Depth joint source channel coding modulation system
By employing a modulation system with deep joint source-channel coding, and utilizing technologies such as the CIFAR-10 neural network, encoder, channel selector, and DQAM module, the contradiction between transmission load and performance in end-to-end image transmission is resolved, achieving efficient image restoration and transmission.
Patent Information
- Application Number
- CN202411183509.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-12
AI Technical Summary
Existing end-to-end image transmission systems struggle to effectively reduce transmission load while maintaining excellent performance during modulation.
A modulation system employing deep joint source-channel coding is trained using a CIFAR-10 neural network. The Encoder module, consisting of convolutional layers, generalized division normalization layers, and PRelu layers, performs feature extraction and size compression. The analog signal is quantized to constellation points using a channel selector and a DQAM module. The image is then recovered by the Decoder module after transmission through the AWGN channel.
While reducing the transmission load, the performance of image transmission is improved, and the image restoration effect is optimized by calculating quantization error loss and mean square error.
Smart Images

Figure CN118944812A_ABST
Abstract
Description
Technical Field
[0001] The present invention designs a modulation system of deep joint source-channel coding, which belongs to the field of image processing and image transmission. Background Art
[0002] Due to the rapid development of deep learning, joint source channel coding (JSCC) based on deep learning has attracted widespread attention in end-to-end image transmission. In order to modulate and quantize the analog quantity output by the encoder, a modulation system DQAM based on deep joint source channel coding is proposed. The distance matrix is obtained by calculating the distance between the analog quantity and the constellation point by Euclidean distance, and then the gumbel-softmax technique is used to obtain the one-hot matrix. Finally, the modulation constellation point corresponding to the analog value is selected and replaced by the one-hot matrix. The proposed scheme achieves excellent performance while reducing the transmission load. Summary of the invention
[0003] A modulation system for deep joint source channel coding, characterized by comprising the following steps:
[0004] The CIFAR-10 neural network system is used for training. The CIFAR-10 dataset uses an image size of 32×32, and the input image is divided by 255 to normalize the pixel value.
[0005] Furthermore, the Encoder encoding module consisting of multiple convolutional layers, generalized division normalization layers (GDN) and PRelu layers is used to extract features and compress the input image.
[0006] Furthermore, the feature data obtained by the encoder module is input into the channel selector, which first performs mean pooling and connects the SNR information, and then passes through two linear layers to obtain the weight value corresponding to each channel, and selects the channel corresponding to the (modulation order / 32) largest weight;
[0007] Furthermore, the analog feature data obtained by the Encoder encoding module is input into the DQAM module, and the constellation point coordinates are first initialized, and the Euclidean distance between the constellation point coordinates and the analog feature data is calculated to obtain a distance matrix between each data point and each constellation point.
[0008] Furthermore, gumbel-softmax is performed on the obtained distance matrix to obtain the corresponding one-hot matrix. The row in this one-hot matrix represents the number of data points that are closest to the constellation point, and the column represents the number of constellation points that the data point in this row is closest to.
[0009] Furthermore, the characteristic data coordinate points of the analog quantity are quantized into corresponding constellation point coordinate points through a one-hot matrix, that is, the modulated transmission data is obtained;
[0010] Further, the channel data coordinates obtained by the channel selector and the constellation point coordinates are used to perform quantization error loss calculation to obtain a quantization error loss value;
[0011] Furthermore, the modulated data is transmitted through an AWGN channel of [0,20]dB;
[0012] Furthermore, the data received after passing through the AWGN channel is restored and reconstructed through a Decoder module consisting of a deconvolution layer, an inverse GDN layer, and a PRelu layer;
[0013] Furthermore, the mean square error between the image restored by the Decoder module and the original image is calculated, and the quantization error loss value is added to be used for the training of the overall model; BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is an overall structural diagram of a modulation system with deep joint source-channel coding;
[0015] Figure 2 It is a structural diagram of a channel selector;
[0016] Figure 3 64-QAM constellation point distribution diagram DETAILED DESCRIPTION
[0017] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0018] The present invention provides an embodiment: a modulation system for deep joint source channel coding. The scheme structure is composed of Figure 1 As shown, the specific steps include:
[0019] The neural network system is trained using CIFAR-10, which has an image size of 32×32 and contains 50,000 training images and 10,000 test images.
[0020] The input image is divided by 255 to normalize the pixel values. The Encoder module consisting of a convolutional layer, a generalized division normalization layer (GDN) and a PRelu layer is used to extract features and compress the input image.
[0021] The feature data obtained by the Encoder module enters the channel selector module. The structure diagram of the channel selector module is shown in Figure 2. First, the input is average pooled and the SNR information is concatenated. Then, the corresponding channel weight value is obtained through two layers of fully connected layers and Relu and Sigmoid activation functions, and the data of the channel corresponding to the (modulation order / 32) maximum weight is selected. The average pooling is shown as follows:
[0022]
[0023] Among them, F C Corresponding to the value obtained by the cth channel, H×W represents the height and width of the image, u c represents the original image;
[0024] The data coordinate points obtained by the channel selector module are reshaped into a matrix in the form of (-1, 2) and input into the DQAM module (-1 means that the number of rows is automatically adjusted while ensuring that the number of columns is 2). The DQAM module first generates the initialized constellation point coordinates. The constellation point coordinate matrix shape is (M, 2), where M is the modulation order. Then, the constellation point coordinates and the input data point coordinates are calculated by Euclidean distance to obtain the distance matrix:
[0025]
[0026] where a i1 ,a i2 represents the point in the i-th row, the first column and the second column of the data point; b m1 ,b m2 represents the point in the i-th row, the point in the 1st column and the point in the 2nd column in the constellation point. D represents the Euclidean distance matrix between the data point and the constellation point. In the CIFAR-10 dataset, i=512; m is the modulation order;
[0027] The distance matrix is converted to the corresponding one-hot vector by gumbel-softmax. The shape of the one-hot matrix is (H, K), where H is the number of data points and K is the number of constellation points. Therefore, each row in the one-hot vector is a one-hot vector, so the one-hot matrix contains the information of each data point and its corresponding closest constellation point;
[0028] Finally, the characteristic data coordinate points of the simulation quantity are replaced with the corresponding constellation point coordinates through the one-hot matrix to obtain the quantized data; y =
[0029] After the quantized data is power normalized and passed through a random AWGN channel with SNR = [0,20]dB, the power normalization formula is:
[0030]
[0031] Where z represents the n-dimensional vector of the real-valued image x mapped to the k-dimensional vector of the complex-valued channel input sample z. z* represents the complex conjugate of z, and k represents the amount of data in the input channel after the original image x passes through the compression module;
[0032] The data after passing through the AWGN channel is received, and the received data is restored and reconstructed through the Decoder module composed of a deconvolution layer, an inverse GDN layer, and a PRelu layer;
[0033] In the last layer of the Decoder module, the Sigmoid function is used to limit the output range to between 0 and 1, and the pixel value of the obtained image is multiplied by 255 to restore it to the original pixel value range to obtain the restored and reconstructed image;
[0034] The mean square error is calculated between the restored image and the original image, and the quantization error loss is calculated between the important data points obtained by the channel selector and the constellation points. The quantization loss is to calculate the Euclidean distance between the data coordinates of the corresponding channel with the largest weight in the channel selector and the coordinates of the constellation points, and take the minimum distance of each point for averaging, and finally control the distance value through a logarithmic function with a base greater than 0 and less than 1:
[0035]
[0036] Where X is the data point coordinate matrix, Y is the constellation point coordinate matrix, i is the number of data points, and j is the number of constellation points;
[0037]
[0038] Among them, min D i is to take the minimum value of each row of the distance matrix D, d min Represents the average of the minimum distance values of each row of the distance matrix;
[0039] Finally, the quantization error loss value is obtained:
[0040]
Claims
1. A modulation system for deep joint source channel coding, characterized in that: A channel selector and DQAM module are added at the output of the Encoder, the analog quantity is modulated by a neural network, and the constellation point coordinates are updated through training to adapt to different channel conditions. After the data is transmitted through the channel, the received data is restored and reconstructed by the Decoder, and the loss value of the training process is composed of the mean square error and the weighted quantization loss.
2. The modulation system of deep joint source channel coding according to claim 1 is characterized in that The channel is an additive white Gaussian noise channel in the range of [0, 20] dB, and the power of the transmitted data is normalized.
3. According to claim 1, a modulation system for deep joint source channel coding is characterized in that The Encoder consists of five repeated convolutional layers, generalized division normalization layers and PRelu layers. The size of the convolutional layer is f×f×c|s , where f is the convolution kernel size, c is the number of convolution kernel output channels, and s is the step size in the channel direction.
4. According to claim 1, a modulation system for deep joint source channel coding is characterized in that The Decoder consists of five repeated deconvolution layers, inverse generalized division normalization layers and PRelu layers. The size of the deconvolution layer is f×f×c|s , where f is the convolution kernel size, c is the number of convolution kernel output channels, and s is the step size in the channel direction.
5. A modulation system for deep joint source channel coding according to claim 1, characterized in that The channel selector is used to average-pool the input data and then splice the SNR information, obtain the channel weights through two fully connected layers and Relu and Sigmoid activation functions, and select (modulation order / 32) channels as important channels.
6. A modulation system for deep joint source channel coding according to claim 1, characterized in that The function of the DQAM is to obtain a distance matrix by performing Euclidean distance calculation on the trainable constellation coordinates and the input analog coordinates, and obtain a one-hot distance matrix by performing gumbel-softmax on the distance matrix, and finally quantize the analog coordinates into the coordinates of the nearest constellation point in the matrix through the one-hot distance matrix.
7. The Encoder module of the decoder according to claim 3, characterized in that The s of the first three layers is 2, the s of the last two layers is 1, and f is 5.
8. The decoder module of the decoder according to claim 3, characterized in that The first three layers have s=1, the last two layers have s=2, and f=5.
9. The loss value according to claim 1. Characterized in that It is composed of the mean square error and the weighted quantization loss. The quantization loss is to calculate the Euclidean distance between the data coordinates of the corresponding channel with the largest weight in the channel selector and the coordinates of the constellation point, and take the average of the minimum distance in each point. Finally, the distance value is controlled by a logarithmic function with a base greater than 0 and less than 1.
