Wireless communication signal specification recognition method and system based on deep learning
Through the deep learning feature fusion network method, the problem of indistinguishable signals in the wireless communication signal specification recognition is solved, and accurate recognition is achieved in a low signal-to-noise ratio environment, providing a reliable identification solution.
Patent Information
- Application Number
- CN202111543403.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-16
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-12-16
AI Technical Summary
The existing wireless communication signal specification recognition technology is difficult to distinguish different signals of the same modulation method, and the recognition effect is poor in low signal-to-noise ratio environments, especially the burst signal cannot obtain complete frame head structure data, which leads to difficulty in identification.
Using a feature fusion network method based on deep learning, the feature fusion network is connected by building densely connected convolutional units, using feature multiplexing mechanism to perform feature fusion, and deep learning and mapping of signal data flow and vector diagrams are achieved to achieve signal specifications Accurate identification.
Effectively identifying signals of different specifications of the same modulation method improves the identification accuracy and noise immunity in a low signal-to-noise ratio environment, and provides a reliable wireless communication signal specification recognition solution with good application prospects.
Smart Images

Figure CN114358055B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wireless communication signal recognition, and in particular relates to a method and system for wireless communication signal specification recognition based on deep learning. Background Art
[0002] With the rapid development of wireless communication technology in recent years, wireless communication signal specifications and communication equipment have shown the characteristics of high similarity between some signal classes, rich signal types and diversified forms, which brings challenges to traditional signal blind processing and analysis technology. Among them, the recognition technology of wireless communication signal specifications is the basis for a series of subsequent work such as monitoring, positioning, demodulation analysis, etc., and has received extensive attention and research in the industry. At present, domestic and foreign research institutions and institutes have carried out a lot of work. Some methods establish a database of disappearance and duration of wireless signal waveforms, and distinguish and identify them by combining characteristic data such as signal amplitude, frequency and time. However, the clustering algorithm used in such methods cannot effectively distinguish signals with similar duration and disappearance frame lengths, and the recognition effect is not good in low signal-to-noise ratio environments. There are also some methods that use the signal bit stream frame synchronization code as a distinguishing feature to identify the signal specifications transmitted by unauthorized users and authorized users, but such methods require receiving a complete frame synchronization code sequence, which is often difficult to meet in actual environments, and the false detection rate of recognition is high when the signal-to-noise ratio is low.
[0003] In addition, as deep learning technology is introduced into the field of signal processing, some methods use the differences in the visual characteristics of the frame structure time-frequency spectrum of different wireless communication signals as features to identify different signals. Such methods are only effective for signal sets that use different modulation methods or have unique frame header structures and obvious differences in the time-frequency spectrum. In practice, many types of wireless communication signals use the same modulation method and have no obvious difference characteristics in the time-frequency spectrum, or even no difference characteristics at all. At the same time, due to the reception reasons in the actual environment, it is often impossible to obtain complete frame header unique structure data for burst signals, which brings many difficulties to the actual signal specification recognition. Summary of the invention
[0004] To this end, the present invention provides a method and system for identifying wireless communication signal specifications based on deep learning. For the identification of signals with the same modulation mode but different specifications in wireless communication signal analysis and processing, a feature fusion neural network (FFN) is constructed to fuse, learn and map the feature information of the wireless communication signal data stream and the vector diagram, so as to achieve accurate identification of the signal specifications.
[0005] According to the design scheme provided by the present invention, a wireless communication signal specification identification method based on deep learning includes the following contents:
[0006] Construct a feature fusion network model and use sample data for training and learning. In the feature fusion network model, all feature extraction layers are connected through densely connected convolutional units, and feature reuse mechanism is used to fuse the feature data of the current feature extraction layer with the feature data extracted by all previous feature extraction layers.
[0007] The sampled signal data stream is preprocessed to obtain a data stream matrix containing complete difference information of the signal data and a signal vector diagram containing modulation information of the reaction signal;
[0008] The data flow matrix and signal vector diagram obtained after preprocessing are used as the input of the feature fusion network model after training and learning, and the feature fusion network model after training and learning is used to classify and identify the signal specifications of the signal data flow.
[0009] As a wireless communication signal specification identification method based on deep learning of the present invention, further, the feature fusion network model includes: branch one for extracting features from an input data stream matrix, branch two for extracting features from an input signal vector diagram, a fully connected layer for fusing features extracted from branch one and branch two, and a classifier for performing feature classification on fused feature data output by the fully connected layer, wherein branch one and branch two are respectively connected to all feature extraction layers through densely connected convolutional units to extract features from corresponding branch inputs.
[0010] As a wireless communication signal specification identification method based on deep learning of the present invention, further, according to the specified specification standard and random message symbol data sequence, the actual wireless communication channel is simulated through a Gaussian channel to generate a wireless communication signal, and each wireless communication signal is preprocessed to obtain the corresponding data flow matrix and signal vector diagram.
[0011] As a wireless communication signal specification identification method based on deep learning of the present invention, further, by setting a random frequency deviation strategy, a frequency deviation control factor is used to perform a random frequency deviation operation on the simulated wireless communication signal; and each generated wireless communication signal is randomly intercepted to obtain sample data for preprocessing.
[0012] As a wireless communication signal specification identification method based on deep learning of the present invention, further, the frequency deviation factor is randomly selected within the range of 0 to a preset frequency deviation upper limit; the position interval of the signal random length interception is a complete single frame of data or a data fragment with a random length value in a preset length interval in a single frame of data, and the interception starting point of the data fragment is randomly selected from the frame header or the frame tail.
[0013] As a wireless communication signal specification identification method based on deep learning of the present invention, further, the data of each sample point of the signal is arranged according to the rule of its real part-odd column, and its imaginary part-even column, and the complex data stream and single-channel real data stream matrix representation of the signal is obtained by transformation; and the signal vector diagram of the corresponding point is obtained by converting the signal data into a grayscale image.
[0014] As a wireless communication signal specification identification method based on deep learning of the present invention, further, before the signal data is input into the feature fusion network model, the signal data is first normalized, and the numerical intervals of various signal data are moved to between [0,1] so that the characteristic values of each signal specification have the same measurement scale; and the normalized signal data is constructed according to preset organizational rules to form a data matrix used as the model input data structure.
[0015] Furthermore, the present invention also provides a wireless communication signal specification recognition system based on deep learning, comprising: a model training module, a signal acquisition module and a signal recognition module, wherein:
[0016] The model training module is used to build a feature fusion network model and use sample data for training and learning. In the feature fusion network model, all feature extraction layers are connected through densely connected convolutional units, and the feature multiplexing mechanism is used to fuse the current feature extraction layer data with the feature data extracted by all previous feature extraction layers;
[0017] A signal acquisition module, used for preprocessing the sampled signal data stream to obtain a data stream matrix containing complete difference information of the signal data and a signal vector diagram containing modulation information of the reaction signal;
[0018] The signal recognition module is used to use the data flow matrix and signal vector diagram obtained after preprocessing as the input of the trained feature fusion network model, and use the trained feature fusion network model to classify and identify the signal specifications of the signal data flow.
[0019] Beneficial effects of the present invention:
[0020] The present invention aims at the problems of distinguishing different specification signals of the same modulation mode and relying only on single-dimensional features in the current wireless communication signal specification identification, and the like. The present invention obtains a standardized wireless communication signal data matrix and a vector diagram as a feature extraction source by signal preprocessing, and constructs a specification identification model based on a feature fusion network; then, the feature extraction module of the network model is used to perform deep feature extraction on the input signal data matrix and vector diagram by using a densely connected convolutional structure, and two-way feature fusion is realized, and finally the target signal specification identification is completed through learning and mapping; in addition, a random frequency deviation strategy can be added when generating a training signal sample set to further improve the model robustness performance, thereby providing a reliable and practical solution for the specification identification of wireless communication signals, and having good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 The following is a flowchart of a method for identifying wireless communication signal specifications based on deep learning in an embodiment;
[0022] Figure 2 This is a schematic diagram of the sample data preprocessing process in the embodiment;
[0023] Figure 3 A schematic diagram of the process of adding random frequency deviation to sample data in the embodiment;
[0024] Figure 4 This is an example of the feature fusion network model structure in the embodiment;
[0025] Figure 5 This is a schematic diagram of the internal structure of feature extraction in the model example in the embodiment;
[0026] Figure 6 Schematic diagram of the structure of the nonlinear function H(·) in the embodiment;
[0027] Figure 7 The following is a schematic diagram of the signal specification identification process in combination with the model example in the embodiment. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.
[0029] In view of the problems of distinguishing different specification signals of the same modulation mode in the current wireless communication signal specification identification, and relying only on single-dimensional features for identification, the embodiment of the present invention provides a wireless communication signal specification identification method based on deep learning, see Figure 1 As shown, it contains the following contents:
[0030] S101, constructing a feature fusion network model and using sample data for training and learning, wherein in the feature fusion network model, all feature extraction layers are connected through densely connected convolutional units, and feature multiplexing mechanism is used to perform feature fusion on the current feature extraction layer data and feature data extracted by all previous feature extraction layers;
[0031] S102, preprocessing the sampled signal data stream to obtain a data stream matrix containing complete difference information of the signal data and a signal vector graph containing modulation information of the reaction signal;
[0032] S103, using the data flow matrix and signal vector diagram obtained after preprocessing as inputs of the trained feature fusion network model, and using the trained feature fusion network model to classify and identify signal specifications of the signal data flow.
[0033] Through data preprocessing, the sampled complex signal data stream is converted into a real data matrix and a signal vector map. The signal data stream is selected as the main feature extraction source, and the vector map is used as a supplement to improve the signal recognition performance. Among them, the signal vector map can fully reflect the modulation information of the signal, and the data stream matrix can retain the complete difference information of the signal data; and the densely connected convolution structure (DCC) is used as the feature extraction module to construct a feature fusion network model. The densely connected convolution structure is used to fuse, learn and map the feature information of the wireless communication signal data stream and the vector map. By fusing the features extracted from the signal data stream and the vector map, accurate recognition of the signal specifications is achieved, which solves the problems that the existing methods are difficult to distinguish signals of different specifications with the same modulation method, and rely only on single-dimensional features.
[0034] As a wireless communication signal specification identification method based on deep learning in an embodiment of the present invention, further, according to the specified specification standard and random message symbol data sequence, the actual wireless communication channel is simulated through a Gaussian channel to generate a wireless communication signal, and each wireless communication signal is preprocessed to obtain the corresponding data flow matrix and signal vector diagram. Further, by setting a random frequency deviation strategy, the frequency deviation control factor is used to perform a random frequency deviation operation on the simulated wireless communication signal; and each generated wireless communication signal is randomly intercepted to obtain sample data for preprocessing. Among them, the frequency deviation factor is randomly selected within the range of 0 to the preset frequency deviation upper limit; the position interval of the random length interception of the signal is a complete single frame data or a data fragment of a random length value in a preset length interval in a single frame data, and the interception starting point of the data fragment is randomly selected from the frame head or the frame tail. Further, in the process of preprocessing the signal, the data of each sample point of the signal is arranged according to the rule of its real part-odd column and imaginary part-even column, and the complex data stream single-channel real data stream matrix representation of the signal is obtained by transformation; and the signal vector diagram of the corresponding point is obtained by converting the signal data into a grayscale image. Furthermore, before the signal data is input into the feature fusion network model, the signal data is first normalized, and the numerical intervals of various signal data are moved to between [0,1], so that the characteristic values of each signal specification have the same measurement scale; and the normalized signal data is constructed into a data matrix used as the model input data structure according to the preset organization rules. Furthermore, the feature fusion network model includes: a branch 1 for extracting features from the input data flow matrix, a branch 2 for extracting features from the input signal vector map, a fully connected layer for fusing the features extracted from branch 1 and branch 2, and a classifier for feature classification of the fused feature data output by the fully connected layer, wherein branch 1 and branch 2 are respectively connected to all feature extraction layers through densely connected convolutional units to extract features from the corresponding branch inputs.
[0035] See also Figure 2 As shown, according to the specified specifications and random message symbol data sequence, the corresponding wireless communication signal is generated, and the relevant parameter indicators of signal generation are implemented according to the specific standard specifications of each signal type. A frequency deviation relative to the symbol rate is added to the generated standard signal, such as Figure 3As shown in the figure, the specific frequency deviation value added to each signal sample is randomly selected in the range of 0 to λ, where the upper limit of the frequency deviation λ is generally set to 0.04 or 0.06. Subsequently, the data of each signal sample is randomly intercepted, which is divided into two interception situations: the intercepted signal includes a complete single frame or a part of a single frame. If a part of a single frame is intercepted, the intercepted signal length is set in the range of 30% to 100%, and the interception starting point is randomly selected at the head or tail of a frame. After random interception, the generated signal samples are closer to the actual situation, and the practical performance of the method is improved. Then, a Gaussian channel is used to simulate the actual wireless communication channel environment, and then the signal sample data is converted into a vector diagram based on a grayscale image. Combined with the processing requirements of the network model, the vector diagram is set to a size of 128×128. In view of the internal structure of the neural network, the complex data of the signal is converted into a single-channel form. The specific method is to separate the real and imaginary parts of the sampling values of each sample point, and arrange them in the order of imaginary part-even column and real part-odd column to obtain a single-channel real number expression of the complex signal. In order to prevent signals of different intensities from affecting the network processing due to the statistical size level of their amplitudes and to make the constructed network model more effective in feature extraction and learning, the obtained signal data is normalized:
[0036]
[0037] Among them, data1 is the signal data without normalization, and data2 is the result after normalization. Through normalization, the numerical range of each signal sample is moved to the range of [0,1], so that they have the same dimension scale. Finally, the signal data samples are converted into a two-dimensional efficient structure form by arranging every 336 values into a column, and a data matrix of size 336×336 is constructed.
[0038] See also Figure 4 and 5 As shown in the figure, the feature extraction module of the densely connected convolutional unit (Dense Block) adopts the feature reuse mechanism, that is, the feature extraction operation in each layer will reuse the feature processing and extraction results of all previous layers. Specifically, the data of this layer and the feature data extracted from the previous layers are connected in the channel dimension, and then nonlinear operation processing is performed: l =H l ([x0,x1,...,x l-1 ]),l≥1. Therefore, for an L-layer Dense Block module structure, it contains The Dense Block module uses a number of connections, which can directly fuse the features of different layers of the network, enhance the effect of feature reuse, and realize the direct fusion of features of different network layers. When obtaining the same network performance, the number of network parameters used is less than half of the currently widely used residual network (ResNet), which greatly improves the computational efficiency. At the same time, this network structure has stronger anti-overfitting and robust performance. When achieving the same classification accuracy, the number of parameters used by the Dense Block module is less than half of the deep residual network, which greatly improves the computational efficiency. At the same time, its anti-overfitting and generalization performance are also stronger. Considering the internal structure of the network model and the relevant parameters of the actual signal, the input signal data samples can be uniformly set to a 336×336 data matrix form, and the vector map samples can be uniformly set to a size of 128×128. The feature extraction of the signal is realized through 7 feature extraction modules (Dense Block). Each feature extraction module is a densely connected convolution structure, and the number of densely connected convolution nodes contained is an integer multiple of N. According to the complexity of the identified signal, N is 4, 6 or 8 as appropriate. The nonlinear processing function H(·) taken for the input signal data is as follows: Figure 6 As shown in the figure, the input signal data is first processed by batch normalization, then activated by a rectified linear unit (ReLU) and a 1×1 convolution, and then batch normalization, ReLU activation and a 3×3 convolution operation are performed to complete the extraction of signal feature information. Figure 7 As shown in the figure, the model processes and recognizes the input signal samples as follows:
[0039] 1) Perform 7×7 convolution and 3×3 maximum pooling operations on the signal data matrix and vector map respectively to preliminarily extract features and reduce the size of feature data;
[0040] 2) The feature data obtained from the data matrix is passed through 4 Dense Block modules in sequence, and the feature data obtained from the vector map is passed through 3 Dense Block modules in sequence to gradually extract features at different levels. The structure of each DenseBlock module is as follows: Figure 4 As shown in the figure, the Dense Block modules of the two networks contain different numbers of densely connected convolution nodes, which are 6, 12, 32, 32 and 6, 12, 12. In each Dense Block module, the feature data size of each network layer remains consistent, so that the feature data can be connected in the channel dimension;
[0041] 3) Each adjacent Dense Block module is connected through a Transition module. In the Transition module, the signal feature data first undergoes a 1×1 convolution, and then passes through a 2×2 average pooling layer with a step size of 2 to reduce the size of the signal feature data;
[0042] 4) Perform a 7×7 global average pooling operation on the two features obtained in the end to obtain their feature vectors, and then concatenate the two feature vectors to achieve feature fusion;
[0043] 5) The fusion features obtained in the above steps are sent to the fully connected classifier for operation to obtain the result of signal specification recognition.
[0044] Further based on the above process, it is verified through experimental data: In the implementation scheme of this case, the data stream is used as the main feature extraction source, and the vector diagram is used as the supplementary feature. The two-dimensional features of the signal are fused for learning and processing, which can effectively identify the 6 types of signal specifications, 4 of which are 8PSK modulation, which has certain advantages; and while ensuring the accuracy of signal specification recognition, it has good noise resistance and certain frequency deviation resistance. The feature fusion network model has the advantages of small file size and fast computing speed, can be easily deployed and carried out in real time, and has good engineering application prospects.
[0045] Furthermore, based on the above method, an embodiment of the present invention also provides a wireless communication signal specification recognition system based on deep learning, comprising: a model training module, a signal acquisition module and a signal recognition module, wherein:
[0046] The model training module is used to build a feature fusion network model and use sample data for training and learning. In the feature fusion network model, all feature extraction layers are connected through densely connected convolutional units, and the feature multiplexing mechanism is used to fuse the current feature extraction layer data with the feature data extracted by all previous feature extraction layers;
[0047] A signal acquisition module, used for preprocessing the sampled signal data stream to obtain a data stream matrix containing complete difference information of the signal data and a signal vector diagram containing modulation information of the reaction signal;
[0048] The signal recognition module is used to use the data flow matrix and signal vector diagram obtained after preprocessing as the input of the trained feature fusion network model, and use the trained feature fusion network model to classify and identify the signal specifications of the signal data flow.
[0049] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0050] Based on the above method and / or system, an embodiment of the present invention further provides a server, comprising: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.
[0051] Based on the above method and / or system, an embodiment of the present invention further provides a computer-readable medium having a computer program stored thereon, wherein the above method is implemented when the program is executed by a processor.
[0052] In all examples shown and described herein, any specific values should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.
[0053] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0054] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A method for identifying wireless communication signal specifications based on deep learning, characterized in that: Contains the following: Construct a feature fusion network model and use sample data for training and learning. In the feature fusion network model, all feature extraction layers are connected through densely connected convolutional units, and feature reuse mechanism is used to fuse the feature data of the current feature extraction layer with the feature data extracted by all previous feature extraction layers. According to the specified specification standard and random message symbol data sequence, the wireless communication signal is generated by simulating the actual wireless communication channel through a Gaussian channel, and by setting a random frequency deviation strategy, the frequency deviation control factor is used to perform a random frequency deviation operation on the simulated wireless communication signal; and each generated wireless communication signal is randomly intercepted to obtain sample data for preprocessing, and each wireless communication signal is preprocessed to obtain the corresponding data flow matrix containing complete difference information of the signal data and the signal vector diagram containing the response signal modulation information; wherein the preprocessing includes: arranging the data of each sample point of the signal according to the rule of the real part-odd column and the imaginary part-even column, obtaining the complex data flow single-channel real data flow matrix representation of the signal by conversion, and obtaining the signal vector diagram of the corresponding point by converting the signal data into a grayscale image; before the signal data is input into the feature fusion network model, the signal data is first normalized, and the numerical range of each type of signal data is moved to between [0,1], so that the eigenvalues of each signal specification have the same measurement scale, and the normalized signal data is constructed according to the preset organization rule to serve as the data structure of the model input data. The data flow matrix and signal vector diagram obtained after preprocessing are used as the input of the trained feature fusion network model, and the trained feature fusion network model is used to classify and identify the signal specifications of the signal data flow.
2. The method for identifying wireless communication signal specifications based on deep learning according to claim 1, characterized in that: The feature fusion network model includes: branch one for extracting features from an input data stream matrix, branch two for extracting features from an input signal vector diagram, a fully connected layer for fusing features extracted from branch one and branch two, and a classifier for performing feature classification on fused feature data output by the fully connected layer, wherein branch one and branch two are respectively connected to all feature extraction layers through densely connected convolutional units to extract features from corresponding branch inputs.
3. The method for identifying wireless communication signal specifications based on deep learning according to claim 1, characterized in that: The frequency deviation factor is randomly selected within the range of 0 to the preset frequency deviation upper limit; the position interval of the signal random length interception is a complete single frame of data or a data segment with a random length value in the preset length interval in the single frame of data, and the interception starting point of the data segment is randomly selected from the frame head or frame tail.
4. A wireless communication signal specification recognition system based on deep learning, characterized in that: It includes: model training module, signal acquisition module and signal recognition module, among which, The model training module is used to build a feature fusion network model and use sample data for training and learning. In the feature fusion network model, all feature extraction layers are connected through densely connected convolutional units, and the feature multiplexing mechanism is used to fuse the current feature extraction layer data with the feature data extracted by all previous feature extraction layers; The signal acquisition module is used to generate wireless communication signals according to the specified specification standards and random message symbol data sequences, and simulate the actual wireless communication channel through a Gaussian channel, and perform random frequency deviation operations on the simulated wireless communication signals by setting a random frequency deviation strategy and using a frequency deviation control factor; and randomly intercept each generated wireless communication signal to obtain sample data for preprocessing, and preprocess each wireless communication signal to obtain the corresponding data flow matrix containing complete difference information of the signal data and a signal vector diagram containing the modulation information of the response signal; wherein the preprocessing includes: arranging the data of each sample point of the signal according to the rule of the real part-odd column and the imaginary part-even column, obtaining the complex data flow single-channel real data flow matrix representation of the signal by conversion, and obtaining the signal vector diagram of the corresponding point by converting the signal data into a grayscale image; before the signal data is input into the feature fusion network model, the signal data is first normalized, and the numerical range of each type of signal data is moved to between [0,1], so that the eigenvalues of each signal specification have the same measurement scale, and the normalized signal data is constructed according to the preset organization rules to serve as the data structure of the model input; The signal recognition module is used to use the data flow matrix and signal vector diagram obtained after preprocessing as the input of the trained feature fusion network model, and use the trained feature fusion network model to classify and identify the signal specifications of the signal data flow.
5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 3 is implemented.
6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 3 is implemented.
Citation Information
Patent Citations
Radar radiation source identification method and device based on extended residual network
CN110147812A
Load identification method based on feature fusion and deep learning and application thereof
CN110443302A
Multi-modal feature fusion modulation recognition method and system based on neural network
CN110490095A
Feature fusion method based on deep learning
CN111382803A
Depth information completion model training method and device, equipment and storage medium
CN112560875A
