CSI matrix compression feedback method based on image processing

Abstract

The invention belongs to the technical field of image processing, and discloses a CSI matrix compression feedback method based on image processing, and the specific technical scheme is as follows: calculating a modulus value corresponding to each element in a CSI matrix; all values of the real part elements, the imaginary part elements and the module value elements are scaled to be within the range from 0 to 255 through normalization; creating an image three-dimensional matrix, and converting the image three-dimensional matrix into an unsigned 8-bit integer; the three-dimensional matrix is converted and stored as an image in a. Png form, and the compressed CSI image is transmitted through a channel; a picture generated by a CSI matrix is put into a high-resolution recovery image model to obtain a recovered Recover picture, two picture matrixes are recovered into a real part matrix, an imaginary part matrix and a module value matrix by using a normalized proportion and an offset of non-negative processing, and the CSI matrix is recovered by using the real part matrix and the imaginary part matrix, so that the interference of a system is effectively reduced, and the reliability of the system is improved. And the total throughput of the system is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a CSI matrix compression feedback method based on image processing. Background Art

[0002] With the advancement of mobile communication technology, especially 5G and the upcoming 6G technology, the data transmission demand of wireless networks has increased dramatically. Multi-antenna technology (MIMO) provides a good solution to meet the data transmission rate or quality requirements. Compared with single-antenna technology, MIMO technology changes the channel state information (CSI) from a single value to a matrix. And the more antennas there are, the larger the CSI matrix is, which brings a lot of overhead to CSI information feedback. If it is not processed, it may not meet the real-time requirements. As a result, CSI compression technology has become a hot topic of research. At present, common CSI compression technologies include discrete cosine transform (DCT) and singular value decomposition (SVD), which use the sparse characteristics of the CSI matrix to achieve efficient compression through technologies such as compressed sensing.

[0003] Existing CSI compression feedback technologies mainly have feedback overhead: In MIMO (multiple input multiple output) systems, CSI needs to be fed back from the receiver to the transmitter. As the number of antennas increases, the dimension of CSI also increases significantly, resulting in a significant increase in feedback overhead. Computational complexity: The acquisition, processing, and update of CSI involve complex calculations, especially when there are many users and large antenna arrays, which puts pressure on the computing power and power consumption of the device. Summary of the invention

[0004] In order to solve the technical problems existing in the prior art, the present invention proposes a CSI matrix compression feedback method based on image processing. The algorithm decomposes the real and imaginary parts of the CSI matrix, and corresponds the real part, imaginary part, and modulus value to the RGB three channels respectively, converting them into RGB images for image processing, and uses the RGB image and the values ​​at various pixel points to correspond to the real and imaginary modulus values ​​of the CSI matrix, thereby realizing compression feedback of the CSI matrix, effectively reducing the interference of the system and improving the total throughput of the system.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a CSI matrix compression feedback method based on image processing, the specific steps are as follows:

[0006] Step 1: Assume that the CSI matrix is:

[0007]

[0008] Step 2: The CSI matrix is ​​a complex matrix with m rows and n columns. Read each element in the matrix A row by row and column by column to extract the real part a in the CSI matrix.mn and the imaginary part b mn , and using the modulus formula To calculate the modulus value corresponding to each element in the CSI matrix;

[0009] Step 3: Take the maximum and minimum values ​​of the real elements, and perform non-negative processing on the minimum values ​​of all real elements, that is, add 0 to all elements and subtract the minimum value to ensure that all real elements are non-negative values, so as to carry out the next step of RGB image conversion, and use normalization to scale all values ​​of real elements to the range of 0 to 255;

[0010] Step 4: Take the maximum and minimum values ​​of the imaginary elements, and perform non-negative processing on the minimum values ​​of all imaginary elements to ensure that all imaginary elements are non-negative values ​​for the next step of RGB image conversion. Use normalization to scale all values ​​of the imaginary elements to the range of 0 to 255.

[0011] Step 5: Take the maximum and minimum values ​​of the modulus elements, and perform non-negative processing on the minimum values ​​of all modulus elements to ensure that all modulus elements are non-negative values, so as to carry out the next step of RGB image conversion, and use normalization to scale all values ​​of the modulus elements to the range of 0 to 255;

[0012] Step 6. Create a new three-dimensional image matrix, put the two-dimensional matrix of real elements into the first dimension of the new three-dimensional image matrix, put the two-dimensional matrix of imaginary elements into the second dimension of the new three-dimensional image matrix, put the two-dimensional matrix of modulus elements into the third dimension of the new three-dimensional image matrix, and convert the three-dimensional image matrix into an unsigned 8-bit integer, leaving only positive integers in the range of 0 to 255, and discarding the decimal data after the integer.

[0013] Step 7. Generate a .png image using JPEG compression in cv2. The JPEG image quality factor used here is set to 80 (usually the quality factor ranges from 0 to 100. The quality factor determines how much information is lost during the compression process. A lower QF value (the quality factor (QF: Quality Factor) QF value range is (0,100). The larger the QF, the higher the image visual quality and the lower the compression rate.) will lead to a higher compression rate and a smaller file size, but the image quality will be significantly reduced, and blurring and color blocks may occur. A higher QF value can retain more image details, but the image size will increase accordingly. Choosing 80 as the QF value can usually achieve a moderate file size while maintaining relatively good image quality.

[0014] The three-dimensional matrix in step 6 is converted and saved as a .png image, and the compressed CSI image is transmitted through the channel;

[0015] Step 8. Put the image generated by the CSI matrix into the high-resolution recovery image model to obtain the recovered Recover image, and use the normalized ratio and non-negative offset to restore the two image matrices into real matrix, imaginary matrix, and modulus matrix. During the recovery process, the real matrix and imaginary matrix have been obtained. Multiply all elements of the imaginary matrix by the imaginary unit i, add all corresponding elements of the real matrix and the processed imaginary matrix, and the real matrix and imaginary matrix can be used to restore the original matrix.

[0016] In step three, the specific formula for normalization is: 255 / (MAX-MIN), where MAX is the maximum value and MIN is the minimum value.

[0017] In step eight, the training steps of the high-resolution restored image model are:

[0018] Step 1), putting the JPEG compressed RGB image set generated based on the CSI matrix into step 2) to train a high-resolution image restoration model;

[0019] Step 2) Read the model and set the yml file, including the dataset path, model name, etc.

[0020] Step 3), using the settings to read the status information of the trained model, and further deepening the training according to the status information of the trained model;

[0021] Step 4), initialize the training log;

[0022] Step 5), create a training data reader and a validation data reader;

[0023] Step 6) Find the model class in the settings and load it into the model object, including network structure information, loss function, learning rate, etc.

[0024] Step 7) Create a log generator and initialize the dataset pre-reader (no special model);

[0025] Step 8) Initialize the training iterator and training batch size using the settings in step 2) and enter the training loop (no special model);

[0026] Step 9) Initialize the learning rate according to the set parameters, fix the learning rate in the first cycle, and continue to iterate. In the second cycle, use the cosine annealing strategy, change the learning rate, and continue to iterate (no special model);

[0027] Step 10), using the low-quality compressed image to progressively train and approximate the original image, and updating the batch of the progressively approximated low-quality images;

[0028] Step 11), the model reads the updated batch image data, and inputs the low-quality image (lq) and the original image (gt) (adopting a step-by-step learning strategy, using the low-quality image lq to gradually restore the original image gt, so that the compressed low-quality image lq can achieve a better recovery effect);

[0029] Step 12), optimize the model parameters, call the net_g network forward propagation to obtain the result (that is, to obtain the loss function), call the loss function back propagation to adjust the network parameters, the loss function uses the L1 loss function, the weight is 1, and the L1 formula is:

[0030]

[0031] Where x and y are two input tensors, N is the number of elements in the tensor, and ε is a small positive number used to avoid gradient explosion at zero points.

[0032] Step 13) Verify the model performance, such as loss, accuracy, PSNR (peak signal-to-noise ratio), SSIM (structural similarity), etc., and save the results to the image;

[0033] Step 14), read the next batch for training and return;

[0034] Step 15), save model parameters and information;

[0035] Step 16), put a picture generated by a new CSI matrix into the pre-trained model test to obtain a recovered Recover picture.

[0036] The test steps of the high-resolution restored image model are:

[0037] Step 21), load the yml setting file to obtain model information including network structure, pre-trained network parameters, etc., and set it to non-training mode;

[0038] Step 22), initialize the log system;

[0039] Step 23), initialize the test dataset loader (no special model);

[0040] Step 24), read and create a model according to the setting information, read the network structure and pre-trained network parameters;

[0041] Step 25), read the test data set and enter the network to run, output the results, verify the performance, and save the results to the picture;

[0042] Step 26), extract the image of the original data set and a Recover image, compare the two data, calculate the MSE (mean square error), and output the comparison result graph.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention can process compressed feedback data of multiple CSI (channel state information) matrices at the same time, significantly improving data processing efficiency. In this way, feedback overhead can be more effectively reduced and overall system performance can be improved. It is particularly suitable for large-scale antenna systems or complex network environments that require frequent feedback. This method not only simplifies the channel information transmission process, but also optimizes bandwidth utilization while ensuring information accuracy, further improving the stability and reliability of the communication system.

[0045] Second, the present invention forms an efficient data compression scheme by converting the matrix into an image form and combining it with JPEG compression technology. After these two steps, the data compression effect is significant, which can greatly reduce the resource occupation during the transmission process, especially the bandwidth and storage space. This method minimizes redundant information and improves transmission efficiency while ensuring data integrity. It is very suitable for application in resource-constrained or high-load communication environments. This innovative compression mechanism provides effective technical support for improving the overall performance of the system.

[0046] 3. The present invention performs well in recovering the CSI (channel state information) matrix, and the recovered matrix has high precision and small error. This good recovery effect ensures the accurate transmission of channel information and helps to improve the performance and stability of the communication system. Through the optimized compression and recovery mechanism, the present invention minimizes the information loss in the recovery process while ensuring efficient compression. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall flow chart of the present invention.

[0048] Figure 2 Generate a plot for the original matrix.

[0049] Figure 3 It is the compression map of the compressed CSI image entering the channel.

[0050] Figure 4 The image generated by the CSI matrix is ​​put into the high-resolution recovery image model to obtain the recovered Recover image.

[0051] Figure 5 This is a comparison chart of the real part of the output results.

[0052] Figure 6 This is a comparison chart of the imaginary part of the output results.

[0053] Figure 7 This is a comparison chart of the module values ​​of the output results.

[0054] exist Figure 5-7 In the example, target is the target and recover is the recovery value. DETAILED DESCRIPTION

[0055] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] A CSI matrix compression feedback method based on image processing, the basic network structure of the image recovery model is:

[0057] a) Input embedding layer: maps the input image to a high-dimensional feature space. This mapping is usually achieved through a convolutional layer or a linear layer.

[0058] b) Encoder: The encoder consists of multiple Transformer layers, each of which contains a multi-head self-attention machine and a feedforward neural network. The encoder is responsible for extracting multi-scale features of the input image.

[0059] c) Decoder: The decoder structure is similar to the encoder. It is mainly used to restore the features extracted by the encoder to the image space. The decoder is also composed of multiple Transformer layers and uses jump connections to fuse features.

[0060] d) Output layer: The output layer maps the features recovered by the decoder back to the original image space to generate the final restored image.

[0061] like Figure 1 As shown, the specific steps of the compression feedback method are as follows:

[0062] Step 1: Figure 2 As shown, assuming that the CSI matrix is:

[0063]

[0064] Step 2: The CSI matrix is ​​a complex matrix with m rows and n columns. Read each element in the matrix A row by row and column by column to extract the real part a in the CSI matrix. mn and the imaginary part b mn , and using the modulus formula To calculate the modulus value corresponding to each element in the CSI matrix;

[0065] Step 3: Take the maximum and minimum values ​​of the real elements, and perform non-negative processing on the minimum values ​​of all real elements, that is, add 0 to all elements and subtract the minimum value to ensure that all real elements are non-negative values, so as to carry out the next step of RGB image conversion, and use normalization to scale all values ​​of real elements to the range of 0 to 255;

[0066] Step 4: Take the maximum and minimum values ​​of the imaginary elements, and perform non-negative processing on the minimum values ​​of all imaginary elements to ensure that all imaginary elements are non-negative values ​​for the next step of RGB image conversion. Use normalization to scale all values ​​of the imaginary elements to the range of 0 to 255.

[0067] Step 5: Take the maximum and minimum values ​​of the modulus elements, and perform non-negative processing on the minimum values ​​of all modulus elements to ensure that all modulus elements are non-negative values, so as to carry out the next step of RGB image conversion, and use normalization to scale all values ​​of the modulus elements to the range of 0 to 255;

[0068] Step 6. Create a new three-dimensional image matrix, put the two-dimensional matrix of real elements into the first dimension of the new three-dimensional image matrix, put the two-dimensional matrix of imaginary elements into the second dimension of the new three-dimensional image matrix, put the two-dimensional matrix of modulus elements into the third dimension of the new three-dimensional image matrix, and convert the three-dimensional image matrix into an unsigned 8-bit integer, leaving only positive integers in the range of 0 to 255, and discarding the decimal data after the integer.

[0069] Step 7. Generate a .png image using JPEG compression in cv2. The JPEG image quality factor used here is set to 80 (usually the quality factor ranges from 0 to 100. The quality factor determines how much information is lost during the compression process. A lower QF value will result in a higher compression ratio and a smaller file size, but the image quality will be significantly reduced, and blurring and color blocks may appear. A higher QF value can retain more image details, but the image size will increase accordingly. Choosing 80 as the QF value can usually maintain relatively good image quality while obtaining a moderate file size);

[0070] The three-dimensional matrix in step six is ​​converted and saved as a .png image. The compressed CSI image is transmitted through the channel. The specific compression diagram is as follows Figure 3 As shown;

[0071] Step 8: Put the image generated by the CSI matrix into the high-resolution recovery image model to obtain the recovered Recover image. The specific image is as follows: Figure 4As shown, the normalized ratio and non-negative offset are used to restore the two image matrices into real matrix, imaginary matrix and modulus matrix. In the restoration process, the real matrix and imaginary matrix have been obtained. All elements of the imaginary matrix are multiplied by the imaginary unit i, and all corresponding elements of the real matrix and the processed imaginary matrix are added. The real matrix and the imaginary matrix can be used to restore the original matrix.

[0072] In step three, the specific formula for normalization is: 255 / (MAX-MIN), where MAX is the maximum value and MIN is the minimum value.

[0073] In step eight, the training steps of the high-resolution restored image model are:

[0074] Step 1), putting the JPEG compressed RGB image set generated based on the CSI matrix into step 2) to train a high-resolution image restoration model;

[0075] Step 2) Read the model and set the yml file, including the dataset path, model name, etc.

[0076] Step 3), using the settings to read the status information of the trained model, and further deepening the training according to the status information of the trained model;

[0077] Step 4), initialize the training log;

[0078] Step 5), create a training data reader and a validation data reader;

[0079] Step 6) Find the model class in the settings and load it into the model object, including network structure information, loss function, learning rate, etc.

[0080] Step 7), create a log generator and initialize the data set pre-reader;

[0081] Step 8), use the setting information in step 2) to initialize the training iterator and training batch size, and enter the training loop;

[0082] Step 9) Initialize the learning rate according to the set parameters, fix the learning rate of the first cycle, the fixed learning rate of the first cycle is 0.0003, and lasts for 9200 iterations. In the second cycle, the cosine annealing strategy is used, and the learning rate is reduced from 0.0003 to 0.000001, which lasts for 20800 iterations.

[0083] Step 10), using the low-quality compressed image to progressively train and approximate the original image, and updating the batch of the progressively approximated low-quality images;

[0084] Step 11), the model reads the updated batch image data and passes in the low-quality image (lq) and the original image (gt);

[0085] Step 12), optimize the model parameters, call the net_g network forward propagation to obtain the result (that is, obtain the loss function), call the loss function back propagation to adjust the network parameters, the loss function uses the L1 loss function, the weight is 1, and the L1 formula is:

[0086]

[0087] Among them, x and y are two input tensors, N is the number of elements in the tensor, and ε is a small positive number used to avoid gradient explosion at zero point.

[0088] Step 13) Verify the model performance, such as loss, accuracy, PSNR (peak signal-to-noise ratio), SSIM (structural similarity), etc., and save the results to the image;

[0089] Step 14), read the next batch for training and return;

[0090] Step 15), save model parameters and information;

[0091] Step 16), pre-training model test, put a picture generated by a new CSI matrix into the pre-training model test to obtain a recovered Recover picture.

[0092] The test steps of the high-resolution restored image model are:

[0093] Step 21), load the yml setting file to obtain model information including network structure, pre-trained network parameters, etc., and set it to non-training mode;

[0094] Step 22), initialize the log system;

[0095] Step 23), initialize the test data set loader;

[0096] Step 24), read and create a model according to the setting information, read the network structure and pre-trained network parameters;

[0097] Step 25), read the test data set and enter the network to run, output the results, verify the performance, and save the results to the picture;

[0098] Step 26), extract the original data set image and a Recover image, compare the two data, and calculate the MSE (mean square error), output the comparison result graph, the real part comparison graph is as follows Figure 5 As shown, the imaginary part comparison diagram is as follows Figure 6 As shown in the figure, the module value comparison chart is as follows Figure 7Verify the recovery accuracy: the MSE (mean square error) of its real part matrix is ​​0.0002394341491162777, the MSE (mean square error) of the imaginary part matrix is ​​0.00015335251577198506, and the MSE (mean square error) of the modulus matrix is ​​0.00014324307441711427.

[0099] Matrix to RGB image: Create a new three-dimensional image matrix, put the two-dimensional matrix of the real part into the first dimension of the new three-dimensional image matrix, put the two-dimensional matrix of the imaginary part into the second dimension of the new three-dimensional image matrix, put the two-dimensional matrix of the modulus value into the third dimension of the new three-dimensional image matrix, and convert the three-dimensional image matrix into an unsigned 8-bit integer.

[0100] RGB color is called "true color". It is the default color used in Photoshop and is also the most commonly used color mode. RGB mode images are composed of three color channels, namely red channel (Red), green channel (Green) and blue channel (Blue). Each channel uses 8-bit color information. The value range of each color is 0 to 255. The combination of these three channels can produce more than 16.7 million different colors.

[0101] PNG is a lossless compression web page format that combines the advantages of GIF and JPEG formats. It is not only lossless compression, smaller in size, but also supports transparency and alpha channels. Since PNG is not fully applicable to all browsers, it is less used than GIF and JPEG in web pages. However, with the development of the network and the improvement of Internet transmission speed, PNG will be a standard image format used in future web pages.

[0102] JPEG format is a lossy compression web page format that does not support alpha channels or transparency. Its biggest feature is that the file is relatively small and can be compressed at a high rate, so it is widely used in fields that focus on file size. For example, images in the web page production process, such as banner ads, product images, and larger illustrations, can all be saved in JPG format.

[0103] JPEG Compression:

[0104] Step 1: Image segmentation;

[0105] In the first step of the JPEG algorithm, the image is divided into small blocks of size 8X8, and these small blocks are processed separately during the entire compression process.

[0106] Step 2: Color space conversion RGB->YCbCr;

[0107] The so-called "color space" refers to a mathematical model for expressing color. For example, the common "RGB" model decomposes color into three components: red, green, and blue. In this way, a picture can be decomposed into three grayscale images. Mathematically, each 8X8 pattern can be expressed as three 8X8 matrices, and the numerical range of the matrices is generally between [0,255].

[0108] Different color models have different application scenarios. For example, the RGB model is suitable for self-luminous patterns such as displays. In the printing industry, ink is used for printing, and the color of the pattern is generated when reflecting light. The CMYK model is usually used. In the JPEG compression algorithm, the pattern needs to be converted into the YCbCr model, where Y represents brightness (Luminance), and Cb and Cr represent the "color difference values" of green and red respectively.

[0109] Step 3: Discrete cosine transform;

[0110] When the present invention is no longer dealing with a function, but a bunch of discrete data, and if these data are symmetrical, then the function transformed by Fourier transformation only contains cosine terms, and this transformation is called discrete cosine transform. The original data Xi can be transformed by the inverse transform (IDCT) of discrete cosine transform. In other words, after DCT transformation, an array can be decomposed into the sum of several arrays. If the array is regarded as a one-dimensional matrix, the result can be regarded as the sum of a series of matrices. In the JPEG compression process, after the conversion of the color space, each 8X8 image block is represented as 3 8X8 matrices in data.

[0111] Step 4: Data quantification;

[0112] Through discrete cosine transform, although the image data has been completely changed, it is still in a "reversible" state, which means it has not yet entered the "lossy" stage. In the case of losing some precision, quantization is used to store these floating point numbers in less space.

[0113] Step 5: Huffman coding;

[0114] The last step of JPEG compression is to perform Huffman encoding on the data. Huffman is the basis of almost all compression algorithms. Its basic principle is to adjust the encoding length of the element according to the usage frequency of the element in the data to obtain a higher compression ratio.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A CSI matrix compression feedback method based on image processing, characterized in that: The specific steps are as follows: Step 1: Assume that the CSI matrix is: Step 2: The CSI matrix is ​​a complex matrix with m rows and n columns. Read each element in the matrix A row by row and column by column to extract the real part a in the CSI matrix. mn and the imaginary part b mn , and using the modulus formula To calculate the modulus value corresponding to each element in the CSI matrix; Step 3: Take the maximum and minimum values ​​of the real elements, perform non-negative processing on the minimum values ​​of all real elements to ensure that all real elements are non-negative values, and use normalization to scale all values ​​of the real elements to the range of 0 to 255; Step 4: Take the maximum and minimum values ​​of the imaginary elements, perform non-negative processing on the minimum values ​​of all imaginary elements to ensure that all imaginary elements are non-negative values, and use normalization to scale all values ​​of the imaginary elements to the range of 0 to 255; Step 5: Take the maximum and minimum values ​​of the modulus elements, perform non-negative processing on the minimum values ​​of all modulus elements to ensure that all modulus elements are non-negative values, and use normalization to scale all values ​​of the modulus elements to the range of 0 to 255; Step 6: Create a new three-dimensional image matrix, put the two-dimensional matrix of real elements into the first dimension of the new three-dimensional image matrix, put the two-dimensional matrix of imaginary elements into the second dimension of the new three-dimensional image matrix, put the two-dimensional matrix of modulus elements into the third dimension of the new three-dimensional image matrix, and convert the three-dimensional image matrix into an unsigned 8-bit integer; Step 7: Convert the three-dimensional matrix in step 6 and save it as a .png image. The compressed CSI image is transmitted through the channel. Step 8: Put the image generated by the CSI matrix into the high-resolution recovery image model to obtain the recovered Recover image, and use the normalized ratio and the non-negative offset to restore the two image matrices into a real matrix, an imaginary matrix, and a modulus matrix; During the recovery process, the real matrix and the imaginary matrix have been obtained. All elements of the imaginary matrix are multiplied by the imaginary unit i, and all corresponding elements of the real matrix and the processed imaginary matrix are added. The real matrix and the imaginary matrix can be used to restore the original matrix.

2. The CSI matrix compression feedback method based on image processing according to claim 1, characterized in that: In step three, the specific formula for normalization is: 255 / (MAX-MIN), where MAX is the maximum value and MIN is the minimum value.

3. The CSI matrix compression feedback method based on image processing according to claim 1, characterized in that: In step eight, the training steps of the high-resolution restored image model are: Step 1), putting the JPEG compressed RGB image set generated based on the CSI matrix into step 2) to train a high-resolution image restoration model; Step 2) Read the model and set the yml file; Step 3), read the status information of the trained model, and perform further in-depth training based on the status information of the trained model; Step 4), initialize the training log; Step 5), create a training data reader and a validation data reader; Step 6) Find the model class in the settings and load it into the model object; Step 7), create a log generator and initialize the data set pre-reader; Step 8), use the setting information in step 2) to initialize the training iterator and training batch size, and enter the training loop; Step 9) Initialize the learning rate according to the set parameters, fix the learning rate in the first cycle, and continue to iterate. In the second cycle, use the cosine annealing strategy, change the learning rate, and continue to iterate; Step 10), using the low-quality compressed image to progressively train and approximate the original image, and updating the batch of the progressively approximated low-quality images; Step 11), read the updated batch image data, and input the low-quality image and the original image; Step 12) Optimize the model parameters, call the network forward propagation to obtain the results, call the loss function back propagation to adjust the network parameters, the loss function uses the L1 loss function, the weight is 1, and the L1 formula is: Among them, x and y are two input tensors, N is the number of elements in the tensor, and ε is a positive number; Step 13) Verify the model performance and save the results to the image; Step 14), read the next batch for training and return; Step 15), save model parameters and information; Step 16), pre-training model test, put the image generated by the CSI matrix into the pre-training model test to obtain the recovered Recover image.

4. The CSI matrix compression feedback method based on image processing according to claim 3, characterized in that: The test steps of the high-resolution restored image model are: Step 21), load the yml setting file to obtain model information, and set it to non-training mode; Step 22), initialize the log system; Step 23), initialize the test data set loader; Step 24), read and create a model according to the setting information, read the network structure and pre-trained network parameters; Step 25), read the test data set and enter the network to run, output the results, verify the performance, and save the results to the picture; Step 26), extract the image of the original data set and a Recover image, compare the two data, calculate the MSE (mean square error), and output the comparison result graph.

Images