A data processing method and system based on Chinese character message transmission

Abstract

The present invention relates to the field of wireless transmission technology, and in particular to a data processing method and system based on Chinese character message transmission, the method comprising obtaining a CP16 signal data set covering a variety of Chinese characters and signal conditions, and inputting a target detection model for training, the target detection model comprising a multi-layer convolutional bidirectional long short-term memory network and a target recognition network; using the trained target detection model as a teacher model, compressing the target detection model to obtain a sub-model as a student model, and using the CP16 signal data set to train the sub-model to obtain a text recognition model; obtaining a CP16 signal, and inputting the CP16 signal into a text recognition model to identify the Chinese character information contained therein; and converting the identified Chinese character information into audio information. According to the scheme of the present invention, the text recognition model can effectively adapt to the embedded operating environment of the CP16 wireless receiver, and can accurately identify the text information corresponding to the CP16 signal.

Description

Technical Field

[0001] The present invention relates to the field of wireless transmission technology. More specifically, the present invention relates to a data processing method and system based on Chinese character message transmission. Background Art

[0002] CP16 (Chinese Pattern) Chinese character message transmission is a new domestic communication mode, and also a "cross-border" mode that transmits in digital mode and receives in analog mode. The CP16 system includes a transmitter and a receiver. The transmitter sends out a radio frequency signal of a certain frequency band, and after the receiver receives the radio frequency signal, it is displayed through a time-frequency waterfall (the horizontal axis is time and the vertical axis is frequency). The shape of Chinese characters is displayed in the form of a two-dimensional bright spot distribution in its interface.

[0003] CP16 can be thought of as 16 CW signals (i.e., "continuous wave") with similar transmission frequencies. It is a radio transmission method. CW signals transmit information through a stable sine wave of a single frequency. The display process of Chinese characters is that the opening and closing of a certain signal at different times represent whether 16 pixels of a row of a specific frequency of the 16×16 Chinese character dot matrix on the time-spectrum "waterfall diagram" are lit. The opening and closing of 16 signals of different frequencies at the same time correspond to whether 16 pixels of a column of the 16×16 Chinese character dot matrix on the time-spectrum "waterfall diagram" are lit at that moment. As time goes by, 16 different columns are displayed from left to right in turn, forming a complete Chinese character graphic, and finally the message is displayed in the form of a "flowing light" on the receiver's time-spectrum "waterfall diagram" (abbreviated as waterfall diagram). Of course, the receiver can also be a spectrum-time "waterfall diagram", which displays 16 rows of Chinese character dot matrix row by row from top to bottom as time passes.

[0004] The current CP16 system has the following main problems in emergency communications:

[0005] 1) Visual observation limitation: First, the CP16 system displays the received information in the form of Chinese characters at the receiving end, which requires the person receiving the information to observe the information visually. In an emergency communication environment, visual observation is severely limited by factors such as ambient light, distance, and obstructions, while sound wave transmission is much less restricted by the environment. In addition, visual observation requires the receiver to keep an eye on the display screen at all times, which is not suitable for use in emergency situations where other tasks need to be performed simultaneously.

[0006] 2) Signal strength differences and difficulty in adjustment: The signal strength of different channels varies greatly. For example, the difference between the 1st and 3rd channels from low to high channel frequency is 43dB, which is equivalent to using 1000W transmission power for the next channel and only 0.5W for the other channel at the same communication distance. This strength difference requires the receiving device to have a very high dynamic range, and the receiver needs to constantly adjust the receiving parameters to adapt to different signal strengths and interference situations. In an emergency communication environment, time is tight, and the receiving personnel may not have time to carefully adjust various receiving parameters according to the strength of the transmitted signal and the wireless environment of the receiving location.

[0007] In order for users to obtain accurate messages in a timely manner in emergency situations, designers try to obtain the corresponding text information by directly performing text recognition from the received CP16 signal, and convert it into a voice signal to achieve message acquisition in emergency situations. For example, the Chinese invention patent application with application number 202210000584.7 and invention name Text Recognition Model Training, Text Recognition Method, Device, Equipment and Medium discloses a method for training a text recognition model. In order to improve the accuracy of text recognition, it is necessary to train to reach the corresponding state. The corresponding model weights will occupy a large amount of storage space and require a system with strong computing power as a support. Based on this, when using this method to realize text information recognition, it is necessary to train a large multimodal language model or use a large model in the cloud, but this method is not applicable in field environments with poor communication quality.

[0008] Therefore, how to solve the problem of poor communication effect of the current CP16 system in emergency communication environment is the focus of current research. Summary of the invention

[0009] In order to solve the problem that the CP16 system has poor communication effect in an emergency communication environment, the present invention provides solutions in the following aspects.

[0010] In the first aspect, the present invention provides a data processing method based on Chinese character message transmission, comprising: obtaining a CP16 signal data set covering a variety of Chinese characters and signal conditions, and inputting it into a target detection model for training, wherein the target detection model comprises a multi-layer convolutional bidirectional long short-term memory network and a target recognition network, the multi-layer convolutional network bidirectional long short-term memory network is used to extract multi-scale spectral features and time features, and fuse the multi-scale spectral features and time features to obtain a comprehensive feature vector, the target recognition network performs feature mapping, dimension conversion and classification on the comprehensive feature vector to output a Chinese character recognition result; based on the knowledge distillation method, the trained target detection model is used as a teacher model, and the sub-model obtained after compressing the target detection model is used as a student model, and the sub-model is trained using the CP16 signal data set to obtain a text recognition model; the CP16 signal is obtained, and the CP16 signal is input into the text recognition model to identify the Chinese character information contained therein; and the identified Chinese character information is converted into audio information.

[0011] The present invention constructs a target detection model of a multi-layer convolutional bidirectional long short-term memory network in a CP16 system, constructs a training set in combination with a CP16 data set of multiple Chinese characters and signal conditions, and trains the target detection model. The trained large model is used as a teacher model, and the compressed sub-model is used as a student model. The design of a special model in a CP16 receiver is realized by knowledge distillation, which effectively reduces the size of the model and ensures the accuracy of text recognition, so that the text recognition model can effectively adapt to the operating environment of the CP16 wireless receiver, and fully utilizes the algorithm resources of the CP16 system, so that the CP16 receiver can directly perform effective text recognition from the received CP16 signal under limited computing resources, instead of the receiving personnel carefully adjusting the receiver parameters and then visually observing and identifying the text dot matrix on the time-frequency waterfall diagram, thereby improving the real-time performance and reliability of the CP16 system in an emergency communication environment. Furthermore, the present invention extracts global features and local detail features in the two dimensions of frequency and time and fuses multi-dimensional features to simulate the effect of the receiving personnel using their prior knowledge of various deformed and even incomplete text images to automatically repair and complete the dot matrix image of Chinese characters in the waterfall diagram in their minds. This can directly capture richer text details and structures from the received signal, thereby achieving a more accurate text recognition process.

[0012] Preferably, the learning rate of the target detection model training process gradually decays, and its calculation formula is:

[0013] ;

[0014] In the formula, represents the current learning rate, and Respectively represent the maximum and minimum number of iterations allowed, represents the initial learning rate, k represents the attenuation factor, t Indicates the current training cycle number, T Indicates the expected total number of training epochs.

[0015] By combining CP16 received signals with a variety of Chinese characters and signal conditions to construct a training set, the learning rate gradually decays during the model training process, ensuring the rapid convergence of the model during the iteration process, so that an effective balance can be achieved between model accuracy and system computing power.

[0016] Preferably, the multi-layer convolutional bidirectional long short-term memory network includes a feature extraction layer and a feature fusion layer, wherein the feature extraction layer includes a spectrum feature extraction layer and a time feature extraction layer, the spectrum feature extraction layer uses convolution kernels of different sizes to extract multi-scale spectrum features from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received; the time feature extraction layer uses a bidirectional long short-term memory network to extract forward time features and reverse time features from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received; the feature fusion layer fuses the multi-scale spectrum features and time features by means of feature weighted splicing to form a comprehensive feature vector.

[0017] Preferably, the feature fusion layer fuses the multi-scale spectral features and time features by means of feature weighted splicing, including: using an attention mechanism to determine the importance of the mutual influence between the multi-scale spectral features and the time features; using the importance to weight the multi-scale spectral features and the time features, and splicing the weighted multi-scale spectral features and the time features to form a comprehensive feature vector.

[0018] Preferably, the target recognition network includes a fully connected layer and an output layer. The fully connected layer adopts a multi-layer fully connected network. The comprehensive feature vector is input into the fully connected layer for mapping and dimension conversion to obtain the comprehensive feature vector after dimensionality reduction. The output layer is a Softmax layer, which uses the comprehensive feature vector after dimensionality reduction for classification and recognition, and outputs the Chinese character recognition result.

[0019] The present invention combines features of different scales together with emphasis by using a weighted splicing feature fusion method, so that the model can more effectively capture the global features and local detail features of the CP16 signal, thereby making the text content corresponding to the CP16 signal easier to recognize, thereby improving the accuracy and reliability of the text recognition process.

[0020] Preferably, obtaining a CP16 signal data set covering a variety of Chinese characters and signal conditions includes: adding one or more of frequency disturbance, time disturbance, and noise to the CP16 signal corresponding to each Chinese character.

[0021] By collecting a large amount of real CP16 Chinese message data, covering various signal strengths and interference conditions, the accuracy of model recognition results can be further improved, and the communication effect in outdoor environments can be improved.

[0022] Preferably, the method of compressing the target detection model includes one or more of the following methods: optimizing the structure of the target detection model: reducing the number of network layers, reducing the size of the convolution kernel, and reducing the number of channels; parameter quantization: quantizing the weights in the target detection model into low-bit-width fixed-point numbers or binary numbers; model pruning: determining the importance of the weights of each part in the target detection model, and removing the weights whose importance is lower than a set threshold; or determining the importance of channels or network layers in the target detection model, and removing the weights whose importance is lower than a set value.

[0023] By adopting a variety of ways to compress the model volume, a text recognition model specifically for the CP16 receiver operating environment can be designed, so that the corresponding text information can be directly obtained through the CP16 signal, achieving reliable communication in the field environment.

[0024] Preferably, the sub-model is trained using the CP16 signal data set, including: inputting the CP16 signal data set into the sub-model and performing iterative training; extracting intermediate feature quantities from at least one target layer in the sub-model, and calculating the difference between the intermediate feature quantities and the feature vector corresponding to the target detection model; calculating the loss during the iterative training process based on the difference, and adjusting the iteration step size based on the learning rate; in response to the loss being less than a set value or the number of iterations reaching a set value, stopping the training and outputting the trained sub-model as a text recognition model.

[0025] Preferably, the calculation formula of the loss function for calculating the loss is:

[0026] ;

[0027] ;

[0028] In the formula, represents the loss function, represents the distance between the probability distributions output by the teacher model and the student model, represents the loss function corresponding to the student model, represents the hyperparameter, represents the total number of training samples in the training data set, Indicates iThe true value corresponding to the training, Indicates i The predicted value corresponding to the training, Indicates j The intermediate feature vector during the training iteration, Indicates j The feature vector corresponding to the target detection model during the training iteration, n Indicates the number of iterations.

[0029] The present invention also designs the loss function in the model training process. Through the output of the teacher model and the student model, and the corresponding loss function of the student model, a comprehensive loss function can be obtained. Training is performed according to the loss function to effectively improve the training effect. The features corresponding to the standard text signal are used as a reference to compare the differences in feature quantities, thereby effectively improving the accuracy of the training results.

[0030] In a second aspect, the present invention also provides a data processing system based on Chinese character message transmission, comprising: a processor; a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the data processing method based on Chinese character message transmission described in one or more of the above embodiments is implemented.

[0031] The beneficial effects of the present invention are as follows: by constructing a target detection model such as a multi-layer convolutional bidirectional long short-term memory network in the CP16 system, and combining a CP16 signal with a variety of Chinese characters and signal conditions to construct a training set, the learning rate gradually decays during the model training process, ensuring the rapid convergence of the model during the iteration process, so that an effective balance can be achieved between the model accuracy and the computing power of the system. The text recognition model constructed according to the method of the present invention can effectively adapt to the operating environment of the CP16 wireless receiver, and fully utilizes the algorithm resources of the CP16 system, thereby improving the real-time performance and reliability of the CP16 system in an emergency communication environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0033] Figure 1 is a schematic diagram schematically illustrating an interference environment of a wireless communication system according to an embodiment of the present invention;

[0034] Figure 2 is a flow chart schematically illustrating a data processing method based on Chinese character message transmission according to an embodiment of the present invention;

[0035] Figure 3 is a flow chart schematically illustrating a method for training a target detection model according to an embodiment of the present invention;

[0036] Figure 4 It is a block diagram schematically showing the structure of the data processing system based on Chinese character message transmission according to this embodiment. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0038] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] The communication system in the outdoor environment is usually based on embedded systems, and its cost, volume, capacity and power consumption are all limited by the environment, which makes the system unable to participate in large parameter quantities and calculations. Then, in this case, it will be impossible to use big data deep neural networks for text recognition. In order to realize automatic text recognition at the CP16 receiving end, the present invention directly trains the teacher model with the received CP16 signal data, and uses knowledge distillation to train the text recognition model after the compressed volume, thereby effectively reducing the model volume while ensuring the accuracy of Chinese character recognition, so that the network model can also be effectively applied in embedded environments, improving the wireless communication effect in outdoor environments. Next, the scheme of the present invention will be described in detail in conjunction with the communication process of the CP16 system.

[0040] Figure 1 FIG. 4 is a schematic diagram schematically illustrating an interference environment of a wireless communication system according to an embodiment of the present invention.

[0041] like Figure 1As shown, the CP16 (Chinese Pattern) Chinese character message transmission system (hereinafter referred to as the CP16 system) includes a transmitter and a receiver. When the receiver receives the signal sent by the transmitter, it can pre-process the received signal and enhance the signal based on the software radio SDR (Software Defination Radio) and the waterfall chart software, so as to realize the display of the received signal in the waterfall chart. According to the context of the present invention, the present invention takes the CP16 transmitter and the CP16 receiver as an example to illustrate the scheme of the present invention. After receiving the wireless transmission signal sent by the transmitter, the receiver constructs a text recognition model dedicated to identifying the text information in the CP16 signal in the receiver, so as to realize automatic recognition and display of text in the field environment, so that the CP16 system gets rid of the problem of difficult interaction in the field environment and improves the communication effect. The method will be described in detail below.

[0042] Figure 2 FIG. 2 is a flow chart schematically showing a data processing method 200 based on Chinese character message transmission according to an embodiment of the present invention.

[0043] like Figure 2 As shown, at step S201, a CP16 signal data set covering a variety of Chinese characters and signal conditions is obtained and input into the target detection model for training. In some embodiments, the method of obtaining the CP16 signal data set covering a variety of Chinese characters and signal conditions includes: adding one or more of frequency disturbance, time disturbance, and noise to the CP16 signal corresponding to each Chinese character, thereby obtaining time series data of multiple signals and the Chinese character content corresponding to the data.

[0044] The target detection model includes a multi-layer convolutional bidirectional long short-term memory network and a target recognition network. The multi-layer convolutional bidirectional long short-term memory network is used to extract multi-scale spectral features and time features, and fuse the multi-scale spectral features and time features to obtain a comprehensive feature vector. The target recognition network performs feature mapping, dimension conversion and classification on the comprehensive feature vector to output a Chinese character recognition result.

[0045] The above-mentioned multi-layer convolutional bidirectional long short-term memory network includes a feature extraction layer and a feature fusion layer, wherein the feature extraction layer includes a spectrum feature extraction layer and a time feature extraction layer. The spectrum feature extraction layer extracts multi-scale spectrum features from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received; the time feature extraction layer adopts a bidirectional long short-term memory network to extract forward time features and reverse time features from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received.

[0046] The feature fusion layer fuses the multi-scale spectral features and the time features by means of feature weighted splicing to form a comprehensive feature vector. In an application scenario, the feature fusion layer fuses the multi-scale spectral features and the time features by means of feature weighted splicing, including: using the attention mechanism to determine the importance of the mutual influence between the multi-scale spectral features and the time features; using the importance to weight the multi-scale spectral features and the time features, and splicing the weighted multi-scale spectral features and the time features to form a comprehensive feature vector.

[0047] The target recognition network includes a fully connected layer and an output layer. The fully connected layer adopts a multi-layer fully connected network. The comprehensive feature vector is input into the fully connected layer for mapping and dimension conversion to obtain the comprehensive feature vector after dimensionality reduction. The output layer is the Softmax layer, which uses the comprehensive feature vector after dimensionality reduction for classification and recognition, and outputs the Chinese character recognition result.

[0048] Furthermore, the target recognition network can also adopt a generative adversarial network. By training the generator and the adversary, the accuracy of text recognition can be improved.

[0049] In step S202, the trained target detection model is used as a teacher model, and the sub-model obtained by compressing the target detection model is used as a student model. The sub-model is trained using the CP16 signal data set to obtain a text recognition model.

[0050] The present invention adopts the method of knowledge distillation to transfer the performance of a large model of a mature model to another relatively small sub-model. For the teacher-student model, the teacher model is a more complex network. In this embodiment, the trained target detection model is used as the teacher model. Its advantages are that it can automatically learn features from data and can capture the detailed features of Chinese characters. Its disadvantages are high complexity and high computing resources required for the training and reasoning processes. By simplifying and compressing the structure of the target detection model and training it as a student model, the model size can be effectively reduced while retaining the recognition accuracy, so that the text recognition model can adapt to the embedded application environment.

[0051] In some embodiments, in order to make the model converge quickly during the iteration process, the learning rate of the target detection model training process gradually decays, and its calculation formula is:

[0052] ;

[0053] In the formula, represents the current learning rate, and Respectively represent the maximum and minimum number of iterations allowed, represents the initial learning rate,k represents the attenuation factor, t Indicates the current training cycle number, T Indicates the expected total number of training epochs.

[0054] By compressing the target detection model, a sub-model of suitable volume and computational complexity can be obtained. In this embodiment, the target detection model is compressed in one or more of the following ways:

[0055] (1) Optimize the target detection model structure: reduce the number of network layers, reduce the size of convolution kernels, and reduce the number of channels;

[0056] (2) Parameter quantization: quantize the weights in the target detection model into low-bit-width fixed-point numbers or binary numbers;

[0057] (3) Model pruning: Determine the importance of the weights of each part in the target detection model and remove the weights whose importance is lower than the set threshold; or determine the importance of the channel or network layer in the target detection model and remove the weights whose importance is lower than the set value.

[0058] At step S203, the CP16 signal is acquired and input into the text recognition model to identify the Chinese character information contained therein. After the sub-model, i.e., the text recognition model, is trained, the text recognition model can be imported into the CP16 receiver so that the CP16 receiver can automatically recognize the Chinese character information after receiving the CP16 signal. In this way, when performing radio communication, the user does not need to carefully adjust various parameters of the receiver, does not need to monitor the text displayed on the waterfall chart for a long time, and gets rid of the interference of the outdoor environment.

[0059] At step S204, the recognized Chinese character information is converted into audio information. The recognized text information is converted into audio through text-to-speech (TTS) technology and broadcast through a loudspeaker. At the same time, the generated audio signal can also be encoded into other forms for secondary wireless transmission. This method can effectively overcome the limitations of ambient light and line of sight conditions, making information dissemination more efficient and reliable.

[0060] Through the above data processing method, the CP16 system not only maintains the original signal transmission mode, but also greatly improves the automation level of signal processing and information interpretation at the receiving end, so that the emergency wireless communication system can transmit Chinese character information more quickly and reliably in emergency situations such as sudden disasters, meeting the needs of emergency communications.

[0061] Figure 3 FIG. 3 is a flowchart schematically illustrating a method 300 for training a target detection model according to an embodiment of the present invention.

[0062] like Figure 3 As shown, at step S301, the CP16 signal data set is input into the sub-model and iterative training is performed. In some embodiments, after the dedicated model architecture is constructed in the above manner, the pre-trained weights can be loaded, and the corresponding loss function, optimizer and learning rate strategy can be set. In this embodiment, a gradually decaying learning rate and an improved loss function are set. The training data set is divided into a training set, a validation set and a test set and sent to the text recognition model for iterative training.

[0063] At step S302, an intermediate feature quantity is extracted from at least one target layer in the sub-model, and the difference between the intermediate feature quantity and the feature vector corresponding to the target detection model is calculated.

[0064] At step S303, the loss in the iterative training process is calculated according to the difference, and the iteration step size is adjusted according to the learning rate. The calculation formula of the loss function for calculating the loss is:

[0065] ;

[0066] ;

[0067] In the formula, represents the loss function, represents the distance between the probability distributions output by the teacher model and the student model, represents the loss function corresponding to the student model, represents the hyperparameter, represents the total number of training samples in the training data set, Indicates i The true value corresponding to the training, Indicates i The predicted value corresponding to the training, Indicates j The intermediate feature vector during the training iteration, Indicates j The feature vector corresponding to the target detection model during the training iteration, n Represents the number of iterations. The distance between the probability distributions output by the teacher model and the student model can be calculated using existing methods, such as calculating the KL divergence, which will not be repeated here.

[0068] In the process of iterative training using the above loss function, learning rate, etc., you can monitor the change in loss during training and the performance on the validation set. After all the training data have been trained once, save the weights of the model for subsequent use or continued training.

[0069] In an application scenario, in order to compress the model size, you can also use the method of quantizing weights: reduce the precision of weights and activation values, for example, from 32-bit floating point numbers to 8-bit integers.

[0070] At step S304, in response to the loss being less than a set value or the number of iterations reaching a set value, the training is stopped and the trained sub-model is output as a text recognition model.

[0071] Next, the solution of the present invention will be further described in conjunction with specific embodiments.

[0072] 1. Data preparation.

[0073] Collect training data: Collect a large amount of CP16 signal data containing a variety of Chinese characters and signal conditions. These signal data are preferably diverse to ensure the quality and diversity of the data set, including different fonts, sizes, backgrounds, noise, font integrity, etc. In this embodiment, a large amount of real CP16 Chinese character message data can be collected, covering various signal strengths and interference conditions. It is also possible to artificially generate various types of damaged data (such as signal loss, noise interference, etc.) based on the real data, and mark the correct Chinese character information of the damaged part.

[0074] In order to improve the training effect, data enhancement can also be performed, such as adding noise of different types and intensities to the training data to simulate the interference in the actual receiving environment, or randomly deleting some Chinese character information to simulate signal loss.

[0075] Data preprocessing: preprocess the received CP16 signal, such as denoising, smoothing and enhancement.

[0076] 2. Build a model.

[0077] Model selection: Choose the appropriate model architecture based on the task requirements. For example, for spectral feature recognition, you can consider using a convolutional model such as CNN, and for temporal features, you can consider using a sequence model such as LSTM.

[0078] Model architecture: Build a deep neural network model, which usually includes feature extraction layer, classification layer, etc. The classification layer includes feature fusion layer, recognition layer, etc.

[0079] Classification layer (or character recognition layer): Use a fully connected layer or a sequence model such as LSTM to predict the category of each Chinese character.

[0080] 3. Feature extraction.

[0081] (1) Spectral feature extraction: For spectral features, a multi-layer convolutional neural network is used as a feature extractor to extract the features of each character from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received. Local frequency feature extraction is the first step in character shape recognition and restoration. The purpose is to extract local features from time-spectrum (or spectrum-time) data. This step is usually achieved through sliding window technology and convolutional neural network (CNN):

[0082] 1) Sliding window technique: Use a sliding window to scan row by row or column by column along the time axis or frequency axis. The window size is usually consistent with the number of columns of the Chinese character dot matrix corresponding to the waterfall chart (for example, 16 columns).

[0083] 2) Convolutional Neural Network (CNN): At each window position, CNN is used to extract local frequency features. CNN can capture local patterns and features in spectrum-time data, such as the strokes and structure of Chinese characters. Convolution kernels of different sizes (such as 3x3, 5x5) are used to extract multi-scale features. Time series features are extracted to capture changes in the signal in the time dimension. Local frequency feature extraction provides basic features for subsequent multi-scale feature fusion.

[0084] (2) Multi-scale feature fusion is based on the results of local frequency feature extraction, which further improves the effect of Chinese character shape structure recognition and restoration. Its main idea is to combine feature information at different scales to capture richer details and structures corresponding to Chinese characters:

[0085] 1) Construct a multi-scale feature pyramid: Extract frequency features at different scales, such as from detail scale to global scale. Multi-scale feature extraction can be achieved by adjusting the convolution kernel size and stride, or using a multi-layer network structure.

[0086] 2) Multi-scale feature fusion: Fuse feature information of different scales to construct a higher resolution Chinese character image. Features of different scales can be combined together through methods such as feature splicing and feature weighting. Multi-scale feature fusion enables the model to capture the global structure and local details of the Chinese character image, thereby repairing the damaged part more effectively.

[0087] (3) Time series feature extraction: For time series features, a bidirectional long short-term memory network is used as a feature extractor to extract the features of the dependencies between each point in the generated time-frequency matrix of each character from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal. This step is usually implemented using two LSTM networks:

[0088] The bidirectional LSTM consists of two LSTM layers: one is the forward LSTM, which processes the input sequence from the first time step to the last time step; the other is the reverse LSTM, which processes the input sequence from the last time step to the first time step. This bidirectional structure enables the model to capture both past and future information. In the output layer of the bidirectional LSTM, the forward and reverse outputs are concatenated or weighted summed to form the final feature representation.

[0089] (4) Spectral-temporal feature fusion

[0090] Feature concatenation and weighting: Concatenate spectral features and temporal features, and use a weighting mechanism to form a comprehensive feature vector to improve the feature fusion effect. The weighted fusion here is achieved using the attention mechanism:

[0091] Here, each spectral and temporal feature vector contains a complete time series. Each feature vector containing a complete time step is taken as a whole for vector embedding. The attention mechanism is used to model the importance of the mutual influence between each feature for the embedded vector.

[0092] Specifically, weighted features = Attention(Q,K,V) = softmax((Q·K) / (E^(1 / 2)))V

[0093] Where Q is the query vector, K is the key vector, V is the value vector, and E is the dimension of the embedding vector.

[0094] (5) Pre-trained model: Pre-trained models, such as ResNet, VGG, etc., can be used as the basis for feature extraction.

[0095] 4. Model training.

[0096] Training: Use the labeled data set to train the model and optimize the loss function (such as cross entropy loss). In this embodiment, the model training is performed by adopting an improved mean square error loss function. Furthermore, during the training process, the differences between the 16-channel spectrum-time signal features corresponding to the damaged text and the 16-channel spectrum-time signal features corresponding to the complete text can be calculated, and these differences can be used as loss functions to guide the training of the model. The loss function constructed based on the difference between the features in this embodiment has been described in the above content and will not be repeated here.

[0097] Model tuning: Adjust model parameters and hyperparameters based on training results to achieve better performance.

[0098] 5. After the target detection model is trained using the above steps 1-4, the knowledge distillation method is used to use the above target detection model as the teacher model, and the sub-model obtained after the target detection model is compressed as the learning model, and the text recognition model is obtained through training through the constructed loss function, so as to obtain the text recognition model for the CP16 receiver. Since the above process has been described in detail, it will not be repeated here.

[0099] 6. Text reorganization.

[0100] Sequence prediction: Input the segmented text into the trained text recognition model one by one to predict the category of each text.

[0101] Text reconstruction: Reconstruct complete sentences or paragraphs based on the prediction results.

[0102] 7. Text conversion.

[0103] Text-to-speech and secondary transmission: The recognized text information is converted into audio through text-to-speech (TTS) technology and broadcast through a loudspeaker. At the same time, the generated audio signal can also be encoded into other forms for secondary wireless transmission. This method can effectively overcome the limitations of ambient light and line of sight conditions, making information transmission more efficient and reliable.

[0104] Through the above improvements, the CP16 system not only maintains the original signal transmission mode, but also greatly improves the automation level of signal processing and information interpretation at the receiving end. This enhanced emergency wireless communication system can transmit text information more quickly and reliably in emergency situations such as sudden disasters, meeting the needs of emergency communications.

[0105] Figure 4 It is a block diagram schematically showing the structure of the data processing system based on Chinese character message transmission according to this embodiment.

[0106] The present invention also provides a data processing system based on Chinese character message transmission. Figure 4 As shown, the system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a data processing method based on Chinese character message transmission as described above is implemented.

[0107] The system also includes other components familiar to those skilled in the art, such as a communication bus and a communication interface, whose configuration and functions are known in the art and thus will not be described in detail here.

[0108] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus or device. For example, a computer-readable storage medium may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0109] In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0110] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A data processing method based on Chinese character message transmission, characterized in that: include: A CP16 signal data set covering a variety of Chinese characters and signal conditions is obtained, and the data is input into a target detection model for training, wherein the target detection model includes a multi-layer convolutional bidirectional long short-term memory network and a target recognition network, the multi-layer convolutional network is used to extract multi-scale spectral features and time features, and fuse the multi-scale spectral features and time features to obtain a comprehensive feature vector, and the target recognition network performs feature mapping, dimension conversion and classification on the comprehensive feature vector to output a Chinese character recognition result; Based on the knowledge distillation method, the trained target detection model is used as the teacher model, the sub-model obtained by compressing the target detection model is used as the student model, and the sub-model is trained using the CP16 signal dataset to obtain a text recognition model; Acquire a CP16 signal, and input the CP16 signal into the text recognition model to recognize the Chinese character information contained therein; Convert the recognized Chinese character information into audio information; The multi-layer convolutional bidirectional long short-term memory network includes a feature extraction layer and a feature fusion layer, wherein the feature extraction layer includes a spectrum feature extraction layer and a time feature extraction layer, the spectrum feature extraction layer uses convolution kernels of different sizes to extract multi-scale spectrum features from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received; the time feature extraction layer uses a bidirectional long short-term memory network to extract forward time features and reverse time features from the time-frequency two-dimensional signal converted from the original 16-channel CP16 signal received; The feature fusion layer fuses multi-scale spectral features and temporal features by weighted concatenation to form a comprehensive feature vector.

2. The data processing method based on Chinese character message transmission according to claim 1 is characterized in that: The learning rate of the target detection model training process gradually decays, and its calculation formula is: ； In the formula, represents the current learning rate, and Respectively represent the maximum and minimum number of iterations allowed, represents the initial learning rate, k represents the attenuation factor, t Indicates the current training cycle number, T Indicates the expected total number of training epochs.

3. The data processing method based on Chinese character message transmission according to claim 1 is characterized in that: The feature fusion layer combines multi-scale spectral features and temporal features by weighted splicing, including: Use the attention mechanism to determine the importance of the mutual influence between multi-scale spectral features and temporal features; The multi-scale spectrum features and time features are weighted using the importance, and the weighted multi-scale spectrum features and time features are concatenated to form a comprehensive feature vector.

4. The data processing method based on Chinese character message transmission according to claim 1 is characterized in that: The target recognition network includes a fully connected layer and an output layer. The fully connected layer adopts a multi-layer fully connected network. The comprehensive feature vector is input into the fully connected layer for mapping and dimension conversion to obtain the comprehensive feature vector after dimensionality reduction. The output layer is the Softmax layer, which uses the comprehensive feature vector after dimensionality reduction for classification and recognition, and outputs the Chinese character recognition result.

5. The data processing method based on Chinese character message transmission according to claim 1 is characterized in that: The CP16 signal data set covering a variety of Chinese characters and signal conditions is obtained, including: One or more of frequency disturbance, time disturbance and noise is added to the CP16 signal corresponding to each Chinese character.

6. The data processing method based on Chinese character message transmission according to claim 1 is characterized in that: Methods for compressing the object detection model include one or more of the following methods: Optimize the target detection model structure: reduce the number of network layers, reduce the size of convolution kernels, and reduce the number of channels; Parameter quantization: quantize the weights in the target detection model into low-bit-width fixed-point numbers or binary numbers; Model pruning: Determine the importance of each part of the target detection model and remove the weights whose importance is lower than the set threshold; or determine the importance of the channel or network layer in the target detection model and remove the weights whose importance is lower than the set value.

7. The data processing method based on Chinese character message transmission according to claim 1 is characterized in that: The sub-model is trained using the CP16 signal data set, including: Inputting the CP16 signal data set into the sub-model and performing iterative training; Extracting intermediate feature quantities from at least one target layer in the sub-model, and calculating the difference between the intermediate feature quantities and the feature vector corresponding to the target detection model; The loss during the iterative training process is calculated based on the difference, and the iteration step size is adjusted based on the learning rate; In response to the loss being less than a set value or the number of iterations reaching a set value, the training is stopped and the trained sub-model is output as a text recognition model.

8. The data processing method based on Chinese character message transmission according to claim 7 is characterized in that: The calculation formula of the loss function for calculating the loss is: ； ； In the formula, represents the loss function, represents the distance between the probability distributions output by the teacher model and the student model, represents the loss function corresponding to the student model, represents the hyperparameter, represents the total number of training samples in the training data set, Indicates i The true value corresponding to the training, Indicates i The predicted value corresponding to the training, Indicates j The intermediate feature vector during the training iteration, Indicates j The feature vector corresponding to the target detection model during the training iteration, n Indicates the number of iterations.

9. A data processing system based on Chinese character message transmission, characterized in that: include: processor; A memory storing computer program instructions, which, when executed by the processor, implements the data processing method based on Chinese character message transmission according to any one of claims 1 to 8.

Images