CN115695810A - Low bit rate image compression coding method based on semantic communication - Google Patents

Abstract

The invention proposes a low-bit-rate image compression coding method based on semantic communication. The present invention combines semantic coding technology and LDPC channel coding technology, and considers the weak computing power of hardware equipment. On the one hand, it realizes a compression ratio and definition higher than that of traditional image coding technology, and on the other hand, it ensures that the data in the actual wireless channel Reliable transmission enables the performance-limited sensor network to meet the needs of reliable transmission of large data volume services. And thanks to the generalization ability of the neural network model, the invention frees the image decoding process from the limitation of the encoding format, and can still ensure that the picture is restored normally and meets the requirements of Certain clarity.

Description

A Low Bit Rate Image Compression Coding Method Based on Semantic Communication

technical field

The invention belongs to the technical fields of deep learning and wireless communication, and in particular relates to a low-bit-rate image compression coding method based on semantic communication.

Background technique

In recent years, wildfires have brought severe disasters to the natural world. Traditional satellite and UAV-based detection technologies are inefficient and costly. With the development of communication technology and Internet of Things (IoT) technology, people have begun to use large-scale Wireless sensor network (WirelessSensorNetwork, WSN) for real-time monitoring, however, the performance of sensor networks in remote areas is limited, it is difficult to carry the transmission of large data volume services, especially pictures and other effective information that can provide on-site environmental changes. In addition, after decades of development, the entropy-based image compression coding method has gradually approached the Shannon limit, and the semantic communication technology with the help of deep learning will compress and code based on image semantics to achieve beyond the Shannon limit, and then make the edge network transmission Lots of pictures possible.

At present, most of the research on semantic communication is focused on lightweight data sets such as text, and they are all implemented in simulation software for a single channel. For the establishment and update of shared knowledge bases related to image semantics, end-to-end transmission affects the time-varying physical wireless channels, There are few related studies on the adaptation of frequency-selective features, that is, there is no successful case of image semantic encoding and transmission in an actual wireless environment.

Contents of the invention

The purpose of the present invention is to solve the problems of low compression rate and low definition of traditional image compression coding technology, and the difficulty of direct application of semantic communication technology in actual wireless environment, and propose a low bit rate image compression coding method based on semantic communication .

The present invention is achieved through the following technical solutions. The present invention proposes a semantic communication-based low-bit-rate image compression coding method for edge devices, the method comprising:

Step 1. Prepare before network training, design the encoder, quantizer, generator, discriminator required in the encoding and decoding process, and the optimization function and optimizer of the network training process, and then initialize the respective network parameters; preprocessing of the provided dataset, shuffling, batching and normalizing it;

Step 2. Input the batched data into the training network built in step 1, pass through the encoder, quantizer and generator in turn, and then input the original image and the reconstructed image output by the generator into the discriminator to calculate the corresponding Loss function value, and use stochastic gradient descent method and backpropagation to update network parameters;

Step 3. Determine whether the loss function value converges to the preset value. If so, terminate the training in advance and save the corresponding model parameters for the edge device to work offline. Otherwise, repeat step 2;

Step 4: The edge device starts, automatically loads the pre-trained encoder and quantizer saved in step 3, passes the captured pictures through the encoder, quantizer and LDPC encoder in sequence, and selects the appropriate modulation mode to modulate and send data to the wireless environment;

Step 5: The receiving end loads the pre-training generator saved in step 3, and sequentially demodulates the received signal, performs LDPC decoding, and generates a reconstructed image.

Further, in the encoder and generator network structure described in step 1:

input indicates the input layer; reflectionpadding() refers to the reflection padding layer, and the padding size is in parentheses; h×w×cconv,strides refers to the convolutional layer with the convolution kernel size h×w, the number of channels is c, and the step size is k. It is followed by an instance normalization layer and a ReLU activation layer; the number of channels in the last convolutional layer in the encoder structure C refers to the bottleneck layer, which is used to control the compression ratio; Residual bloock refers to the residual network block, where BatchNorm refers to batch return One chemical layer.

Further, the optimization function is:

表示寻找能使损失函数取极小值时对应的编码器E和生成器G以及能使损失函数取极大值的鉴别器D过程中所使用的极小化极大算法；f(·)和g(·)表示衡量样本真实程度的辅助函数，d(·)表示原图与生成图片的失真函数，H(·)表示熵编码算法，即量化后数据表示所需要的比特开销，

表示取期望值，λ和β表示失真函数项和熵编码项的权重，x和

表示原图与生成图，z表示接收到的信号样本，y表示量化后数据。where D( ) and G( ) represent the functions corresponding to the discriminator network and the generator network, respectively,

Indicates the minimax algorithm used in the process of finding the corresponding encoder E and generator G that can make the loss function take the minimum value and the discriminator D that can make the loss function take the maximum value; f( ) and g(·) represents the auxiliary function to measure the realness of the sample, d(·) represents the distortion function of the original image and the generated image, H(·) represents the entropy coding algorithm, that is, the bit overhead required for quantized data representation,

Represents the expected value, λ and β represent the weight of the distortion function item and the entropy coding item, x and

Represents the original image and the generated image, z represents the received signal sample, and y represents the quantized data.

Further, the implementation process of the quantizer in the step 2 includes the following steps:

The first part is the forward derivation process:

L表示量化集的长度；Among them, z _i represents the i-th data sample in the i-th data stream, and c _j represents the elements in the quantized set, satisfying

L represents the length of the quantization set;

The second part is the backpropagation process:

Among them, exp( ) represents the exponential function, and σ represents the temperature hyperparameter in the softmax function.

Further, the loss function corresponding to the discriminator is:

Among them, k represents the number of discriminators. Each discriminator has the same structure and is independent of each other. For the kth discriminator, its input is an image pair that performs a downsampling operation on the original image with a factor of 2 ^k-1 . Each time Downsampling can provide a high degree of abstraction of the global features of the image pair to ensure high fidelity from local features to global features between the original image and the generated image.

Further, the loss function corresponding to the generator is:

Further, the Adam optimization algorithm is used to backpropagate the gradient value of the obtained loss function value, and the trainable parameters in the discriminator network and the generator network are updated respectively. The specific process of the Adam algorithm is as follows:

v _k ＝β ₁ v _k-1 +(1-β ₁ )g _k

θ _k ＝θ _k-1 -Δg _k

和

表示偏差修正后的动量变量与累加变量，常数β₁和β₂分别为梯度指数加权移动平均的超参数和梯度平方指数加权移动平均的超参数，η为优化器的学习率，常数ε表示防止分母为0所添加的一个极小值，取值为10^-8。Among them, g _k represents the stochastic gradient of the k-th batch of data, v _k represents the momentum variable corresponding to the gradient of the k-th batch of data, s _k represents the cumulative variable of the gradient square of the k-th batch of data,

and

Indicates the momentum variable and accumulation variable after bias correction, the constants β ₁ and β ₂ are the hyperparameters of the gradient exponentially weighted moving average and the hyperparameter of the gradient squared exponentially weighted moving average, respectively, η is the learning rate of the optimizer, and the constant ε represents the prevention of A minimum value added when the denominator is 0, the value is 10 ^-8 .

Further, the selection of a suitable modulation mode for modulation, sending data to the wireless environment, the specific formula is:

s _data = modulate(LDPC(y;r))

Among them, s _data represents the transmitted data, modulate(·) represents the modulation process, LDPC(·) represents the LDPC code encoding process, r represents the code rate, and y represents the encoded and quantized data.

Further, the specific process formula of said step five is:

表示生成器的输入；Among them, demodulate(·) represents the demodulation process, LDPC ^-1 (·) represents the LDPC decoding process, r _data represents the data after passing through the wireless channel,

represents the input to the generator;

The sending data and receiving data satisfy r _data = hs _data + n;

In the formula, h represents the channel state information, and n represents the additive noise of the channel.

The beneficial effects of the present invention are:

The present invention combines semantic coding technology and LDPC channel coding technology, and considers the weak computing power of hardware equipment. On the one hand, it realizes a compression ratio and definition higher than that of traditional image coding technology, and on the other hand, it ensures that the data in the actual wireless channel Reliable transmission enables the performance-limited sensor network to meet the needs of reliable transmission of large data volume services. And thanks to the generalization ability of the neural network model, the invention frees the image decoding process from the limitation of the encoding format, and can still ensure that the picture is restored normally and meets the requirements of Certain clarity.

Description of drawings

Fig. 1 is a system frame diagram of image semantic compression coding method.

Figure 2 is a network structure diagram of the encoder and generator.

Fig. 3 is the algorithm flow chart of neural network model training.

Figure 4 is a comparison diagram of the effect of the semantic compression coding method and the traditional image coding method BPG.

Fig. 5 is a comparison diagram of different channel transmission effects between the semantic compression coding method and the traditional image coding method BPG.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In combination with Figures 1-5, the present invention proposes a semantic communication-based low-bit-rate image compression coding method for edge devices, including the following steps:

Step 1: Prepare for network training. Design the encoder, quantizer, generator, discriminator required in the encoding and decoding process, as well as the optimization function and optimizer of the network training process, and then initialize the respective network parameters; preprocess the provided data set, the It scrambles, batches, and normalizes.

The network structure of the encoder and generator is shown in Figure 2, where input represents the input layer; reflectionpadding() refers to the reflection padding layer, and the padding size is in parentheses; h×w×cconv, strides refers to the convolution kernel size is h × w, the number of channels is c, the convolutional layer with a step size of k, followed by the instance normalization layer and the ReLU activation layer; the number of channels C in the last convolutional layer in the encoder structure refers to the bottleneck layer, which is used for Control the compression ratio; Residualbloock refers to the residual network block, and the specific structure is shown in the third picture in Figure 2, where BatchNorm refers to the batch normalization layer.

The optimization function is as follows:

表示寻找能使损失函数取极小值时对应的编码器E和生成器G以及能使损失函数取极大值的鉴别器D过程中所使用的极小化极大算法。f(·)和g(·)表示衡量样本真实程度的辅助函数。d(·)表示原图与生成图片的失真函数，H(·)表示熵编码算法，即表示量化后数据表示所需要的比特开销，

表示取期望值，λ和β表示失真函数项和熵编码项的权重，x和

表示原图与生成图，z表示接收到的信号样本，y表示量化后数据。where D( ) and G( ) represent the functions corresponding to the discriminator network and the generator network, respectively,

Indicates the minimax algorithm used in the process of finding the corresponding encoder E and generator G that can make the loss function take the minimum value, and the discriminator D that can make the loss function take the maximum value. f( ) and g( ) represent auxiliary functions to measure the realness of the sample. d(·) represents the distortion function of the original image and the generated image, H(·) represents the entropy coding algorithm, that is, the bit overhead required for quantized data representation,

Represents the expected value, λ and β represent the weight of the distortion function item and the entropy coding item, x and

Represents the original image and the generated image, z represents the received signal sample, and y represents the quantized data.

The optimizer adopts Adam algorithm optimizer. The optimizer selection is not unique and can be selected according to requirements.

In this step, the relevant objects required for the training process of the confrontation generative network are mainly initialized. The specific network structure can be built according to the performance of different hardware devices, and the distortion function item and entropy encoding item in the optimization function can also be customized according to requirements. The selection is the same for the optimizer. The key basis for selection is to ensure that the model can be trained stably and converges, and finally achieve the desired effect. In addition, the selection of data sets can be selected according to different application scenarios to obtain better training results, and the batch size and normalization process can be selected according to the configuration used for training.

Step 2. Input the batched data into the training network built in step 1, pass through the encoder, quantizer, and generator in turn, and then input the original image and the reconstructed image output by the generator into the discriminator to calculate the corresponding Loss function value, and use stochastic gradient descent method and backpropagation to update network parameters, including the following steps:

Step 21. Input the current batch data into the encoder and the quantizer, as shown in the following formula:

y=q(E(x;θ))(2)

where E represents the encoder network, θ represents the parameters of the encoder network, and q represents the quantization process.

Among them, in order to ensure that the gradient link does not experience gradient interruption during the backpropagation process, especially after passing through the quantizer, the implementation logic of the quantizer is divided into two parts:

The first part is the forward derivation process:

L表示量化集的长度。Among them, z _i represents the i-th data sample in the i-th data stream, and c _j represents the elements in the quantized set, satisfying

L represents the length of the quantization set.

The second part is the backpropagation process:

Among them, exp( ) represents the exponential function, and σ represents the temperature hyperparameter in the softmax function.

Step 22: Use the quantified data in step 21 as the input of the generator to perform reconstruction from semantics to images, specifically:

表示生成器网络中的可训练参数，

表示步骤二一中量化后的数据，

表示生成图像。Among them, G( ) represents the generator network,

Denotes the trainable parameters in the generator network,

Represents the quantified data in step 21,

Indicates the generated image.

Step two and three, combine the original image of the data set in step one with the generated image obtained in step two and two, and use the image pair together as the input of the discriminator, and find the corresponding discriminator loss function value:

Among them, k represents the number of multi-scale discriminators, and each discriminator has the same structure and is independent of each other. For the kth discriminator, its input is an image pair that performs a downsampling operation on the original image with a factor of 2 ^k-1 , Each downsampling can provide a high degree of abstraction of the global features of the image pair to ensure high fidelity from local features to global features between the original image and the generated image.

The corresponding generator loss function value is:

Step two and four, using the Adam optimization algorithm to backpropagate the gradient value of the loss function value obtained in step two and three, respectively updating the trainable parameters in the discriminator network and the generator network, wherein the specific process of the Adam algorithm is as follows:

和

表示偏差修正后的动量变量与累加变量，常数β₁和β₂分别为梯度指数加权移动平均的超参数和梯度平方指数加权移动平均的超参数，η为优化器的学习率，常数ε表示防止分母为0所添加的一个极小值，通常为10^-8。Among them, g _k represents the stochastic gradient of the k-th batch of data, v _k represents the momentum variable corresponding to the gradient of the k-th batch of data, s _k represents the cumulative variable of the gradient square of the k-th batch of data,

and

Indicates the momentum variable and accumulation variable after bias correction, the constants β ₁ and β ₂ are the hyperparameters of the gradient exponentially weighted moving average and the hyperparameter of the gradient squared exponentially weighted moving average, respectively, η is the learning rate of the optimizer, and the constant ε represents the prevention of A minimum value added by a denominator of 0, usually 10 ^-8 .

In this step, this step mainly carries out the specific training of the confrontation generation network model in the system. For the network structure built in step 1, the optimizer and learning rate, the preprocessed data set, etc. are input into the confrontation generation network framework, And the encoder, generator and discriminator are alternately trained under the guidance of the optimization function.

Step 3. Determine whether the loss function value converges to the preset value. If yes, terminate the training in advance and save the corresponding model parameters. Otherwise, repeat step 2.

In this step, judge according to the loss value obtained in step 2 to determine whether it is converged. For different network parameters and data set inputs, even the same network structure may converge to different values, so the convergence of the optimization function is only the neural network After learning a feature, the obtained model and its effects need to be judged based on some additional tests, such as whether the quality of the pictures actually generated by the model in the test data set and verification data set is also up to expectations, and whether it is necessary to terminate the training early and save the network parameter.

Step 4, the edge device starts, automatically loads the pre-trained encoder and quantizer saved in step 3, passes the captured pictures through the encoder, quantizer, and Low Density Parity Check Code (LowDensityParityCheckCode, LDPC) encoder in sequence, and Select an appropriate modulation mode for modulation and send data to the wireless environment. The specific process is:

s _data = modulate(LDPC(y;r))(9)

Among them, s _data represents the sent data, modulate(·) represents the modulation process, LDPC(·) represents the LDPC code encoding process, r represents the code rate, and y represents the encoded and quantized data. The trained encoder and The quantizer output is obtained, that is, formula (2).

In this step, the edge device in this step directly loads the encoder and quantizer that have been trained in step 2. The pictures captured by the device are encoded and quantized. At the same time, in order to ensure that the data can be reliably transmitted in the actual wireless environment, channel Coding and modulation, where the modulation method can be BPSK, 16QAM, etc. The LDPC coding rate can also be set according to actual needs. Usually, the higher the code rate, the higher the transmission efficiency, but the lower the reliability.

Step 5. The receiving end loads the pre-training generator saved in step 3, demodulates the received signal, LDPC decodes, and generates a reconstructed image in sequence. The specific process is as follows:

表示生成器的输入。Among them, demodulate(·) represents the demodulation process, LDPC ^-1 (·) represents the LDPC decoding process, r _data represents the data after passing through the wireless channel,

Represents the input to the generator.

The sending data and receiving data satisfy r _data = hs _data + n;

In the formula, h represents the channel state information, and n represents the additive noise of the channel.

In this step, the receiver device first demodulates and decodes the received signal to restore the semantics of the image for use by the subsequent generator. The channel noise here should be abstract semantic noise rather than traditional physical noise. The former causes Semantic errors occur in the information, and the size of the error is determined by the shared semantic library of the sender and receiver. It actually refers to the adaptability of the trainable parameters of the encoder network and the generator network to new samples, but it can still ensure that the received semantics are normally restored to The corresponding picture, and the latter causes a bit error phenomenon. When a bit error occurs in the format header, the entire data packet cannot be recognized and decoded.

Example 1:

Carry out image semantic compression according to the method for specific embodiment, as follows:

长度L＝5，优化器选取Adam算法，学习率为0.0002，数据集选取FLAME数据集，分批数量为1；优化函数中，对抗生成网络的损失函数采用最小二乘损失，即令f(x)＝(x-1)²和g(x)＝x²，失真函数选取均方误差函数，权重λ＝10，熵编码部分暂不考虑，即令β＝0；多尺度鉴别器个数为3，对原图对、下采样因子为2的图像对、下采样因子为4的图像对进行鉴别。(1) The first step is to determine the hyperparameters required for model training. The training set is selected from the FLAME data set, the bottleneck layer C is 4 and 8 respectively, and the quantization set

The length L=5, the optimizer selects the Adam algorithm, the learning rate is 0.0002, the data set selects the FLAME data set, and the number of batches is 1; in the optimization function, the loss function of the confrontation generation network adopts the least square loss, that is, f(x) =(x-1) ² and g(x)=x ² , the distortion function selects the mean square error function, the weight λ=10, the entropy coding part is not considered temporarily, that is, β=0; the number of multi-scale discriminators is 3, Identify the original image pair, the image pair with a downsampling factor of 2, and the image pair with a downsampling factor of 4.

(2) The second step is to start training. After the optimization function converges, the model and related trainable parameters are saved and stored in the edge device.

(3) In the third step, the edge device loads the model, captures pictures, codes, quantizes, LDPC codes, the code rate is 1/2, BPSK modulation, and the channel model selects AWGN channel and Rayleigh channel respectively.

(4) In the fourth step, the receiving end receives the data, performs BPSK demodulation, LDPC decoding, and generates a picture.

It can be seen from the above process that the semantic coding method proposed by the present invention has no specific format requirements, so it is more universal than the traditional image coding method.

Example 2:

Perform simulation verification according to the method in the specific implementation:

The simulation conditions are: the image format is RGB format, the resolution is 256×256px, the number of channels is 3, and the maximum value range of pixels on each channel is 255, that is, the storage overhead of the image is 24 bits per pixel. The reference image coding format used for comparison is BPG format, and the image quality metrics use Peak Signal to Noise Ratio (PSNR) and Multiscale Structural Similarity (MSSSIM).

It can be seen from FIG. 4 that the present invention can generate images with higher definition when the bit overhead is slightly smaller. And the bit overhead required to generate a picture of similar definition is smaller.

It can be seen from Figure 4 that the semantic coding compression method is superior to the BPG coding method in terms of compression ratio and image quality.

It can be seen from Figure 5 that under the AWGN channel and Rayleigh channel, the semantic coding method can provide higher image quality than BPG under different SNR environments.

It can be seen from Figure 5 that the semantic coding method has higher adaptability to different wireless channel environments.

A kind of low-bit-rate image compression coding method based on semantic communication proposed by the present invention has been introduced in detail above. In this paper, specific examples are used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for To help understand the method of the present invention and its core idea; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation and scope of application. In summary, the content of this specification It should not be construed as a limitation of the invention.

Claims (9)

1. A low bit rate image compression coding method of edge equipment based on semantic communication is characterized in that: the method comprises the following steps:

firstly, preparing before network training, designing an encoder, a quantizer, a generator and a discriminator required in the encoding and decoding process, and an optimization function and an optimizer in the network training process, and then initializing respective network parameters; preprocessing the provided data set, scrambling, batching and normalizing the data set;

inputting the batched data into the training network established in the step one, sequentially passing through an encoder, a quantizer and a generator, then inputting the original image and the reconstructed image output by the generator into a discriminator together, calculating a corresponding loss function value, and updating network parameters by using a random gradient descent method and back propagation;

step three, judging whether the loss function value converges to a preset value, if so, terminating the training in advance and storing corresponding model parameters for the edge equipment to work offline, otherwise, repeating the step two;

starting edge equipment, automatically loading a pre-training encoder and a quantizer stored in the step three, enabling the captured picture to sequentially pass through the encoder, the quantizer and the low-density parity check code (LDPC) encoder, selecting a proper modulation mode for modulation, and sending data to a wireless environment;

and step five, loading the pre-training generator stored in the step three at the receiving end, and sequentially demodulating, decoding the LDPC and generating a reconstructed image for the received signal.

2. The method of claim 1, wherein in the encoder and generator network structure of step one:

input represents an input layer; reflection padding () refers to the reflective padding layer, with the padding size in parentheses; h multiplied by w multiplied by c conv, stride s refers to a convolution layer with convolution kernel size of h multiplied by w, channel number of c and step length of k, and is connected with an example normalization layer and a ReLU activation layer; the channel number C in the last convolutional layer in the encoder structure refers to the bottleneck layer and is used for controlling the compression ratio; residual block refers to a Residual network block, where Batch Norm refers to the Batch normalization layer.

3. The method of claim 1, wherein the optimization function is:

wherein D (-) and G (-) represent functions corresponding to the network of the discriminator and the network of the generator, respectively,

a minimization maximum algorithm used in the process of searching for the corresponding encoder E and generator G which can make the loss function take the minimum value and the discriminator D which can make the loss function take the maximum value is shown; f (-) and g (-) denote auxiliary functions for measuring the trueness of the sample, d (-)) Represents the distortion function of the original and the generated picture, H (-) represents the entropy coding algorithm, i.e. the bit overhead required for the quantized data representation,

representing the expectation, λ and β represent the weights of the distortion function term and the entropy coding term, x and

the original image is represented and the image is generated, z represents the received signal samples, and y represents the quantized data.

4. The method of claim 1, wherein the quantizer in step two implements a process comprising the steps of:

the first part is a forward derivation process:

wherein z is _i Representing the ith data sample in the ith data stream, c _j Representing elements in the quantization set, satisfy

L represents the length of the quantization set;

the second part is a backward propagation process:

where exp (·) represents an exponential function and σ represents a temperature hyperparameter in the softmax function.

5. The method of claim 1, wherein the penalty function for a corresponding discriminator is:

wherein k represents the number of discriminators, each discriminator has the same structure and is independent of each other, and for the kth discriminator, the input is to perform a factor of 2 on the original image ^k-1 Each downsampling operation can provide high abstraction of global features of the image pair, so that high fidelity from local features to global features between the original image and the generated image is guaranteed.

6. The method of claim 3, wherein the penalty function for a corresponding generator is:

7. the method of claim 1, wherein the trainable parameters in the discriminator network and the generator network are updated respectively by back-propagating the gradient values of the obtained loss function values using an Adam optimization algorithm, wherein the Adam optimization algorithm is specifically performed as follows:

v _k ＝β ₁ v _k-1 +(1-β ₁ )g _k

θ _k ＝θ _k-1 -Δg _k

wherein, g _k Random gradient, v, representing the kth data _k Representing the momentum variable, s, corresponding to the gradient of the kth data _k An accumulated variable representing the square of the gradient of the kth batch of data,

and

representing the deviation-corrected momentum and accumulated variables, constant beta ₁ And beta ₂ The method comprises the steps of respectively weighting a hyperparameter of a gradient index weighted moving average and a hyperparameter of a gradient square index weighted moving average, wherein eta is a learning rate of an optimizer, a constant epsilon represents a minimum value added when a denominator is 0, and a value is 10 ^-8 。

8. The method of claim 1, wherein the selecting of the suitable modulation mode for modulation and sending of the data to the wireless environment is according to the following specific formula:

s _data ＝modulate(LDPC(y；r))

wherein s is _data Represents the transmitted data, modulated (-) represents the modulation process, LDPC (-) represents the LDPC code encoding process, r represents the code rate, and y represents the encoded and quantized data.

9. The method according to claim 1, wherein the step five specific process formula is:

wherein, demodulation (. Cndot.) denotes the demodulation process, LDPC ^-1 (. Cndot.) denotes an LDPC decoding process,r _data representing the data after it has passed through the radio channel,

represents an input to a generator;

wherein the transmission data and the reception data satisfy r _data ＝hs _data +n；

In the formula, h represents channel state information, and n represents additive noise of a channel.

Images