image to a complex-valued channel input symbol of a first training task among the plurality of training tasks to obtain an output signal;

[0024] Performing source-channel coding on the original input image using the second training model to obtain a first coding result;

[0025] transmitting the output signal on a channel according to the first encoding result to obtain damaged output data;

[0026] performing approximate reconstruction on the original input image according to the damaged output data to obtain a reconstructed image;

[0027] The second training model is modified according to the original input image and the reconstructed image to obtain the third training model.

[0028] This preferred solution is to train and optimize the third training model through actual training tasks. By combining the first encoding result of the second training model with the output signal transmitted over the channel, actual damaged data can be obtained. Approximate reconstruction is performed based on this data, and the resulting reconstructed image can reflect the difference between itself and the original input image. Therefore, these two sets of images can be used as control groups to correct the second training model, thereby optimizing the encoding capability of the third training model.

[0029] In addition, this method of training through actual channels helps the model better capture the dynamic characteristics of the channel, improves the reliability of the communication system in complex or unstable communication environments, and reduces the overfitting of the model network to specific channel conditions, thereby improving the robust performance of the third training model under various channel conditions.

[0030] As a preferred solution, the several training tasks are constructed based on the Rayleigh slow fading model and a preset channel signal-to-noise ratio, specifically:

[0031] An expression for the multiplication effect of channel gain on the transmission signal is established through a preset channel transfer function and a Gaussian channel transfer function;

[0032] Establishing a channel signal-to-noise ratio based on the average power of the channel input signal;

[0033] The plurality of training tasks are established according to the multiplication effect expression and the channel signal-to-noise ratio.

[0034] In this preferred solution, since the Rayleigh fading model is suitable for describing wireless channels in densely built urban center areas, several training tasks can be quickly constructed based on its channel transfer function and Gaussian channel transfer function, combined with the channel signal-to-noise ratio. The task construction method is simple and fast; at the same time, these established training tasks have different channel conditions, which can help the model perform adaptive training and improve the model's generalization ability.

[0035] As a preferred solution, the target model is obtained by performing end-to-end training on a preset neural network model through a deep joint source-channel coding model according to several gradients of several meta-learning tasks, so that the external network parameters of the neural network model are iteratively updated, specifically:

[0036] Obtaining a first gradient of a first meta-learning task among the plurality of meta-learning tasks, and initializing internal network parameters of the neural network model;

[0037] According to the first gradient, the neural network model is trained end-to-end using the deep joint source-channel coding model to update external network parameters of the neural network model to obtain a first model;

[0038] The first model is used to traverse the remaining tasks in the plurality of meta-learning tasks, and external network parameters of the first model are iteratively updated and fine-tuned to obtain the target model.

[0039] This preferred solution performs end-to-end training on the neural network model through the deep joint source channel coding model. It can fully utilize the structural information of the deep joint source channel coding model to train and optimize the neural network model without providing other additional information, so that the obtained first model can learn the excellent generalization ability of the deep joint source channel coding model, allowing it to quickly adapt to new tasks with a small number of samples; by traversing several meta-learning tasks, it can further enable the target model to have a higher adaptability to limited samples, thereby enhancing the performance of the target model under unknown channel conditions.

[0040] As a preferred solution, the fine-tuning is specifically as follows:

[0041] Using a preset fine-tuning data set to perform data adjustment on the iteratively updated first model, and using a preset loss function to perform back-propagation optimization on the iteratively updated first model;

[0042] The data adjustment includes learning rate adjustment, algorithm optimization and regularization of the internal structure of the model.

[0043] This preferred solution can make the target model better adapt to the task requirements of the dynamic channel environment by fine-tuning the first model after iterative updating, so as to improve the coding capability of the target model.

[0044] As a preferred solution, the several meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, specifically:

[0045] By changing the size of the noise variance to set different average channel signal-to-noise ratios, several average channel signal-to-noise ratios are obtained;

[0046] Get a dataset containing labeled samples and predicted samples;

[0047] The plurality of meta-learning tasks are established according to the plurality of average channel signal-to-noise ratios and the dataset.

[0048] The present invention also provides a joint source-channel coding device based on meta-learning, comprising an acquisition module, a coding module and a transmission module;

[0049] Wherein, the acquisition module is used to acquire the target image;

[0050] The encoding module is configured to perform source-channel coding on the target image using a preset target model to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model using a deep joint source-channel coding model based on multiple gradients of multiple meta-learning tasks, so as to iteratively update external network parameters of the neural network model; the multiple meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, and the deep joint source-channel coding model is obtained by performing inner and outer loop training on the preset model;

[0051] The transmission module is used to transmit the target image according to the encoding result.

[0052] As a preferred solution, the encoding module includes an internal updating unit, an external updating unit, a task training unit and a model acquisition unit;

[0053] The internal updating unit is configured to initialize the internal network parameters of the preset model, and optimize and update the initialized internal network parameters according to a preset first loss function using a preset support set to obtain a first training model;

[0054] The external updating unit is configured to optimize and update the external network parameters of the first training model according to a preset meta-loss function using a preset query set to obtain a second training model;

[0055] The task training unit is configured to train the second training model using a first training task among a plurality of training tasks to obtain a third training model; wherein the plurality of training tasks are constructed based on a Rayleigh slow fading model and a preset channel signal-to-noise ratio;

[0056] The model acquisition unit is used to iteratively optimize and update the internal network parameters and external network parameters of the third training model, and use the optimized and updated third training model to traverse the remaining tasks in the several training tasks to obtain the deep joint source channel coding model.

[0057] As a preferred solution, the external updating unit includes a first updating subunit and a second updating subunit;

[0058] The first updating subunit is configured to establish the meta-loss function using a loss function of the external network parameters of the first training model on the query set after the first training task;

[0059] The second updating subunit is used to optimize and update the external network parameters of the first training model with the goal of minimizing the meta-loss function to obtain the second training model.

[0060] As a preferred solution, the task training unit includes a first training subunit, a second training subunit, a third training subunit, a fourth training subunit and a fifth training subunit;

[0061] The first training subunit is configured to map a preset original input image to a complex-valued channel input symbol of a first training task among the plurality of training tasks to obtain an output signal;

[0062] The second training subunit is configured to perform source channel coding on the original input image using the second training model to obtain a first coding result;

[0063] The third training subunit is configured to transmit the output signal on a channel according to the first encoding result to obtain damaged output data;

[0064] The fourth training subunit is configured to perform approximate reconstruction of the original input image based on the damaged output data to obtain a reconstructed image;

[0065] The fifth training subunit is used to correct the second training model according to the original input image and the reconstructed image to obtain the third training model.

[0066] As a preferred solution, the task training unit includes a first task unit, a second task unit and a third task unit;

[0067] The first task subunit is configured to establish an expression for the multiplication effect of the channel gain on the transmission signal through a preset channel transfer function and a Gaussian channel transfer function;

[0068] The second task subunit is configured to establish a channel signal-to-noise ratio according to an average power of a channel input signal;

[0069] The third task subunit is used to establish the multiple training tasks according to the multiplication effect expression and the channel signal-to-noise ratio.

[0070] As a preferred solution, the encoding module includes an initialization unit, a first updating unit and a second updating unit;

[0071] The initialization unit is configured to obtain a first gradient of a first meta-learning task among the plurality of meta-learning tasks, and initialize internal network parameters of the neural network model;

[0072] The first updating unit is configured to perform end-to-end training on the neural network model through the deep joint source-channel coding model according to the first gradient, so as to update external network parameters of the neural network model to obtain a first model;

[0073] The second updating unit is configured to use the first model to traverse the remaining tasks in the plurality of meta-learning tasks, iteratively update and fine-tune the external network parameters of the first model, and obtain the target model.

[0074] As a preferred solution, the fine-tuning is specifically as follows:

[0075] Using a preset fine-tuning data set to perform data adjustment on the iteratively updated first model, and using a preset loss function to perform back-propagation optimization on the iteratively updated first model;

[0076] The data adjustment includes learning rate adjustment, algorithm optimization and regularization of the internal structure of the model.

[0077] As a preferred solution, the encoding module includes a first establishing unit, a second establishing unit and a third establishing unit;

[0078] The first establishing unit is configured to set different average channel signal-to-noise ratios by changing the size of the noise variance to obtain a plurality of average channel signal-to-noise ratios;

[0079] The second establishing unit is used to obtain a data set including label samples and prediction samples;

[0080] The third establishing unit is configured to establish the plurality of meta-learning tasks according to the plurality of average channel signal-to-noise ratios and the data set.

[0081] The present invention also provides a storage medium having a computer program stored thereon, wherein the computer program is called and executed by a computer to implement the above-mentioned joint source-channel coding method based on meta-learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 1 is a flow chart of a joint source-channel coding method based on meta-learning provided by an embodiment of the present invention;

[0083] Figure 2 This is a system architecture diagram of a deep joint source-channel coding solution based on meta-learning provided by an embodiment of the present invention;

[0084] Figure 3 This is a graph showing the convergence of the algorithm provided by an embodiment of the present invention;

[0085] Figure 4 This is a comparison chart of the impact of the number of training samples on image transmission performance provided by an embodiment of the present invention;

[0086] Figure 5 is a diagram showing the relationship between the channel environment and image transmission performance provided by an embodiment of the present invention;

[0087] Figure 6 It is a structural diagram of a joint source-channel coding device based on meta-learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0088] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0089] In the description of this application, it should be understood that the terms "first," "second," ..., and "fifth" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," ..., and "fifth" may explicitly or implicitly include one or more of such features. In the description of this application, unless otherwise specified, "several" means two or more.

[0090] The meta-learning-based joint source-channel coding method described in the embodiment of the present invention is mainly used in situations where wireless image transmission is required, especially when the model has fewer training samples. This solution can fully utilize limited small samples to train the model, thereby improving the reliability and efficiency of the model for image transmission.

[0091] In the description of this application, it should be noted that SNR is the abbreviation of SIGNAL-NOISE RATIO, which means signal-to-noise ratio.

[0092] Example 1:

[0093] See also Figure 1 The embodiment of the present invention provides a joint source-channel coding method based on meta-learning, including S1 to S3. The specific implementation steps are as follows:

[0094] S1. Acquire the target image.

[0095] Step S1 of the embodiment of the present invention is specifically as follows:

[0096] A target image is acquired, where the target image refers to an image to be wirelessly transmitted.

[0097] S2. Use a preset target model to perform source channel coding according to the target image to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model through a deep joint source channel coding model based on several gradients of several meta-learning tasks, so that the external network parameters of the neural network model are iteratively updated; several meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, and the deep joint source channel coding model is obtained by performing inner and outer loop training on the preset model.

[0098] Step S2 of the embodiment of the present invention is specifically as follows:

[0099] Use a preset target model to perform source channel coding according to the target image to obtain a coding result;

[0100] The target model construction process includes S2.1 to S2.11, specifically:

[0101] S2.1. Establish an expression for the multiplication effect of channel gain on the transmitted signal using a preset channel transfer function and a Gaussian channel transfer function;

[0102] According to the average power of the channel input signal and by changing the noise variance σ 2 The average channel signal-to-noise ratio (SNR) is set to 0dB, 5dB, 10dB, 15dB, and 20dB respectively.

[0103] Establish several training tasks based on the multiplication effect expression and the average channel signal-to-noise ratio;

[0104] The multiplier effect expression is:

[0105] η(z)=η n (η h (z))=hz+n

[0106] Among them, the multiplication effect of channel gain on the transmission signal is given by the channel transfer function η h (z) = hz capture, is a complex normal random variable, and the Gaussian channel transfer function is η n (z) = z + n, Indicates noise.

[0107] The average channel signal-to-noise ratio is:

[0108]

[0109] Wherein, P represents the average power of the channel input signal after the power normalization layer is applied at the encoder, which is set to P=1; the unit of SNR is decibel (dB).

[0110] The channel for each training task constructed in this embodiment is a Rayleigh slow fading model. Since the Rayleigh fading model is suitable for describing wireless channels in densely built urban center areas, several training tasks can be quickly constructed based on its channel transfer function and Gaussian channel transfer function, combined with the channel signal-to-noise ratio. The task construction method is simple and fast. At the same time, these established training tasks have different channel conditions, which can help the model perform adaptive training and improve the model's generalization ability.

[0111] S2.2. Update the internal network parameters of the preset model through an inner loop. That is, at the beginning of each inner loop, use the external network parameters φ to initialize the internal network parameters θ of the preset model. Use the Adam optimizer to optimize and update the initialized internal network parameters based on the preset support set according to the preset first loss function to obtain a first training model.

[0112] Among them, the first loss function is:

[0113]

[0114] in, represents the mean square error distortion, x i For sample i in the support set, is the estimated value of sample i; N is the number of samples.

[0115] The specific process of optimizing and updating the initialized internal network parameters is as follows: for the hth task in the inner loop of several training tasks, its internal network parameter θ h , first by using the first loss function Calculate the gradient, and then use the Adam optimizer to optimize the internal network parameters θ based on the gradient of the current parameters and the average of the historical gradients, and use the gradient descent method h Make updates;

[0116] The above internal network parameters θ h The updating process can be expressed as:

[0117]

[0118] in, Represents the loss function during the inner loop The gradient of the parameter θ, D Sup (h) is the support set of the h-th task, β is the learning rate of the inner loop, θ hare the internal network parameters of the h-th task.

[0119] It should be noted that the preset model is a network model composed of the parameters φ of the external network and the parameters θ of the internal network;

[0120] The external network parameters φ are global parameters in the entire meta-learning process and are optimized in the outer loop of meta-learning. The parameters of the external network are responsible for learning common data information from multiple tasks so that they can quickly adapt to new tasks or environments. The external network parameters φ specifically refer to the parameters of the convolutional layers, batch normalization layers, and fully connected layers in the network model.

[0121] The internal network parameters θ are specific to a single training task and are optimized in the inner loop of meta-learning. In each specific task (image transmission task under specific channel signal-to-noise ratio conditions), the θ parameters are quickly adjusted to adapt to the requirements of the task; these internal network parameters θ are derived from the φ parameters and are optimized for the specific task.

[0122] In this embodiment, initializing the internal network parameters of the preset model can enable the preset model to better generalize to new tasks and concepts; moreover, the parameters of the inner loop are optimized and updated by observing the data of the support set. These parameters focus more on the details of specific tasks, so they can help the preset model learn more connections and differences between information in training tasks, thereby enhancing the model's adaptability.

[0123] S2.3. Update the external network parameters through the outer loop, i.e., establish a meta-loss function using the loss function of the external network parameters φ of the first training model on the query set after the first training task;

[0124] With the goal of minimizing the meta-loss function, the external network parameters of the first training model are optimized and updated to obtain the second training model;

[0125] Among them, the meta-loss function is:

[0126]

[0127] in, is the loss function associated with the training query dataset and the meta-training source task-specific parameters θ, θ h is the internal network parameter after optimization in the hth training task, The optimized parameter θ for the hth training task h The loss function on the query set, h is the ordinal number of the training task, is the total number of tasks.

[0128] The first training model of this embodiment uses a query set to optimize its performance. Since a query set usually contains multiple tasks or concepts and covers a wide range of data information, by calculating the meta-loss function on the query set and then using this loss to update the parameters of the first training model, the loss of the second training model obtained on the entire query set can be reduced, so that the second training model has good generalization ability.

[0129] S2.4, the preset original input image x (where x∈R n ) is mapped to the complex-valued channel input symbol z of the first training task in several training tasks (where z∈C k ), the output signal z ′ ; Wherein, the size of the original input image x is H (height) × W (width) × C (number of channels), n = H × W × C, R represents the set of real numbers, C represents the set of complex numbers, and k is the size of the channel input symbol;

[0130] Performing source-channel coding on the original input image using the second training model to obtain a first coding result;

[0131] According to the first encoding result, the output signal z ′ Transmitted on the channel, corrupted output data is obtained;

[0132] The original input image is approximately reconstructed based on the damaged output data to obtain the reconstructed image. (in, );

[0133] The second training model is modified according to the original input image and the reconstructed image to obtain a third training model.

[0134] To apply the embodiments of the present invention, please refer to Figure 2 , Figure 2 This is a system architecture diagram for a deep joint source channel coding scheme based on meta-learning. It provides a specific process for reconstructing images. By establishing a wireless image transmission system under multiple channel environments and setting the channel signal-to-noise ratio to simulate the changing transmission channel environment, the wireless image transmission system is designed using the joint source channel coding (JSCC) technology. Specifically, an encoder is deployed at the transmitting end in combination with a second training model to perform source coding and channel coding on the original input image x, convert the image data into coded data suitable for transmission, and compress and encode the image; the wireless communication channel is modeled as a series of non-trainable layers and incorporated into the entire neural network architecture; a decoder is set at the receiving end to reverse the operations performed by the encoder through a series of transposed convolutional layers (with nonlinear activation functions) to restore the image and obtain a reconstructed image.

[0135] This embodiment is to train and optimize the third training model through the actual training task; wherein, the original input image x is mapped to the complex channel input symbol z, and then the output signal z is ′ It is then transmitted over the channel. This approach can control the amplitude and frequency distribution of the image signal within a reasonable range, thereby reducing the bandwidth and signal-to-noise ratio requirements during transmission. Furthermore, since the complex-valued channel input symbol z can offset the influence of channel noise by adjusting its phase, the reliability of transmission can be improved.

[0136] By combining the first encoding result of the second training model with the output signal transmitted over the channel, the actual damaged data can be obtained. Approximate reconstruction is performed based on this data. The resulting reconstructed image can reflect the difference between itself and the original input image. Therefore, these two sets of images can be used as a control group to correct the second training model, thereby optimizing the coding capability of the third training model.

[0137] In addition, this method of training through actual channels helps the model better capture the dynamic characteristics of the channel, improves the reliability of the communication system in complex or unstable communication environments, and reduces the overfitting of the model network to specific channel conditions, thereby improving the robust performance of the third training model under various channel conditions.

[0138] S2.5. Iteratively optimize and update the internal network parameters and external network parameters of the third training model, and use the optimized and updated third training model to traverse the remaining tasks in the multiple training tasks until the meta-loss function Convergence, a preliminary deep joint source-channel coding model is obtained.

[0139] S2.6. Use the meta-learned external network parameters φ to adjust the deep joint source channel coding model parameters θ under a preset specific target channel SNR channel environment. k Perform adaptive fine-tuning; where the parameter θ k It is the model parameter corresponding to the deep joint source-channel coding model when transmitting images in a new target environment;

[0140] The gradient descent method is used to adjust θ on the preset adaptive dataset D(k). k Adaptive fine-tuning is performed again to obtain a further improved deep joint source-channel coding model;

[0141] The above about parameter θ k The fine-tuning optimization process can be expressed as:

[0142]

[0143] Among them, γ represents the learning rate of the model in the target environment, Indicates D Ad The loss function on (k), Represents the loss function About parameter θ k gradient.

[0144] It should be noted that the performance of the above-mentioned deep JSCC algorithm and all benchmark solutions is quantified based on the Peak Signal-to-Noise Ratio (PSNR). The PSNR metric measures the ratio between the maximum possible power of the signal and the noise power of the interfering signal, specifically expressed as:

[0145]

[0146] in, is the original input image x and the reconstructed image The mean square error between them; MAX is the maximum possible value of the image pixel. For RGB images, MAX = 2 8 -1=255.

[0147] From an overall perspective, in this embodiment S2.2 to S2.6, using the parameters of the outer loop to initialize the parameters of the inner loop enables the model to better generalize to new tasks or concepts; the parameters of the outer loop are adjusted through a global optimization algorithm on a larger query set, and these parameters include the generalization ability of the model on various tasks; while the parameters of the inner loop are updated by observing the data of the support set, and these parameters focus more on the details of the specific task. This method of updating the internal network parameters and external network parameters of the preset model through alternating training can continuously optimize the model, improve the ability to adapt to new tasks using a small number of samples, and enable the final deep joint source channel coding model to have both good image coding transmission capabilities and adaptability, and can be used for subsequent training of the neural network model to enable the neural network model to quickly adapt to different channel environments;

[0148] Moreover, through the final fine-tuning step, the model can update its parameters in the new target environment to have more reliable image transmission capabilities.

[0149] S2.7, by changing the noise variance σ 2 The size of is set to different average channel signal-to-noise ratios (SNRs) to obtain several average channel signal-to-noise ratios;

[0150] The N-way K-shot method is used to obtain data from N categories (for example, airplanes, cars, birds, cats, ships, and trucks). Each category contains K labeled samples, and q samples are extracted from each category as prediction samples.

[0151] The dataset consisting of labeled samples is defined as the support set D sup (support set), the dataset consisting of prediction samples is defined as query set D Que (query set).

[0152] Among them, the support set D sup and query set D Que They are:

[0153]

[0154] Among them, x i and y i They represent the i-th sample in the label sample and its corresponding label category, N is the number of categories, K is the number of samples of each category in the support set, and q is the number of predicted samples of each category in the query set.

[0155] Several meta-learning tasks are established based on several average channel signal-to-noise ratios and datasets containing support sets and query sets.

[0156] S2.8. Obtain the first gradient of the first meta-learning task among several meta-learning tasks, and initialize the internal network parameters of the neural network model to a set of random parameters θ h Among them, the choices of neural network models include convolutional neural network (CNN), recurrent neural network (RNN), Transformer model and long short-term memory network (LSTM);

[0157] According to the first gradient, the neural network model is trained end-to-end through the deep joint source channel coding model, and the external network parameter φ of the neural network model is adjusted by combining the gradient descent algorithm. h Update to get the first model;

[0158] Use the first model to traverse the rest of the meta-learning tasks, and adjust the external network parameters φ of the first model h Perform iterative updates until the first model converges or reaches a preset number of update training times.

[0159] This embodiment performs end-to-end training on the neural network model through the deep joint source channel coding model. It can fully utilize the structural information of the deep joint source channel coding model to train and optimize the neural network model without providing other additional information, so that the obtained first model can learn the excellent generalization ability of the deep joint source channel coding model, allowing it to quickly adapt to new tasks with a small number of samples; by traversing several meta-learning tasks, it can further enable the target model to have a higher adaptability to limited samples, thereby enhancing the performance of the target model under unknown channel conditions.

[0160] S2.9. Obtain a small number of sample image data as a fine-tuning dataset and divide the dataset into a training set, a validation set, and a test set;

[0161] Among them, the training set is used for model fine-tuning, the validation set is used to adjust model parameters and select appropriate hyperparameters for the model, and the test set is used to finally evaluate the model performance.

[0162] S2.10. Using the training set, perform data adjustment on the iteratively updated first model, and use a preset loss function to perform backpropagation optimization on the iteratively updated first model to obtain a target model, i.e., a meta-learning-based JSCC model.

[0163] Among them, data adjustment includes learning rate adjustment, algorithm optimization and regularization of the internal structure of the model.

[0164] This embodiment fine-tunes the first model after iterative updating, so that the target model can better adapt to the task requirements of the dynamic channel environment, thereby improving the coding capability of the target model.

[0165] S2.11. Use the validation set and test set to perform performance evaluation tests on the target model.

[0166] Specifically, during the experimental evaluation phase, appropriate evaluation metrics need to be selected and experiments conducted to assess the performance of meta-learning-based pre-training methods in small-sample wireless image transmission tasks in dynamic environments. Peak signal-to-noise ratio (PSNR) is selected as a transmission quality metric to evaluate the model's transmission performance and signal recovery capabilities. Image quality metrics such as the structural similarity index (SSIM) and mean square error (MSE) are used to evaluate the quality of model-generated images on both validation and test sets.

[0167] This study compares the performance of our meta-learning and transfer learning-based methods with the baseline JSCC model, analyzing the impact of pre-training methods on small-sample wireless image transmission tasks in dynamic environments. A comparative experiment compares the meta-learning-based pre-training method with other traditional methods or baseline models using different training sample scenarios. The performance differences between the meta-learning and transfer learning-based methods are evaluated by comparing performance metrics on a test set, such as transmission quality and image quality. The experimental results are analyzed to explore the effectiveness and advantages of meta-learning and transfer learning methods for small-sample wireless image transmission tasks in dynamic environments.

[0168] To apply the embodiments of the present invention, please refer to Figure 3 , Figure 3The algorithm convergence graph is provided, showing the convergence curve of our meta-learning approach (DeepJSCC-ML) during the fine-tuning phase under a small sample size, 10-way, 20-shot model, and a channel signal-to-noise ratio (SNR) of 10dB. This is compared with the convergence curve of transfer learning during the fine-tuning phase (DeepJSCC-TL). This shows that the meta-learning approach is more effective, with stronger model performance and generalization, and rapid adaptation and generalization to new tasks.

[0169] To apply the embodiments of the present invention, please refer to Figure 4 , Figure 4 A comparative chart comparing the impact of the number of training samples on image transmission performance is provided. It demonstrates that, under the 10-way K-shot and 0dB channel signal-to-noise ratio (SNR) conditions, the peak signal-to-noise ratio (PSNR) obtained by different schemes increases with increasing K. The image transmission performance of the meta-learning method (DeepJSCC-ML) is superior to that of the transfer learning method (DeepJSCC-TL) and the baseline method (DeepJSCC-NT). In low SNR and small sample sizes, data is very limited, and traditional training methods may struggle to obtain sufficient training samples. In contrast, meta-learning-based methods can be trained with a small number of source channel condition samples and learn from them the knowledge and experience to adapt to different SNR channel environments. This rapid adaptability makes meta-learning-based methods more advantageous in low SNR and small sample sizes.

[0170] To apply the embodiments of the present invention, please refer to Figure 5 , Figure 5 A graph showing the relationship between channel environment and image transmission performance is provided, demonstrating the image transmission performance of different schemes under different channel environments using a small sample size 10-way 20-shot scenario. Clearly, the proposed scheme's meta-learning-based approach (DeepJSCC-ML) outperforms baseline schemes, including transfer learning (DeepJSCC-TL) and no transfer (DeepJSCC-NT). The meta-learning-based JSCC model learns under multiple source channel conditions to acquire source channel coding and decoding strategies and rapidly adapt to new channel conditions. Under varying SNR channel conditions, the meta-learning-based JSCC model can adjust the encoding and decoding strategies based on the channel conditions to improve transmission efficiency and reliability.

[0171] In general, in dynamic communication scenarios, traditional deep learning methods, due to the difficulty and high cost of data acquisition, can only obtain limited small sample data, which may lead to overfitting or insufficient generalization. Meta-learning methods, on the other hand, learn multiple small sample tasks from large datasets during the pre-training phase, giving the model prior knowledge and generalization capabilities, allowing it to quickly adapt to new small sample tasks, thus performing better in small sample situations in dynamic communication environments.

[0172] Meta-learning methods can acquire certain knowledge and experience from large-scale datasets by pre-training models. This prior knowledge includes information on channel conditions, noise levels, transmission parameters, and other aspects. When faced with small sample tasks in dynamic environments, the model can use this prior knowledge to better understand and handle the current image transmission task, improving transmission performance. Compared to methods that train from scratch, meta-learning methods can more efficiently utilize prior knowledge, thereby achieving better results in dynamic environments.

[0173] Due to the complexity and uncertainty of dynamic communication environments, transmission quality is affected by a variety of factors, such as water quality, propagation loss, and multipath fading. Traditional methods may not be able to adapt well to these changes when faced with new communication tasks. Meta-learning methods, through pre-training and fine-tuning, give models strong generalization capabilities, enabling them to quickly adapt to different communication environments and task requirements, improving transmission performance and stability.

[0174] S3. Transmit the target image according to the encoding result.

[0175] Step S3 of the embodiment of the present invention is specifically as follows:

[0176] The target image is transmitted according to the encoding result.

[0177] In general, the embodiments of the present invention have the following beneficial effects:

[0178] This embodiment aims to train a model network to adapt to small-sample image transmission tasks under different channel conditions through meta-learning and the relationship between the classes and sample numbers of the support set. By incorporating the communication channel as a non-trainable layer into the neural network architecture, the accuracy of the communication channel model can be improved. This helps to better capture the dynamic characteristics of the channel and improve the reliability of the communication system in complex or unstable communication environments. This in turn reduces the model network's overfitting to specific channel conditions, thereby improving performance robustness under various channel conditions.

[0179] Moreover, the internal and external network parameters of the deep joint source channel coding model can be optimized and updated during the training process, so that the deep joint source channel coding model can fully learn the connections and differences between information in limited training tasks, and possess good image coding transmission capabilities and adaptability, so that the neural network model can be trained end-to-end according to several meta-learning tasks, so that the obtained target model can learn the ability of the deep joint source channel coding model to quickly adapt to different channel environments while further improving the image coding transmission capabilities in several meta-learning tasks.

[0180] Example 2:

[0181] See also Figure 6 , an embodiment of the present invention provides a joint source-channel coding device based on meta-learning, comprising an acquisition module 10, an encoding module 20 and a transmission module 30;

[0182] Wherein, the acquisition module 10 is used to acquire the target image;

[0183] An encoding module 20 is configured to perform source-channel coding on a target image using a preset target model to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model using a deep joint source-channel coding model based on a plurality of gradients of a plurality of meta-learning tasks, so as to iteratively update external network parameters of the neural network model; the plurality of meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, and the deep joint source-channel coding model is obtained by performing inner and outer loop training on the preset model;

[0184] The transmission module 30 is used to transmit the target image according to the encoding result.

[0185] In one embodiment, the acquisition module 10 is specifically:

[0186] A target image is acquired, where the target image refers to an image to be wirelessly transmitted.

[0187] In one embodiment, the encoding module 20 is specifically:

[0188] Use a preset target model to perform source channel coding according to the target image to obtain a coding result;

[0189] The target model construction process includes a first task subunit, a second task subunit, a third task subunit, an internal update unit, a first update subunit, a second update subunit, a first training subunit, a second training subunit, a third training subunit, a fourth training subunit, a fifth training subunit, a model acquisition unit, a fine-tuning unit, a first establishment unit, a second establishment unit, a third establishment unit, an initialization unit, a first update unit, a second update unit, a partitioning unit, an optimization unit, and a testing unit.

[0190] The first task subunit is configured to establish an expression for the multiplication effect of the channel gain on the transmission signal through a preset channel transfer function and a Gaussian channel transfer function;

[0191] The second task subunit is used to adjust the average power of the channel input signal and the noise variance σ 2 The average channel signal-to-noise ratio (SNR) is set to 0dB, 5dB, 10dB, 15dB, and 20dB respectively.

[0192] The third task subunit is used to establish a plurality of training tasks according to the multiplication effect expression and the average channel signal-to-noise ratio;

[0193] The multiplier effect expression is:

[0194] η(z)=η n (η h (z))=hz+n

[0195] Among them, the multiplication effect of channel gain on the transmission signal is given by the channel transfer function η h (z) = hz capture, is a complex normal random variable, and the Gaussian channel transfer function is η n (z) = z + n, Indicates noise.

[0196] The average channel signal-to-noise ratio is:

[0197]

[0198] Wherein, P represents the average power of the channel input signal after the power normalization layer is applied at the encoder, which is set to P=1; the unit of SNR is decibel (dB).

[0199] The channel for each training task constructed in this embodiment is a Rayleigh slow fading model. Since the Rayleigh fading model is suitable for describing wireless channels in densely built urban center areas, several training tasks can be quickly constructed based on its channel transfer function and Gaussian channel transfer function, combined with the channel signal-to-noise ratio. The task construction method is simple and fast. At the same time, these established training tasks have different channel conditions, which can help the model perform adaptive training and improve the model's generalization ability.

[0200] An internal update unit is configured to update the internal network parameters of the preset model through an inner loop, that is, at the beginning of each inner loop, the internal network parameters θ of the preset model are initialized using the parameters φ of the external network, and the Adam optimizer is used to optimize and update the initialized internal network parameters based on a preset support set according to a preset first loss function to obtain a first training model;

[0201] Among them, the first loss function is:

[0202]

[0203] in, represents the mean square error distortion, x i For sample i in the support set, is the estimated value of sample i; N is the number of samples.

[0204] The specific process of optimizing and updating the initialized internal network parameters is as follows: for the hth task in the inner loop of several training tasks, its internal network parameter θ h , first by using the first loss function Calculate the gradient, and then use the Adam optimizer to optimize the internal network parameters θ based on the gradient of the current parameters and the average of the historical gradients, and use the gradient descent method h Make updates;

[0205] The above internal network parameters θ h The updating process can be expressed as:

[0206]

[0207] in, Represents the loss function during the inner loop The gradient of the parameter θ, D Sup (h) is the support set of the h-th task, β is the learning rate of the inner loop, θ h are the internal network parameters of the h-th task.

[0208] It should be noted that the preset model is a network model composed of the parameters φ of the external network and the parameters θ of the internal network;

[0209] The external network parameters φ are global parameters in the entire meta-learning process and are optimized in the outer loop of meta-learning. The parameters of the external network are responsible for learning common data information from multiple tasks so that they can quickly adapt to new tasks or environments. The external network parameters φ specifically refer to the parameters of the convolutional layers, batch normalization layers, and fully connected layers in the network model.

[0210] The internal network parameters θ are specific to a single training task and are optimized in the inner loop of meta-learning. In each specific task (image transmission task under specific channel signal-to-noise ratio conditions), the θ parameters are quickly adjusted to adapt to the requirements of the task; these internal network parameters θ are derived from the φ parameters and are optimized for the specific task.

[0211] In this embodiment, initializing the internal network parameters of the preset model can enable the preset model to better generalize to new tasks and concepts; moreover, the parameters of the inner loop are optimized and updated by observing the data of the support set. These parameters focus more on the details of specific tasks, so they can help the preset model learn more connections and differences between information in training tasks, thereby enhancing the model's adaptability.

[0212] A first updating subunit is configured to update the external network parameters through an outer loop, that is, to establish a meta-loss function using the loss function of the external network parameters φ of the first training model on the query set after the first training task;

[0213] A second updating subunit is configured to optimize and update the external network parameters of the first training model with the goal of minimizing the meta-loss function to obtain a second training model;

[0214] Among them, the meta-loss function is:

[0215]

[0216] in, is the loss function associated with the training query dataset and the meta-training source task-specific parameters θ, θ h is the internal network parameter after optimization in the hth training task, The optimized parameter θ for the hth training task h The loss function on the query set, h is the ordinal number of the training task, is the total number of tasks.

[0217] The first training model of this embodiment uses a query set to optimize its performance. Since a query set usually contains multiple tasks or concepts and covers a wide range of data information, by calculating the meta-loss function on the query set and then using this loss to update the parameters of the first training model, the loss of the second training model obtained on the entire query set can be reduced, so that the second training model has good generalization ability.

[0218] The first training subunit is used to convert the preset original input image x (where x∈R n ) is mapped to the complex-valued channel input symbol z of the first training task in several training tasks (where z∈C k ), the output signal z ′ ; Wherein, the size of the original input image x is H (height) × W (width) × C (number of channels), n = H × W × C, R represents the set of real numbers, C represents the set of complex numbers, and k is the size of the channel input symbol;

[0219] A second training subunit is configured to perform source-channel coding on the original input image using a second training model to obtain a first coding result;

[0220] The third training subunit is used to make the output signal z ′ Transmitted on the channel, corrupted output data is obtained;

[0221] The fourth training subunit is used to perform approximate reconstruction of the original input image based on the damaged output data to obtain a reconstructed image (in, );

[0222] The fifth training subunit is used to modify the second training model according to the original input image and the reconstructed image to obtain a third training model.

[0223] To apply the embodiments of the present invention, please refer to Figure 2 , Figure 2 This is a system architecture diagram for a deep joint source channel coding scheme based on meta-learning. It provides a specific process for reconstructing images. By establishing a wireless image transmission system under multiple channel environments and setting the channel signal-to-noise ratio to simulate the changing transmission channel environment, the wireless image transmission system is designed using the joint source channel coding (JSCC) technology. Specifically, an encoder is deployed at the transmitting end in combination with a second training model to perform source coding and channel coding on the original input image x, convert the image data into coded data suitable for transmission, and compress and encode the image; the wireless communication channel is modeled as a series of non-trainable layers and incorporated into the entire neural network architecture; a decoder is set at the receiving end to reverse the operations performed by the encoder through a series of transposed convolutional layers (with nonlinear activation functions) to restore the image and obtain a reconstructed image.

[0224] This embodiment is to train and optimize the third training model through the actual training task; wherein, the original input image x is mapped to the complex channel input symbol z, and then the output signal z is ′ It is then transmitted over the channel. This approach can control the amplitude and frequency distribution of the image signal within a reasonable range, thereby reducing the bandwidth and signal-to-noise ratio requirements during transmission. Furthermore, since the complex-valued channel input symbol z can offset the influence of channel noise by adjusting its phase, the reliability of transmission can be improved.

[0225] By combining the first encoding result of the second training model with the output signal transmitted over the channel, the actual damaged data can be obtained. Approximate reconstruction is performed based on this data. The resulting reconstructed image can reflect the difference between itself and the original input image. Therefore, these two sets of images can be used as a control group to correct the second training model, thereby optimizing the coding capability of the third training model.

[0226] In addition, this method of training through actual channels helps the model better capture the dynamic characteristics of the channel, improves the reliability of the communication system in complex or unstable communication environments, and reduces the overfitting of the model network to specific channel conditions, thereby improving the robust performance of the third training model under various channel conditions.

[0227] The model acquisition unit is used to iteratively optimize and update the internal network parameters and external network parameters of the third training model, and use the optimized and updated third training model to traverse the remaining tasks in the plurality of training tasks until the meta-loss function Convergence, a preliminary deep joint source-channel coding model is obtained.

[0228] The fine-tuning unit is used to use the external network parameters φ after meta-learning to fine-tune the parameters θ of the deep joint source channel coding model under the channel environment of a preset specific target channel SNR. k Perform adaptive fine-tuning; where the parameter θ k It is the model parameter corresponding to the deep joint source-channel coding model when transmitting images in a new target environment;

[0229] The fine-tuning unit is also used to adjust θ on the preset adaptive dataset D(k) by gradient descent method. k Adaptive fine-tuning is performed again to obtain a further improved deep joint source-channel coding model;

[0230] The above about parameter θ k The fine-tuning optimization process can be expressed as:

[0231]

[0232] Among them, γ represents the learning rate of the model in the target environment, Indicates D Ad The loss function on (k), Represents the loss function About parameter θ k gradient.

[0233] It should be noted that the performance of the above-mentioned deep JSCC algorithm and all benchmark solutions is quantified based on the Peak Signal-to-Noise Ratio (PSNR). The PSNR metric measures the ratio between the maximum possible power of the signal and the noise power of the interfering signal, expressed as:

[0234]

[0235] in, is the original input image x and the reconstructed image The mean square error between them; MAX is the maximum possible value of the image pixel. For RGB images, MAX = 2 8 -1=255.

[0236] From a general perspective, the internal update unit ~ adaptive fine-tuning unit of this embodiment uses the parameters of the outer loop to initialize the parameters of the inner loop, which enables the model to better generalize to new tasks or concepts; the parameters of the outer loop are adjusted through a global optimization algorithm on a larger query set, and these parameters include the generalization ability of the model on various tasks; while the parameters of the inner loop are updated by observing the data of the support set, and these parameters focus more on the details of the specific task. This method of updating the internal network parameters and external network parameters of the preset model through alternating training can continuously optimize the model, improve the ability to adapt to new tasks using a small number of samples, and make the final deep joint source channel coding model have both good image coding transmission capabilities and adaptability, and can be used for subsequent training of the neural network model to enable the neural network model to quickly adapt to different channel environments;

[0237] Moreover, through the final fine-tuning step, the model can update its parameters in the new target environment to have more reliable image transmission capabilities.

[0238] The first establishment unit is used to change the noise variance σ 2 The size of is set to different average channel signal-to-noise ratios (SNRs) to obtain several average channel signal-to-noise ratios;

[0239] The second establishment unit is used to obtain data including N categories (for example, airplanes, cars, birds, cats, ships, and trucks) through the N-way K-shot method. Each category of data contains K labeled samples, and q samples are extracted from each category of data as prediction samples;

[0240] The second establishment unit is also used to define the data set consisting of labeled samples as the support set D sup (supportset), the dataset consisting of the prediction samples is defined as the query set D Que (query set).

[0241] Among them, the support set D sup and query set D Que They are:

[0242]

[0243] Among them, x i and y i They represent the i-th sample in the label sample and its corresponding label category, N is the number of categories, K is the number of samples of each category in the support set, and q is the number of predicted samples of each category in the query set.

[0244] The third establishment unit is used to establish a plurality of meta-learning tasks according to a plurality of average channel signal-to-noise ratios and a data set including a support set and a query set.

[0245] Initialization unit, used to obtain the first gradient of the first meta-learning task in several meta-learning tasks, and initialize the internal network parameters of the neural network model to a set of random parameters θ h Among them, the choices of neural network models include convolutional neural network (CNN), recurrent neural network (RNN), Transformer model and long short-term memory network (LSTM);

[0246] The first updating unit is used to perform end-to-end training on the neural network model through the deep joint source channel coding model according to the first gradient, and to combine the gradient descent algorithm to make the external network parameter φ of the neural network model h Update to get the first model;

[0247] The second updating unit is used to use the first model to traverse the remaining tasks in the meta-learning tasks, and to adjust the external network parameters φ of the first model. h Perform iterative updates until the first model converges or reaches a preset number of update training times.

[0248] This embodiment performs end-to-end training on the neural network model through the deep joint source channel coding model. It can fully utilize the structural information of the deep joint source channel coding model to train and optimize the neural network model without providing other additional information, so that the obtained first model can learn the excellent generalization ability of the deep joint source channel coding model, allowing it to quickly adapt to new tasks with a small number of samples; by traversing several meta-learning tasks, it can further enable the target model to have a higher adaptability to limited samples, thereby enhancing the performance of the target model under unknown channel conditions.

[0249] A partitioning unit is used to obtain a number of small sample image data as a fine-tuning dataset and divide the dataset into a training set, a validation set, and a test set;

[0250] Among them, the training set is used for model fine-tuning, the validation set is used to adjust model parameters and select appropriate hyperparameters for the model, and the test set is used to finally evaluate the model performance.

[0251] An optimization unit is used to adjust the data of the iteratively updated first model using the training set, and to perform backpropagation optimization on the iteratively updated first model using a preset loss function to obtain a target model, that is, a JSCC model based on meta-learning;

[0252] Among them, data adjustment includes learning rate adjustment, algorithm optimization and regularization of the internal structure of the model.

[0253] This embodiment fine-tunes the first model after iterative updating, so that the target model can better adapt to the task requirements of the dynamic channel environment, thereby improving the coding capability of the target model.

[0254] The testing unit is used to perform validation performance evaluation tests on the target model using the validation set and the test set;

[0255] Specifically, during the experimental evaluation phase, appropriate evaluation metrics need to be selected and experiments conducted to assess the performance of meta-learning-based pre-training methods in small-sample wireless image transmission tasks in dynamic environments. Peak signal-to-noise ratio (PSNR) is selected as a transmission quality metric to evaluate the model's transmission performance and signal recovery capabilities. Image quality metrics such as the structural similarity index (SSIM) and mean square error (MSE) are used to evaluate the quality of model-generated images on both validation and test sets.

[0256] This study compares the performance of our meta-learning and transfer learning-based methods with the baseline JSCC model, analyzing the impact of pre-training methods on small-sample wireless image transmission tasks in dynamic environments. A comparative experiment compares the meta-learning-based pre-training method with other traditional methods or baseline models using different training sample scenarios. The performance differences between the meta-learning and transfer learning-based methods are evaluated by comparing performance metrics on a test set, such as transmission quality and image quality. The experimental results are analyzed to explore the effectiveness and advantages of meta-learning and transfer learning methods for small-sample wireless image transmission tasks in dynamic environments.

[0257] To apply the embodiments of the present invention, please refer to Figure 3 , Figure 3 The algorithm convergence graph is provided, showing the convergence curve of our meta-learning approach (DeepJSCC-ML) during the fine-tuning phase under a small sample size, 10-way, 20-shot model, and a channel signal-to-noise ratio (SNR) of 10dB. This is compared with the convergence curve of transfer learning during the fine-tuning phase (DeepJSCC-TL). This shows that the meta-learning approach is more effective, with stronger model performance and generalization, and rapid adaptation and generalization to new tasks.

[0258] To apply the embodiments of the present invention, please refer to Figure 4 , Figure 4 A comparative chart comparing the impact of the number of training samples on image transmission performance is provided. It demonstrates that, under the 10-way K-shot and 0dB channel signal-to-noise ratio (SNR) conditions, the peak signal-to-noise ratio (PSNR) obtained by different schemes increases with increasing K. The image transmission performance of the meta-learning method (DeepJSCC-ML) is superior to that of the transfer learning method (DeepJSCC-TL) and the baseline method (DeepJSCC-NT). In low SNR and small sample sizes, data is very limited, and traditional training methods may struggle to obtain sufficient training samples. In contrast, meta-learning-based methods can be trained with a small number of source channel condition samples and learn from them the knowledge and experience to adapt to different SNR channel environments. This rapid adaptability makes meta-learning-based methods more advantageous in low SNR and small sample sizes.

[0259] To apply the embodiments of the present invention, please refer to Figure 5 , Figure 5A graph showing the relationship between channel environment and image transmission performance is provided, demonstrating the image transmission performance of different schemes under different channel environments using a small sample size 10-way 20-shot scenario. Clearly, the proposed scheme's meta-learning-based approach (DeepJSCC-ML) outperforms baseline schemes, including transfer learning (DeepJSCC-TL) and no transfer (DeepJSCC-NT). The meta-learning-based JSCC model learns under multiple source channel conditions to acquire source channel coding and decoding strategies and rapidly adapt to new channel conditions. Under varying SNR channel conditions, the meta-learning-based JSCC model can adjust the encoding and decoding strategies based on the channel conditions to improve transmission efficiency and reliability.

[0260] In general, in the test unit of this embodiment, traditional deep learning methods can only obtain limited small sample data in dynamic communication scenarios due to the difficulty and high cost of data acquisition, which may lead to overfitting or insufficient generalization. Meta-learning methods, on the other hand, learn multiple small sample tasks from large-scale datasets during the pre-training phase, giving the model prior knowledge and generalization capabilities, allowing it to quickly adapt to new small sample tasks, thus performing better in small sample situations in dynamic communication environments.

[0261] Meta-learning methods can acquire certain knowledge and experience from large-scale datasets by pre-training models. This prior knowledge includes information on channel conditions, noise levels, transmission parameters, and other aspects. When faced with small sample tasks in dynamic environments, the model can use this prior knowledge to better understand and handle the current image transmission task, improving transmission performance. Compared to methods that train from scratch, meta-learning methods can more efficiently utilize prior knowledge, thereby achieving better results in dynamic environments.

[0262] Due to the complexity and uncertainty of dynamic communication environments, transmission quality is affected by a variety of factors, such as water quality, propagation loss, and multipath fading. Traditional methods may not be able to adapt well to these changes when faced with new communication tasks. Meta-learning methods, through pre-training and fine-tuning, give models strong generalization capabilities, enabling them to quickly adapt to different communication environments and task requirements, improving transmission performance and stability.

[0263] In one embodiment, the transmission module 30 is specifically:

[0264] The target image is transmitted according to the encoding result.

[0265] In general, the embodiments of the present invention have the following beneficial effects:

[0266] This device aims to train a model network to adapt to small-sample image transmission tasks under different channel conditions through a meta-learning approach and by designing the relationship between the classes of the support set and the number of samples. By incorporating the communication channel as a non-trainable layer into the neural network architecture, the accuracy of the communication channel model can be improved. This helps to better capture the dynamic characteristics of the channel and improve the reliability of the communication system in complex or unstable communication environments. This in turn reduces the model network's overfitting to specific channel conditions, thereby improving performance robustness under various channel conditions.

[0267] Moreover, the internal and external network parameters of the deep joint source channel coding model can be optimized and updated during the training process, so that the deep joint source channel coding model can fully learn the connections and differences between information in limited training tasks, and possess good image coding transmission capabilities and adaptability, so that the neural network model can be trained end-to-end according to several meta-learning tasks, so that the obtained target model can learn the ability of the deep joint source channel coding model to quickly adapt to different channel environments while further improving the image coding transmission capabilities in several meta-learning tasks.

[0268] Example 3:

[0269] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to perform the joint source-channel coding method based on meta-learning;

[0270] Wherein, if the meta-learning-based joint source-channel coding method is implemented in the form of a software functional unit and used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0271] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A joint source-channel coding method based on meta-learning, characterized in that: include: Acquire the target image; performing source-channel coding on the target image using a preset target model to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model using a deep joint source-channel coding model based on multiple gradients of multiple meta-learning tasks, so as to iteratively update external network parameters of the neural network model; and the multiple meta-learning tasks are constructed by setting different average channel signal-to-noise ratios; Transmitting the target image according to the encoding result; The deep joint source channel coding model is obtained by performing inner and outer loop training on a preset model, specifically: Initializing the internal network parameters of the preset model, using a preset support set, and optimizing and updating the initialized internal network parameters according to a preset first loss function to obtain a first training model; Using a preset query set, optimizing and updating the external network parameters of the first training model according to a preset meta-loss function to obtain a second training model; Training the second training model using a first training task among a plurality of training tasks to obtain a third training model; wherein the plurality of training tasks are constructed based on a Rayleigh slow fading model and a preset channel signal-to-noise ratio; Iteratively optimize and update the internal network parameters and the external network parameters of the third training model, and use the optimized and updated third training model to traverse the remaining tasks in the plurality of training tasks to obtain the deep joint source-channel coding model.

2. The meta-learning-based joint source-channel coding method according to claim 1, wherein: Using a preset query set, the external network parameters of the first training model are optimized and updated according to a preset meta-loss function to obtain a second training model, specifically: Establishing the meta-loss function using a loss function of the external network parameters of the first training model on the query set after the first training task; With the goal of minimizing the meta-loss function, the external network parameters of the first training model are optimized and updated to obtain the second training model.

3. The joint source-channel coding method based on meta-learning according to claim 1, wherein: The second training model is trained using the first training task among the plurality of training tasks to obtain a third training model, specifically: Mapping a preset original input image to a complex-valued channel input symbol of a first training task among the plurality of training tasks to obtain an output signal; Performing source-channel coding on the original input image using the second training model to obtain a first coding result; transmitting the output signal on a channel according to the first encoding result to obtain damaged output data; performing approximate reconstruction on the original input image according to the damaged output data to obtain a reconstructed image; The second training model is modified according to the original input image and the reconstructed image to obtain the third training model.

4. The meta-learning-based joint source-channel coding method according to claim 1, wherein: The training tasks are constructed based on the Rayleigh slow fading model and a preset channel signal-to-noise ratio, specifically: An expression for the multiplication effect of channel gain on the transmission signal is established through a preset channel transfer function and a Gaussian channel transfer function; Establishing a channel signal-to-noise ratio based on the average power of the channel input signal; The plurality of training tasks are established according to the multiplication effect expression and the channel signal-to-noise ratio.

5. The meta-learning-based joint source-channel coding method according to claim 1, wherein: The target model is obtained by performing end-to-end training on a preset neural network model through a deep joint source-channel coding model according to several gradients of several meta-learning tasks, so that the external network parameters of the neural network model are iteratively updated. Specifically, Obtaining a first gradient of a first meta-learning task among the plurality of meta-learning tasks, and initializing internal network parameters of the neural network model; According to the first gradient, the neural network model is trained end-to-end using the deep joint source-channel coding model to update external network parameters of the neural network model to obtain a first model; The first model is used to traverse the remaining tasks in the plurality of meta-learning tasks, and external network parameters of the first model are iteratively updated and fine-tuned to obtain the target model.

6. The joint source-channel coding method based on meta-learning according to claim 5, characterized in that: The fine-tuning is specifically as follows: Using a preset fine-tuning data set to perform data adjustment on the iteratively updated first model, and using a preset loss function to perform back-propagation optimization on the iteratively updated first model; The data adjustment includes learning rate adjustment, algorithm optimization and regularization of the internal structure of the model.

7. The joint source-channel coding method based on meta-learning according to claim 1, wherein: The meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, specifically: By changing the size of the noise variance to set different average channel signal-to-noise ratios, several average channel signal-to-noise ratios are obtained; Get a dataset containing labeled samples and predicted samples; The plurality of meta-learning tasks are established according to the plurality of average channel signal-to-noise ratios and the dataset.

8. A joint source-channel coding device based on meta-learning, characterized in that: It includes an acquisition module, an encoding module and a transmission module; Wherein, the acquisition module is used to acquire the target image; The encoding module is configured to perform source-channel coding on the target image using a preset target model to obtain a coding result; wherein the target model is obtained by performing end-to-end training on a preset neural network model using a deep joint source-channel coding model based on multiple gradients of multiple meta-learning tasks, so as to iteratively update external network parameters of the neural network model; the multiple meta-learning tasks are constructed by setting different average channel signal-to-noise ratios, and the deep joint source-channel coding model is obtained by performing inner and outer loop training on the preset model; The transmission module is used to transmit the target image according to the encoding result; The encoding module includes an internal updating unit, an external updating unit, a task training unit and a model acquisition unit; The internal updating unit is configured to initialize the internal network parameters of the preset model, and optimize and update the initialized internal network parameters according to a preset first loss function using a preset support set to obtain a first training model; The external updating unit is configured to optimize and update the external network parameters of the first training model according to a preset meta-loss function using a preset query set to obtain a second training model; The task training unit is configured to train the second training model using a first training task among a plurality of training tasks to obtain a third training model; wherein the plurality of training tasks are constructed based on a Rayleigh slow fading model and a preset channel signal-to-noise ratio; The model acquisition unit is used to iteratively optimize and update the internal network parameters and external network parameters of the third training model, and use the optimized and updated third training model to traverse the remaining tasks in the several training tasks to obtain the deep joint source channel coding model.

9. The joint source-channel coding device based on meta-learning according to claim 8, characterized in that: The external updating unit includes a first updating subunit and a second updating subunit; The first updating subunit is configured to establish the meta-loss function using a loss function of the external network parameters of the first training model on the query set after the first training task; The second updating subunit is used to optimize and update the external network parameters of the first training model with the goal of minimizing the meta-loss function to obtain the second training model.

10. The joint source-channel coding device based on meta-learning according to claim 8, characterized in that: The task training unit includes a first training subunit, a second training subunit, a third training subunit, a fourth training subunit and a fifth training subunit; The first training subunit is configured to map a preset original input image to a complex-valued channel input symbol of a first training task among the plurality of training tasks to obtain an output signal; The second training subunit is configured to perform source channel coding on the original input image using the second training model to obtain a first coding result; The third training subunit is configured to transmit the output signal on a channel according to the first encoding result to obtain damaged output data; The fourth training subunit is configured to perform approximate reconstruction of the original input image based on the damaged output data to obtain a reconstructed image; The fifth training subunit is used to correct the second training model according to the original input image and the reconstructed image to obtain the third training model.

11. The joint source-channel coding device based on meta-learning according to claim 8, characterized in that: The task training unit includes a first task unit, a second task unit and a third task unit; The first task subunit is configured to establish an expression for the multiplication effect of the channel gain on the transmission signal through a preset channel transfer function and a Gaussian channel transfer function; The second task subunit is configured to establish a channel signal-to-noise ratio according to an average power of a channel input signal; The third task subunit is used to establish the multiple training tasks according to the multiplication effect expression and the channel signal-to-noise ratio.

12. The joint source-channel coding device based on meta-learning according to claim 8, characterized in that: The encoding module includes an initialization unit, a first updating unit and a second updating unit; The initialization unit is configured to obtain a first gradient of a first meta-learning task among the plurality of meta-learning tasks, and initialize internal network parameters of the neural network model; The first updating unit is configured to perform end-to-end training on the neural network model through the deep joint source-channel coding model according to the first gradient, so as to update external network parameters of the neural network model to obtain a first model; The second updating unit is configured to use the first model to traverse the remaining tasks in the plurality of meta-learning tasks, iteratively update and fine-tune the external network parameters of the first model, and obtain the target model.

13. The joint source-channel coding device based on meta-learning according to claim 12, characterized in that: The fine-tuning is specifically as follows: Using a preset fine-tuning data set to perform data adjustment on the iteratively updated first model, and using a preset loss function to perform back-propagation optimization on the iteratively updated first model; The data adjustment includes learning rate adjustment, algorithm optimization and regularization of the internal structure of the model.

14. The joint source-channel coding device based on meta-learning according to claim 8, characterized in that: The encoding module includes a first establishing unit, a second establishing unit and a third establishing unit; The first establishing unit is configured to set different average channel signal-to-noise ratios by changing the size of the noise variance to obtain a plurality of average channel signal-to-noise ratios; The second establishing unit is used to obtain a data set including label samples and prediction samples; The third establishing unit is configured to establish the plurality of meta-learning tasks according to the plurality of average channel signal-to-noise ratios and the data set.

15. A storage medium, characterized in that: The storage medium stores a computer program, which is called and executed by a computer to implement any one of the meta-learning-based joint source-channel coding methods according to claims 1 to 7.

Images