A Blind Frame Synchronization Recognition Method Based on Scaling Factor and RGB Image Fusion

Abstract

This invention discloses a blind frame synchronization identification method based on scaling factor and RGB image fusion. The implementation steps include: constructing an analysis matrix with a column count similar to the frame length based on the data distribution pattern of the frame synchronization code in the code domain, and generating grayscale images; calculating the scaling factor for each grayscale image; generating RGB images using the grayscale images and scaling factors; dividing the RGB images into a training set proportionally; training a frame length identification network using the training set; identifying the frame length using the trained frame length identification network; reconstructing the analysis matrix using the frame length estimation results for the code domain data; identifying the frame start point using the first-order cumulant method; and identifying the frame synchronization code using the frame start point. This invention has the advantages of being suitable for blind frame synchronization identification in complex non-cooperative communication environments, reducing identification complexity, and improving error resistance.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and more specifically relates to a blind frame synchronization identification method based on scaling factor and RGB image fusion in the field of non-cooperative frame synchronization communication technology. This invention can be used in complex non-cooperative communication environments to reduce identification complexity and improve the error resistance of blind identification methods. Background Technology

[0002] In digital communication, data is often transmitted in units of "frames." A data frame consists of two parts: a frame synchronization code and an information sequence. The frame synchronization code is a fixed sequence used to distinguish between adjacent data frames, while the information sequence includes the user data to be transmitted. The sending end groups the user data and assembles it into frames according to a certain format. The receiving end distinguishes different data frames and reconstructs them into user data information; this process is called frame synchronization. In non-cooperative communication systems, the receiving end is unaware of the frame synchronization information. Frame synchronization methods without any prior information are called "blind frame synchronization code identification methods." Traditional blind frame synchronization identification techniques typically rely on the autocorrelation characteristics of the synchronization codewords and require multiple sliding traversal calculations to detect the frame synchronization code. These traditional methods perform well in low-error environments, but in high-error environments, the autocorrelation characteristics may be affected, resulting in poor performance of traditional blind identification techniques.

[0003] In his paper "Frame Synchronization Recognition Based on Correlation Filtering and Discreteness Analysis" (Journal of Detection and Control, 2019), Shao Kun proposed a blind frame synchronization recognition method that utilizes correlation filtering to correct threshold values ​​and discreteness analysis to analyze synchronization codes. This method first constructs a padding matrix to initially eliminate the influence of redundant data on frame length recognition; then, it adaptively adjusts the threshold value using cubic correlation filtering to complete frame length recognition; next, it extracts key fields using correlation values; finally, it separates the synchronization code from fixed fields using discreteness analysis to achieve accurate recognition of the synchronization code. While this method solves the problem of difficult threshold selection, it still has shortcomings. When identifying frame length, it requires adaptive threshold adjustment using cubic correlation filtering for threshold decision-making and is limited by the autocorrelation characteristics of the frame synchronization code, resulting in high recognition complexity and a lack of universality.

[0004] In her paper "A Low-Complexity Blind Frame Synchronization Recognition Method" (Proceedings of the 14th National Conference on Signal and Intelligent Information Processing and Applications, 2021), Chen Xiaofang proposed a low-complexity blind frame synchronization recognition method for non-cooperative communication. This method first constructs matrices with different numbers of columns from the received sequence and uses row accumulation and other calculations to obtain the frame length existence probability of each matrix. By analyzing the frame length existence probabilities of all matrices, an estimated value for the frame length is obtained. Then, the matrix is ​​reconstructed using this frame length estimate, and the row accumulation results are analyzed to obtain an estimated value for the frame start point and a coarse estimate for the frame synchronization code. Finally, the strong autocorrelation characteristic of the frame synchronization code is used to analyze the coarse estimate of the frame synchronization code, thus obtaining the frame synchronization code. Although this method can solve the problem of excessive algorithm complexity, it still has shortcomings. The results show that when the bit error rate (BER) is below 0.2, the blind frame synchronization recognition performance is above 95%, but the recognition performance under noisy channels or high BER conditions is still poor, thus limiting its application in high BER environments. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technologies by providing a frame synchronization blind identification method based on scaling factor and RGB image fusion, which can reduce the identification complexity and improve error resistance in non-cooperative communication high-error environments.

[0006] The specific approach to achieving the objective of this invention is as follows: This invention constructs analysis matrices with different column numbers by analyzing the data distribution patterns of frame synchronization codes, and converts these matrices into grayscale images, thereby more clearly revealing features that are not readily apparent in the data domain. When the number of columns in the matrix is ​​an integer multiple of the frame length, the relative positions of the synchronization codes in each row of the matrix are the same. When the matrix is ​​converted into an image, obvious vertical stripes will appear at the positions of the synchronization code columns. When the number of columns in the matrix is ​​close to an integer multiple of the frame length, obvious slanted stripes will appear in the image. When the number of columns in the matrix differs significantly from the frame length, the element values ​​in the matrix tend to be randomly distributed, and the stripe features in the image are not obvious. This solves the problem in existing technologies where a threshold value still needs to be set during frame length recognition, and the lack of universality is due to the limitation of the synchronization code type. After estimating the frame length using a neural network, this invention constructs an analysis matrix with the frame length as the number of columns and combines it with the first-order cumulant method to achieve frame start point and synchronization code recognition. By constructing a matrix with a column count close to the frame length and building RGB images by calculating a scaling factor, the dataset is divided using these RGB images, and a frame length recognition network is trained to obtain an estimated frame length. This frame length is then used as a basis to identify parameters such as the frame start point and synchronization code. This solves the problems of poor frame length recognition performance in high-error-rate environments and the difficulty in recognizing other parameters in existing technologies.

[0007] The method for achieving the objective of this invention includes the following steps:

[0008] Generate grayscale images using an analysis matrix with a column count similar to the frame length; calculate the scaling factor for each grayscale image; generate RGB images using the grayscale images and scaling factors; train a frame length recognition network using the training set generated from the RGB images; identify the frame length using the trained frame length recognition network; identify the frame start point using the first-order cumulant method; and identify the frame synchronization code using the frame start point.

[0009] Compared with the prior art, the present invention has the following advantages:

[0010] First, because this invention analyzes the data distribution pattern of the frame synchronization code and generates grayscale images from an analysis matrix with a column count similar to the frame length, it can more clearly reveal features that are not obvious in the data domain. Then, it calculates the scaling factor for each grayscale image and uses the grayscale images and scaling factors to generate RGB images. This overcomes the shortcomings of existing technologies, such as the need to set a threshold value during frame length recognition and the limitation imposed by the autocorrelation characteristics of the frame synchronization code, resulting in a lack of universality. This invention eliminates the need to set a threshold value during frame length recognition, reducing recognition complexity and enabling effective frame length recognition under different frame length conditions.

[0011] Secondly, because this invention constructs a training set using RGB images and trains a frame length recognition network, it combines a first-order cumulant method to identify the frame start point and uses the frame start point to identify the frame synchronization code. This overcomes the problems of poor frame length recognition performance in high-error-rate environments and the difficulty in identifying other parameters in existing technologies. Therefore, this invention enables effective blind recognition of frame synchronization in high-error-rate environments. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method of the present invention;

[0013] Figure 2 This is the vertical stripe feature map that appears when the number of columns in the analysis matrix is ​​equal to or an integer multiple of the frame length. Figure 2 (a) is the feature map of vertical stripes that appear when the number of columns equals the frame length. Figure 2 (b) is the vertical stripe feature map that appears when the number of columns is an integer multiple of the frame length;

[0014] Figure 3 This is a feature map of diagonal stripes that appears when the number of columns in the analysis matrix is ​​near the frame length, as described in this invention. Figure 3 (a) is the feature map of diagonal stripes that appears when the number of columns in the matrix is ​​less than the frame length. Figure 3 (b) is the diagonal stripe feature map that appears when the number of matrix columns is greater than the frame length;

[0015] Figure 4 These are sample images from a portion of the dataset of this invention, wherein... Figure 4 (a) is a positive sample image. Figure 4(b) is a negative sample image;

[0016] Figure 5 This is the analysis matrix diagram reconstructed by the present invention using frame length identification as the column number;

[0017] Figure 6 This is a confusion matrix diagram of the frame length recognition network of the present invention for the recognition of the validation set;

[0018] Figure 7 This is a confusion matrix diagram of the frame length recognition network of the present invention for the test set. Figure 7 (a) is the confusion matrix after excluding negative samples. Figure 7 (b) is the confusion matrix diagram of the final identification results;

[0019] Figure 8 This is a comparison chart of the frame length recognition results of the present invention and a low-complexity blind frame synchronization recognition method;

[0020] Figure 9 This is a diagram showing the identification results of frame length, frame start point, and synchronization code according to the present invention.

[0021] Figure 10 This is a comparison chart of the recognition results of the present invention and a low-complexity frame synchronization blind recognition method. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Reference Figure 1 The implementation steps of the embodiments of the present invention will be further described below.

[0024] Step 1: Generate a grayscale image from an analysis matrix with a number of columns similar to the frame length.

[0025] The steps for generating a grayscale image are as follows:

[0026] The first step is to generate multiple n×n dimensional analysis matrices with element values ​​of 1 or 0 by analyzing the distribution pattern of frame synchronization codes in wireless digital communication, where (L-3)≤n≤(L+3), L is the frame length in wireless digital communication, and the range of n indicates that the number of columns in the analysis matrix is ​​similar to L.

[0027] The second step is that when the corresponding synchronization symbols in each frame of the digital sequence in wireless digital communication are not in the same column, the first-order cumulant method has failed. However, there are still relatively obvious tilted stripe features in the image. Therefore, we can use the analysis matrix to generate grayscale images with stripe features to express different degrees of tilt. Each grayscale image corresponds to an analysis matrix. In the grayscale image, white corresponds to the element value of 1 in the analysis matrix, and black corresponds to the element value of 0 in the analysis matrix.

[0028] When the number of columns in the matrix is ​​equal to or an integer multiple of the frame length, vertical stripes will appear in the grayscale image, such as... Figure 2 As shown, when the number of columns equals the frame length, the vertical stripe feature appears as follows: Figure 2 As shown in (a), the vertical stripe feature appears when the number of columns is an integer multiple of the frame length, such as Figure 2 As shown in (b), when the number of columns in the matrix approaches the frame length, obvious diagonal stripes will appear in the grayscale image, such as... Figure 3 As shown, when the number of columns in the matrix is ​​less than the frame length, each row of the data matrix is ​​insufficient to store a complete frame. The excess portion of the same frame is filled into the next row of the matrix, thus shifting the frame synchronization code of the next frame backward, as shown below. Figure 3 As shown in (a), when the number of columns in the matrix is ​​greater than the frame length, the data in one physical frame is insufficient to fill a single row of the data matrix. The data in the next frame is shifted forward, as shown in (a). Figure 3 As shown in (b).

[0029] Step 2: Calculate the scaling factor for each grayscale image.

[0030] Because the input to a neural network is deterministic and uniform, images need to be compressed to the same size before being fed into the network. This compression process inevitably leads to the loss of image information. When the neural network classifies and recognizes the scaled samples, confusion and misjudgment will occur.

[0031] To solve this problem, additional parameters are needed to compensate for information loss during image scaling. Therefore, the concept of a scaling factor is introduced, and the scaling factor is calculated as follows:

[0032]

[0033] Where, β i L represents the scaling factor of the i-th grayscale image. min The width of the image represents the minimum width among all grayscale images, d represents an integer selected from [0, |nL|] based on the image size, and L represents the frame length in wireless digital communication. max This represents the width value of the largest grayscale image.

[0034] Step 3: Generate an RGB image using the grayscale image and a scaling factor.

[0035] The process of generating an RGB image involves filling each pixel value of a grayscale image into the corresponding position in the R and G channels of an RGB image, and filling the scaling factor of each image into each pixel position in the B channel of the corresponding RGB image, thus obtaining an RGB image.

[0036] At this point, each pixel value in the R and G channels of the RGB image is the same as the corresponding element in the analysis matrix, while each pixel value in the B channel is equal to β. Therefore, after the bilinear interpolation operation during the scaling process, the original pixel values ​​in the R and G channels will be stretched, which may lead to blurring or distortion, while each pixel value in the B channel remains unchanged, so the image features are preserved.

[0037] Step 4: Construct a training set based on the RGB images and train the frame length recognition network.

[0038] The steps for constructing the training set are as follows:

[0039] The first step is to copy all the RGB images 10 times, and add a bit error rate of 0.02k to the k-th copy image, where k = 1, 2, ..., 10.

[0040] The second step is to label the physical frame length of each copied image to obtain N frame length categories.

[0041] The third step involves labeling each RGB image whose total column value in each analysis matrix is ​​close to the actual frame length value of wireless digital communication as a positive sample, and labeling RGB images whose column values ​​differ significantly from the frame length value as negative samples. All positive and negative samples and their corresponding labels are then combined to form a training set. Some positive sample images are shown below. Figure 4 As shown in (a), the partial negative sample image is as follows: Figure 4 As shown in (b), the positive sample of the image has stripe features, and each row contains at most one complete physical frame. Due to the presence of bit errors, there are scattered spots in the stripes.

[0042] The frame length recognition network uses a ResNet18 residual network, which consists of 17 convolutional layers and 1 fully connected layer connected in series. The kernel size of each convolutional layer is set to 3×3, and the number of output nodes of the fully connected layer is equal to the number of label categories in the training set.

[0043] The training of the frame length recognition network involves inputting the training set into the network and iteratively updating the network parameters using a stochastic gradient descent algorithm until the network's cross-entropy loss function reaches its minimum value, thus obtaining a trained frame length recognition network model. By using the cross-entropy loss function, the model can measure the difference between its prediction results and the actual labels during training, thereby improving the model's performance.

[0044] Step 5: Use the trained frame length recognition network to identify the frame length;

[0045] The frame length identification network identifies frame lengths using the following steps:

[0046] The first step is to convert the intercepted code field data in the wireless digital communication into Q-code. i ×u i A 3D RGB image, where... T∈N and T≤2d,u i =L min +T×(i-1),i={1,2,…,Q};

[0047] The second step is to put Q into a box. i ×u i A 3D RGB image is input into a trained frame length recognition network to obtain the frame length recognition result for each RGB image.

[0048] The third step is to count the number of images identified as frame length t, Num_j, and the maximum probability P_j, where j = 1, 2, ..., N, and N represents the total number of frame length categories. We select Num = max{Num_j} and count the maximum recognition probability P_m for all frame length categories where Num_j = Num, where m is the number of image categories that satisfy Num_j = Num. The frame length corresponding to the highest probability value in P_m is selected as the recognition category z.

[0049] Step 6: Identify the frame start point using the first-order cumulant method.

[0050] The steps for identifying the frame start point using the first-order cumulant method are as follows:

[0051] The first step is to construct an h×g dimensional analysis matrix with elements of either 1 or 0, where the value of g is equal to the identified category z. M represents the total length of code field data intercepted in at least 30 minutes of wireless digital communication. Since the total number of columns g in the analysis matrix equals the identification category z, the frame synchronization code in the analysis matrix appears at a fixed position in each row, such as... Figure 5 As shown, Figure 5 The position highlighted in red in the analysis matrix indicates the location of the frame synchronization code in the analysis matrix. The row vector is obtained by summing the element values ​​in all rows of the analysis matrix. Each element value s in the row vector is a natural number not greater than h.

[0052] The second step is to consider the impact of noise in the channel, set the error tolerance for the data, and determine the value of each element in the frame structure vector of the analysis matrix according to the characteristics of the distribution pattern of the frame synchronization code in the row vector, using the following formula:

[0053]

[0054] Among them, H y1 v represents the value of the y1-th element in the frame structure vector. y This represents the value of the y-th element in the row vector.

[0055] The third step is to search for the "011" or "111" field in the frame structure vector, and take the position of the first field found in the first search as the starting point H of the frame. s Search the "110" field in the frame structure vector, and take the position of the first field found in the first search as the synchronization code end point H of that frame. e .

[0056] Step 7: Identify the frame synchronization code using the frame start point.

[0057] The steps for identifying the frame synchronization code are as follows:

[0058] The first step is to subtract the start point from the end point of the synchronization code of each frame to obtain the synchronization code length of that frame.

[0059] The second step is to take the starting point of each frame as the starting position of the frame in each row of the analysis matrix, and the synchronization code length of that frame as the length of the frame synchronization code in each row of the analysis matrix. The frame synchronization code that appears most frequently in the analysis matrix is ​​taken as the frame synchronization code estimation result.

[0060] The effects of the present invention will be further explained below with reference to simulation experiments.

[0061] 1. Simulation experimental conditions:

[0062] The hardware platform for the simulation experiment of this invention is as follows: the processor is an Intel i5 12500H CPU with a main frequency of 2.5GHz, the memory is 40GB, the GPU is an NVIDIA RTX 3060 Laptop with 6GB of video memory.

[0063] The software platform for the simulation experiment of this invention is: Windows 11 operating system, Python 3.8, and Matlab R2021b.

[0064] The input image used in the simulation experiment of this invention is an RGB image with a frame synchronization code distribution pattern. This data is constructed by an analysis matrix that matches the frame length and a calculated scaling factor. The image size is n×n×3 pixels, n∈[40,100]. The image contains 5 frame length values ​​and the image format is png.

[0065] 2. Simulation content and result analysis:

[0066] The simulation experiment of this invention uses this invention and existing technology (ResNet18 network classification method) to classify an input RGB image with a frame synchronization code distribution pattern, and obtains the classification result, such as... Figure 5 As shown.

[0067] In simulation experiments, one existing technology used is:

[0068] The existing ResNet18 network classification method refers to the residual network classification method proposed by Kaiming He et al. in their paper "Deepresidual learning for image recognition" ([C] / / Proceedings of the IEEE conference on computer vision and pattern recognition.2016:770-778), abbreviated as ResNet18 network classification method.

[0069] The following is combined Figure 6-10 The simulation results further illustrate the effects of the present invention.

[0070] Figure 6 This is a confusion matrix diagram drawn by the network model after identifying the validation set. Figure 6 The horizontal axis represents the class predicted by the network, the vertical axis represents the true class, and the dark blue values ​​on the diagonal represent the probabilities after row-direction normalization, excluding negative samples. Figure 6 As can be seen, the network model has a good classification ability for images with a matching number of columns and frame length, and the erroneous samples are basically misclassified as "NS" category, avoiding confusion between categories. In addition, very few images with negative sample categories are identified as positive sample categories.

[0071] Figure 7 This is a confusion matrix diagram drawn by the network model after identifying the test set at a bit error rate of 0.2. Figure 7 The horizontal axis represents the category predicted by the network, and the vertical axis represents the true category. Figure 7 The dark blue values ​​on the diagonal represent the row-direction normalized probabilities after excluding negative samples. The confusion matrix after excluding negative samples and the confusion matrix of the final identification result after decision analysis are shown below. Figure 7 (a) Figure 7 As shown in (b). Figure 7 (a) is the confusion matrix after excluding negative samples. Figure 7 (b) is the confusion matrix diagram of the final identification results. (Comparison) Figure 7 (a) Figure 7 (b) It can be seen that the method of the present invention significantly improves the recognition effect under high error rate, and the recognition probability of all categories is above 98%.

[0072] The method of this invention was compared with a low-complexity frame synchronization blind recognition method, using the same experimental data. 200 simulation experiments were conducted at different bit error rates. Under the same experimental data, the recognition performance was as follows: Figure 8 As shown.

[0073] Figure 8 The horizontal axis represents the bit error rate, and the vertical axis represents the network's recognition accuracy. The blue line represents the frame length recognition performance curve of this invention, and the red line represents the frame length recognition performance curve of the low-complexity frame synchronization blind recognition method. It can be seen that when the bit error rate is 0.2%, the frame length recognition performance of this invention can be improved by 10% compared to the low-complexity frame synchronization blind recognition method.

[0074] The analysis matrix, with the frame length as the column number, is reconstructed using the estimated frame length. Then, the start and end points of the frame synchronization code are identified using the first-order cumulant method, completing the blind identification of frame synchronization. The identification results are as follows: Figure 9 As shown.

[0075] Figure 9 The horizontal axis represents the bit error rate, and the vertical axis represents the network's recognition accuracy. Figure 9 The blue line represents the performance curve of frame length blind identification in this invention, the red line represents the performance curve of frame start point blind identification in this invention, and the yellow line represents the performance curve of frame synchronization code blind identification in this invention.

[0076] 200 random experiments were conducted at different bit error rates. Under identical experimental data, the method of this invention, compared with the low-complexity blind frame synchronization identification method, achieves the following comprehensive frame synchronization identification results: Figure 10 As shown.

[0077] Figure 10 The horizontal axis represents the bit error rate, and the vertical axis represents the network's recognition accuracy. Figure 10 The blue line represents the recognition performance curve of the present invention, while the red line represents the recognition performance curve of the low-complexity frame synchronization blind recognition method. The recognition performance of the present invention is slightly better than that of the low-complexity frame synchronization blind recognition method.

[0078] This invention combines neural networks with first-order cumulants for blind frame synchronization recognition. Starting from the positional patterns of the frame synchronization code, a grayscale image with stripe features is generated, and a scaling factor is introduced to compensate for information loss caused by image scaling. Simultaneously, an RGB image is constructed using the grayscale image and the scaling factor. A frame length recognition network is trained using a predefined training set. This trained network identifies the frame length, achieving the recognition of the physical frame length. When the bit error rate is 0.2%, this method improves the accuracy of frame length recognition by 10% compared to the first-order cumulant method. Then, the analysis matrix is ​​reconstructed with the frame length as the column number, and the first-order cumulant method is combined to identify the frame start point, which is then used to identify the frame synchronization code.

[0079] The simulation experiments above show that the method of the present invention overcomes the problems of the existing technology, which still requires setting a threshold value in the frame length recognition process and is not universal due to the limitation of the synchronization code type. At the same time, it reduces the recognition complexity and has stronger anti-error performance than the traditional low-complexity frame synchronization blind recognition method, and has better fault tolerance under high error rate.

Claims

1. A frame synchronization blind recognition method based on scaling factor and RGB image fusion, characterized in that, Calculate the scaling factor for each grayscale image, and generate an RGB image using the grayscale image and the scaling factor; The steps of this recognition method include: generating grayscale images using an analysis matrix with a column count similar to the frame length; calculating the scaling factor for each grayscale image; generating RGB images using the grayscale images and scaling factors; training a frame length recognition network using the training set generated from the RGB images; recognizing the frame length using the trained frame length recognition network; recognizing the start point of the frame using the first-order cumulant method; and recognizing the frame synchronization code using the frame start point. The steps for generating a grayscale image are as follows: The first step is to generate multiple elements with values ​​of 1 or 0. 3D analysis matrix, where, L is the frame length in wireless digital communication, and the range of values ​​for n indicates that the number of columns in the analysis matrix is ​​similar to L. The second step is to use the analysis matrix to generate grayscale images with stripe features to represent different degrees of tilt. Each grayscale image corresponds to an analysis matrix, where white in the grayscale image corresponds to the element value of 1 in the analysis matrix, and black corresponds to the element value of 0 in the analysis matrix. The steps for identifying the frame start point using the first-order cumulant method are as follows: The first step is to construct a set of elements with values ​​of 1 or 0. An analysis matrix of dimension , where the value of g is equal to the identified category z. M represents the total length of code field data intercepted in at least 30 minutes of wireless digital communication. The row vector is obtained by summing the element values ​​in all rows of the analysis matrix. Each element value s in the row vector is a natural number not greater than h. The second step is to determine the value of each element in the frame structure vector of the analysis matrix according to the following formula: ； in, This represents the value of the y1-th element in the frame structure vector. This represents the value of the y-th element in the row vector; The third step is to search for the "011" or "111" field in the frame structure vector, and take the position of the first field found in the search as the starting point of the frame. Search the "110" field in the frame structure vector, and take the position of the first field found in the first search as the end point of the synchronization code for that frame. .

2. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The scaling factor is obtained by the following formula: ； in, This represents the scaling factor of the i-th grayscale image. This represents the minimum width value among all grayscale images. Indicates in A randomly selected integer within the range, This represents the width value of the largest grayscale image.

3. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The process of generating an RGB image involves filling each pixel value from a grayscale image into the corresponding position in the R and G channels of an RGB image, and filling the scaling factor of each image into each pixel position in the B channel of the corresponding RGB image, thus obtaining an RGB image.

4. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The steps for generating the training set are as follows: The first step is to copy all RGB images 10 times, and then add a bit error rate of 0.02k to the k-th copy image. ; The second step is to label the physical frame length of each copied image to obtain N frame length categories. The third step is to label each RGB image whose total value in each column of the analysis matrix is ​​close to the actual frame length value of the wireless digital communication as a positive sample, and label the RGB images whose column values ​​differ significantly from the frame length values ​​as negative samples. All positive and negative samples and their corresponding labels are then combined to form a training set.

5. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The frame length recognition network uses a ResNet18 residual network, which consists of 17 convolutional layers and 1 fully connected layer connected in series. The kernel size in each convolutional layer is set to a specific value. The number of output nodes of a fully connected layer is equal to the number of label categories in the training set.

6. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The training of the frame length recognition network involves inputting the training set into the frame length recognition network and iteratively updating the network parameters using the stochastic gradient descent algorithm until the network's cross-entropy loss function converges, thus obtaining a trained frame length recognition network model.

7. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The steps for identifying frame length using the trained frame length recognition network are as follows: The first step is to convert the intercepted code field data in the wireless digital communication into Q-sheets. A 3D RGB image, in which, , and , ; The second step is to transfer Q sheets. The RGB images are input into the trained frame length recognition network to obtain the frame length recognition result for each RGB image; The third step is to count the number of images identified as frame length t, Num_j, and the maximum probability P_j, where j=1,2…N, and N represents the total number of frame length categories. We select Num = max{Num_j} and count the maximum recognition probability P_m for all frame length categories where Num_j=Num, where m is the number of image categories that satisfy Num_j=Num. The frame length corresponding to the highest probability value in P_m is selected as the recognition category z.

8. The frame synchronization blind recognition method based on scaling factor and RGB image fusion according to claim 1, characterized in that, The steps for identifying the frame synchronization code are as follows: The first step is to subtract the start point from the end point of the synchronization code of each frame to obtain the synchronization code length of that frame. The second step is to take the starting point of each frame as the starting position of the frame in each row of the analysis matrix, and the synchronization code length of that frame as the length of the frame synchronization code in each row of the analysis matrix. The frame synchronization code that appears most frequently in the analysis matrix is ​​taken as the frame synchronization code estimation result.

Images