Compressed convolutional neural network and radar radiation source signal identification method

Abstract

The invention belongs to the technical field of image compression and signal recognition, and particularly relates to a compressed convolutional neural network and a radar radiation source signal recognition method. The last residual block in the compressed convolutional neural network is a compressed residual block, and output feature maps of two convolutional layers in the residual block have original resolution; and for the last residual block, the input feature map input to the penultimate convolutional layer is not filled. And setting the convolution step length of the penultimate convolution layer. And performing a convolution operation on the input feature map by using an original convolution kernel of the penultimate convolution layer to obtain an output feature map, the output feature map having a compression resolution less than the original resolution. The radar radiation source signal identification method comprises the following steps: receiving a radar radiation source signal, and preprocessing the radar radiation source signal to obtain a preprocessed image; and taking the preprocessed image as input, and performing identification by adopting a compressed convolutional neural network to obtain the category of the input image.

Description

Technical Field

[0001] The present invention relates to the technical field of image compression, and in particular to a compressed convolutional neural network and a radar radiation source signal recognition method. Background Art

[0002] The intra-pulse modulation classification of radar signals plays an important role in the classification and identification of radar emitters. Accurate intra-pulse modulation classification of received radar signals can estimate the function and potential threat level of enemy radars, and provide intelligent support for the electronic support measures (ESM) system to formulate countermeasures. Therefore, the reliability of ESM is highly dependent on the accuracy and stability of intra-pulse modulation classification. Recently, due to the more complex and uncertain electromagnetic environment, the difficulty of identifying radar modulated signals has increased significantly. Various forms of radar pulse modulation signals have put forward more stringent requirements for accuracy and robustness classification.

[0003] Traditional machine learning (ML) methods used to be relatively effective methods. Usually, they are divided into two steps: the first step is to extract features based on manually designed algorithms, and the second step is to design classifiers using traditional ML algorithms. For example, some scholars have proposed classifiers based on support vector machines (SVMs) and extract instantaneous frequency features, classifiers based on artificial neural networks (ANNs) and extract signal entropy features, and classifiers based on kernel fuzzy mean and extract main ridge features. However, the main limitations of traditional ML methods are threefold: (1) The design of feature extractors requires careful engineering design and considerable domain expertise, resulting in a high application threshold; (2) The generalization ability of the method is poor. When the application environment changes, the performance of the original feature extractor will be greatly reduced; (3) The method is very sensitive to the signal-to-noise ratio. When the signal-to-noise ratio (SNR) of the received signal is low, the classification accuracy cannot be maintained. Obviously, the traditional ML methods for classification and recognition of intra-pulse modulation signals cannot meet the growing requirements of ESM for accuracy, intelligence, and stability.

[0004] In view of this, if good features can be automatically learned using a general learning procedure, the problems existing in the above ML methods can be avoided, which is also the key advantage of deep learning algorithms. In recent years, deep learning algorithms have made breakthrough progress in image classification and recognition. Deep learning (DL)-based algorithms have been increasingly adopted to solve the bottleneck problem in the field of radar signal processing. The common processing procedure is to first convert the original one-dimensional data into a two-dimensional time-frequency image (TFI) through time-frequency analysis, and then apply TFI to drive the automatic training of the CNN model.

[0005] However, these CNN-based methods still have similar defects, and the classification accuracy mainly depends on the architectural design of the CNN model. However, the existing CNN models were originally designed for computer vision tasks, not for TFI recognition. Most studies mainly consider data preprocessing or algorithm fusion, and ignore the design of an efficient CNN model architecture for TFI recognition. The obvious difference between ordinary images and TFI is that TFI contains a large amount of noise signals. Therefore, the presence of a large amount of noise signals in the output feature map is the main reason for the rapid decline in CNN performance. A common phenomenon also exists in a large number of literature studies. Under low signal-to-noise ratio conditions, the classification accuracy of TFI drops rapidly, especially for easily confused signals.

[0006] Based on this, it is urgent to propose a compressed convolutional neural network suitable for TFI classification under low signal-to-noise ratio conditions and a radar emitter signal recognition method using the compressed convolutional neural network. Summary of the invention

[0007] The object of the present invention is to provide a compressed convolutional neural network and a radar emitter signal recognition method suitable for the classification of TFI under low signal-to-noise ratio conditions.

[0008] In a first aspect, the present invention provides a compressed convolutional neural network, which is a residual network. The residual network includes a plurality of residual blocks, and each residual block includes two sequentially connected convolutional layers.

[0009] The compressed convolutional network is generated by the following steps: compressing the last residual block. Before compression, the output feature maps of the two convolutional layers of the last residual block have the original resolution;

[0010] For the last residual block, the input feature map to the penultimate convolutional layer is not padded;

[0011] Set the convolution step size of the penultimate convolutional layer;

[0012] A convolution operation is performed on the input feature map using the original convolution kernel of the penultimate convolution layer to obtain an output feature map having a compressed resolution smaller than the original resolution.

[0013] As a possible implementation method, the original resolution is 7*7, the convolution step size is 1, the original convolution kernel is 3*3, and the compressed resolution is 5*5; or, the original resolution is 7*7, the convolution step size is 2, the original convolution kernel is 3*3, and the compressed resolution is 3*3.

[0014] In a second aspect, the present invention provides another compressed convolutional neural network, the convolutional neural network is ResNet50, ResNet50 includes multiple bottleneck structures, each bottleneck structure includes three sequentially connected convolutional layers;

[0015] A compressed convolutional neural network is generated by the following steps: compressing the last bottleneck structure. Before compression, the output feature maps of the three convolutional layers of the last bottleneck structure have the original resolution;

[0016] For the last bottleneck structure, the input feature map to the penultimate convolutional layer is not padded;

[0017] Set the convolution step size of the penultimate convolutional layer;

[0018] A convolution operation is performed on the input feature map using the original convolution kernel of the penultimate convolution layer to obtain an output feature map having a compressed resolution smaller than the original resolution.

[0019] As a possible implementation method, the original convolution kernels of the three convolution layers from top to bottom are 1*1, 3*3 and 1*1 respectively, the original resolution is 7*7, the convolution step is 1, and the compressed resolution is 5*5; or, the original resolution is 7*7, the convolution step is 2, and the compressed resolution is 3*3.

[0020] In a third aspect, the present invention provides a third compressed convolutional neural network, the compressed convolutional neural network is GoogleNet, and GoogleNet is composed of inception as a basic unit; each inception includes three parallel convolutional layers and one maximum pooling layer; the original convolution kernels of the three convolutional layers are 5*5, 3*3 and 1*1 respectively, and the size of the maximum pooling layer is 3*3; before compression, the output feature maps of the three convolutional layers and one maximum pooling layer have the same original resolution; the compressed convolutional neural network is generated by the following steps:

[0021] Fill the input feature map of the convolution layer with the original convolution kernel of 1*1, set the first convolution step size, perform the convolution operation, and obtain the first output feature map;

[0022] Fill the input feature map of the convolution layer with the original convolution kernel of 5*5, set the second convolution step size, perform the convolution operation, and obtain the second output feature map;

[0023] The input feature map of the convolution layer whose original convolution kernel is 3*3 is not filled, the third convolution step size is set, the convolution operation is performed, and the third output feature map is obtained;

[0024] The input feature map of the maximum pooling layer with a size of 3*3 is not filled, the first pooling step size is set, and the pooling operation is performed to obtain the fourth output feature map;

[0025] The first output feature map, the second output feature map, the third output feature map, and the fourth output feature map are concatenated to obtain an output feature map, where the output feature map has a first compressed resolution that is less than the original resolution.

[0026] As a possible implementation, compression methods include:

[0027] The input feature map of the convolution layer whose original convolution kernel is 1*1 is not filled, the fourth convolution step size is set, the convolution operation is performed, and the fifth output feature map is obtained;

[0028] The input feature map of the convolution layer with the original convolution kernel of 5*5 is not filled, the fifth convolution step size is set, the convolution operation is performed, and the sixth output feature map is obtained;

[0029] The input feature map of the convolution layer with the original convolution kernel of 3*3 is not filled, the sixth convolution step is set, the convolution operation is performed, and the seventh output feature map is obtained;

[0030] The input feature map of the maximum pooling layer with a size of 3*3 is not filled, the second pooling step size is set, and the pooling operation is performed to obtain the eighth output feature map;

[0031] The fifth output feature map, the sixth output feature map, the seventh output feature map, and the eighth output feature map are concatenated to obtain an output feature map, wherein the output feature map has a second compressed resolution that is less than the original resolution.

[0032] As a possible implementation method, the first convolution step size is 2, the second convolution step size is 1, the third convolution step size is 1, and the first pooling step size is 1. At this time, the first compression resolution is 5*5.

[0033] As a possible implementation method, the fourth convolution step size is 3, the fifth convolution step size is 1, the sixth convolution step size is 2, and the second pooling step size is 2. At this time, the first compression resolution is 3*3.

[0034] As a possible implementation method, filling the input feature map of the convolution layer whose original convolution kernel is 1*1 includes: filling the input feature map with a value of 0 for one pixel; filling the input feature map of the convolution layer whose original convolution kernel is 5*5 includes: filling the input feature map with a value of 0 for one pixel.

[0035] In a fourth aspect, the present invention provides a fourth compressed convolutional neural network, the compressed convolutional neural network is MobileNetV2, MobileNetV2 is composed of a linear bottleneck structure with an inverted residual as a basic unit; a linear bottleneck structure includes three convolutional layers connected in sequence;

[0036] The compressed convolutional neural network is generated by the following steps: compressing the last linear bottleneck structure. Before compression, the output feature maps of the three convolutional layers of the last bottleneck structure have the original resolution;

[0037] For the last linear bottleneck structure, the input feature map to the penultimate convolutional layer is not padded;

[0038] Set the convolution step size of the penultimate convolutional layer;

[0039] A convolution operation is performed on the input feature map using the original convolution kernel of the penultimate convolution layer to obtain an output feature map having a compressed resolution smaller than the original resolution.

[0040] As a possible implementation method, the original convolution kernels of the three convolution layers from top to bottom are 1*1, 3*3 and 1*1 respectively, the original resolution is 7*7, the convolution step is 1, and the compressed resolution is 5*5; or, the original resolution is 7*7, the convolution step is 2, and the compressed resolution is 3*3.

[0041] In a fifth aspect, the present invention further provides a radar emitter signal recognition method, comprising the following steps:

[0042] Receiving radar radiation source signals and preprocessing the radar radiation source signals to obtain preprocessed images;

[0043] The preprocessed image is used as an input feature map, and the compressed convolutional neural network provided by the first or fourth aspect is used for recognition to obtain the category of the input image.

[0044] Compared with the prior art, the compressed convolutional neural network provided by the present invention does not change the parameter complexity and floating-point precision calculation complexity of the model. It is a simple and flexible method. The compressed convolutional neural network can effectively suppress the noise signal in the output feature map, highlight the feature signal that is crucial to classification and recognition, and solves the defects of difficulty in identifying radar radiation sources under low signal-to-noise ratio conditions, high complexity of the recognition method and low recognition performance, and has a good recognition effect on radar radiation source signals with greater similarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0046] Figure 1 A flowchart of a compressed residual convolutional neural network provided by an embodiment of the present invention;

[0047] Figure 2The basic component unit of the residual network model provided by the embodiment of the present invention;

[0048] Figure 3 The basic component unit of the ResNet50 network model provided by the embodiment of the present invention;

[0049] Figure 4 The basic component unit of the GoogleNet network model provided by the embodiment of the present invention;

[0050] Figure 5 The basic component unit of the MobileNetV2 network model provided by the embodiment of the present invention;

[0051] Figure 6 A comparison chart of the average recognition rates of 14 types of signals by the compressed residual network model provided by an embodiment of the present invention and the original residual network model;

[0052] Figure 7 A comparison chart of the changes in recognition rates of 14 types of signals under different signal-to-noise ratio conditions between the model designed using the embodiment of the present invention and the original model;

[0053] Figure 8 This is a comparison chart of the recognition accuracy of the compressed convolutional neural network model designed based on the embodiment of the present invention and three classic original convolutional neural network models. DETAILED DESCRIPTION

[0054] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0055] It should be noted that when an element is referred to as being "fixed to" or "disposed on" another element, it can be directly on the other element or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0056] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. The meaning of "several" is one or more, unless otherwise clearly and specifically defined.

[0057] In the description of the present invention, it is necessary to understand that the directions or positional relationships indicated by the terms "up", "down", "front", "back", "left", "right", etc. are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0058] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0059] The embodiment of the present invention provides compression applied to various existing convolutional neural network models, which is used to improve the accuracy and stability of TFI classification. The core idea is to compress the resolution of the output feature map without increasing the depth and complexity of the compressed convolutional neural network model. That is, the embodiment of the present invention provides a new convolutional neural network design concept. At the same time, the compression method provided by the embodiment of the present invention can be applied to multiple types of CNN models, and the specific CNN model includes a residual network (ResNet) model of different depths, where the ResNet is composed of basic blocks as basic units. It also includes ResNet50 composed of a bottleneck structure as a basic unit, as well as the classic open source models GoogleNet and MobileNetV2 designed by Google.

[0060] It needs to be further explained that there are certain differences in the improvement methods for different network models, but the overall design idea remains unchanged, that is, to weaken the adverse effects of noise by compressing the resolution of the output feature map and highlight the effective feature signals in the output feature map.

[0061] First, see Figure 1 and Figure 2 A typical residual network is composed of basic residual blocks, wherein each residual block consists of two convolutional layers, and the original convolution kernels of the two convolutional layers are both 3*3. The embodiment of the present invention provides a compressed convolutional neural network, specifically a compressed residual network.

[0062] The compressed convolutional network is generated by the following steps: compressing the last residual block, before compression, the output feature maps of the two convolutional layers of the last residual block have the original resolution. As an example, the original resolution can be 7*7.

[0063] For the last residual block, the input feature map of the penultimate convolutional layer is not padded, that is, the convolution operation is performed directly using the original 3*3 convolution kernel.

[0064] Set the convolution step size of the penultimate convolution layer. The convolution step size can be set to 1 or 2.

[0065] The input feature map is convolved using the original convolution kernel of the penultimate convolution layer to obtain an output feature map, which has a compressed resolution smaller than the original resolution. As an example, when the convolution step size is 1, the original convolution kernel is 3*3, and the compressed resolution is 5*5. As a second example, when the original resolution is 7*7, the convolution step size is 2, and the original convolution kernel is 3*3, the compressed resolution is 3*3.

[0066] Second, see Figure 3 , an embodiment of the present invention provides a second compressed convolutional neural network, the convolutional neural network is ResNet50, ResNet50 is composed of bottleneck as a basic unit, and the bottleneck includes three layers of convolution. The compressed convolutional neural network provided by the embodiment of the present invention only modifies the last bottleneck structure. The last bottleneck structure contains three convolutional layers, namely, a convolutional layer with an original convolution kernel size of 1*1, a convolutional layer with an original convolution kernel size of 3*3, and a convolutional layer with an original convolution kernel size of 1*1. The output feature map resolution is 7*7. According to the different output feature map resolutions of the compressed convolutional neural network, there are two compression methods:

[0067] In the first compression method, the input feature map of the 3*3 convolution layer is not padded, and the convolution operation is performed directly using the original 3*3 convolution kernel with a convolution step size of 1. Finally, the output feature map size is 5*5.

[0068] The second compression method does not fill the input feature map of the 3*3 convolution layer, but directly uses the 3*3 original convolution kernel to perform the convolution operation, sets the convolution step size to 2, and obtains an output feature map size of 3*3.

[0069] Thirdly, see Figure 4, an embodiment of the present invention provides a third compressed convolutional neural network, and the convolutional neural network is GoogleNet. GoogleNet is composed of inception as a basic unit. Inception includes 4 parts, a bottleneck structure including three convolutional layers and a maximum pooling layer with a size of 3*3. The three convolutional layers are respectively a convolutional layer with an original convolution kernel size of 1*1, a convolutional layer with an original convolution kernel size of 3*3, and a convolutional layer with an original convolution kernel size of 5*5. The working principles are: performing different convolution kernel pooling operations on the input feature map to obtain different output feature maps, and finally fusing the feature maps, that is, taking concatenation. The compressed convolutional neural network provided in the embodiment of the present invention only modifies the last inception structure, and the resolution of its output feature map is 7*7. According to the different resolutions of the output feature map of the compressed convolutional neural network, it can be divided into two compression methods. The specific operation steps are as follows:

[0070] In the first compression method, padding is performed on the input feature map of the 1*1 convolution layer, and a pixel of 0 value is padded around the feature map, and then the convolution stride is set to 2. Padding is performed on the input feature map of the 5*5 convolution layer, and a pixel of 0 value is padded around the feature map, and then the convolution stride is set to 1. The input feature map of the 3*3 convolution layer is not padded, and the convolution operation is performed directly using the 3*3 convolution kernel, and the convolution stride is 1. The input feature map of the 3*3 maximum pooling layer is not padded, and the pooling operation is performed directly, and the pooling stride is 1, and finally the output feature map size is 5*5.

[0071] The second compression method does not pad the input feature map of the 1*1 convolution layer, but directly uses the 1*1 convolution kernel to perform the convolution operation, with a stride of 3. The input feature map of the 3*3 convolution layer is not padded, but directly uses the 3*3 convolution kernel to perform the convolution operation, with a stride of 2. The input feature map of the 5*5 convolution layer is not padded, but directly uses the 5*5 convolution kernel to perform the convolution operation, with a stride of 1. The input feature map of the 3*3 maximum pooling layer is not padded, but directly performs the pooling operation, with a stride of 2, and finally the output feature map size is 3*3.

[0072] Fourthly, see Figure 5, an embodiment of the present invention provides a fourth compressed convolutional neural network, and the convolutional neural network is MobileNetV2. MobileNetV2 is composed of a linear bottleneck structure (inverted bottleneck) with an inverted residual as a basic unit, and an inverted bottleneck structure includes three layers of convolution. An embodiment of the present invention provides a compressed convolutional neural network that only modifies the last bottleneck structure. The last bottleneck structure contains three convolutional layers, namely: a convolutional layer with an original convolution kernel size of 1*1, a deep convolutional layer with an original convolution kernel size of 3*3, and a convolutional layer with an original convolution kernel size of 1*1. The output feature map resolution is 7*7. According to the different output feature map resolutions of the compressed convolutional neural network, it can be divided into two compression methods. The specific operation steps are as follows:

[0073] The first compression method does not fill the input feature map of the 3*3 convolution layer, but directly uses the 3*3 depth original convolution kernel to perform the convolution operation, with a convolution step length of 1, and finally obtains an output feature map size of 5*5.

[0074] The second compression method does not fill the input feature map of the 3*3 convolution layer, but directly uses the 3*3 depth original convolution kernel to perform the convolution operation, sets the stride to 2, and obtains an output feature map size of 3*3.

[0075] The above are four different types of compressed convolutional neural networks. When applied to radar emitter signal recognition, they include the following steps:

[0076] S1. Receive a radar radiation source signal and preprocess the radar radiation source signal to obtain a preprocessed image.

[0077] S2. Use the preprocessed image as an input feature map and use any compressed convolutional neural network provided in the first to fourth aspects for recognition.

[0078] S3. Get the category of the input image.

[0079] Preprocessing includes CWD time-frequency transformation, grayscale conversion, and linear interpolation to adjust the time-frequency image size. The CWD, namely Choi-Williams Distribution, represents Choi-Williams Distribution. CWD can be mathematically expressed as:

[0080]

[0081] Where t and w represent time and angular frequency, and σ is the attenuation coefficient. In order to obtain high time-frequency resolution and suppress cross terms, σ is configured to be 1.

[0082] S1 includes the following sub-steps:

[0083] S10, performing CWD time-frequency transformation on various radar radiation source signals to obtain a transformed time-frequency diagram;

[0084] S11, graying the transformed time-frequency image to obtain a grayscale image;

[0085] The grayscale processing of S11 retains the complete information of the time-frequency image under various signal-to-noise ratio conditions.

[0086] S12, adjusting the grayscale image obtained in S11, resizing the image by using an interpolation method, and obtaining a preprocessed image with a size of 224*224.

[0087] Compared with the prior art, the compressed convolutional neural network provided by the present invention does not change the parameter complexity and floating-point precision calculation complexity of the model. It is a simple and flexible method. The compressed convolutional neural network can effectively suppress the noise signal in the output feature map, highlight the feature signal that is crucial to classification and recognition, and solves the defects of difficulty in identifying radar radiation sources under low signal-to-noise ratio conditions, high complexity of the recognition method and low recognition performance, and has a good recognition effect on radar radiation source signals with greater similarity.

[0088] In order to verify the beneficial effects of the compressed convolutional neural network provided by the embodiment of the present invention, it is explained below in conjunction with specific embodiments.

[0089] 1. Experimental Environment

[0090] The experiments conducted in the embodiments of the present invention are equipped with a computer with AMD Ryzen 5 5600K CPU, 16GB memory and RTX3060GPU hardware capabilities, and the simulation software is MATLAB 2022a software.

[0091] 2. Experimental Dataset Construction

[0092] Generally speaking, in most studies, the signal-to-noise ratio of the data set is -10dB-10dB, and only a few types of easily confused signals are included. In order to be closer to the complex electromagnetic environment and better conduct comparative experiments, the embodiment of the present invention adopts two measures to construct the data set. One measure is to reduce the signal-to-noise ratio of the data set, and the other measure is to increase the difficulty of signal classification by simulating more types of easily confused signals. The pulse width of the simulated signal is 2μs, the sampling frequency fs is 400MHz, and the configuration of other parameters is detailed in Table 1, which is randomly selected within the setting range. The signal-to-noise ratio ranges from -16 to 0dB, with a step size of 2dB, and 1200 signals of various types are randomly generated at each signal-to-noise ratio. All samples are divided into training data set, verification data set and test data set in a ratio of 4:1:1.

[0093] Table 1 Signal parameter settings

[0094]

[0095]

[0096] The radar radiation source signal needs to be preprocessed to obtain a preprocessed image. The preprocessing includes CWD time-frequency transformation, grayscale conversion, and adjustment of the time-frequency image size; the CWD, namely Choi-Williams Distribution, represents Choi-Williams Distribution, which can be mathematically expressed as:

[0097]

[0098] Where t and w represent time and angular frequency, and σ is the attenuation coefficient. In order to obtain high time-frequency resolution and suppress cross terms, σ is configured to be 1.

[0099] 3. Design Concept

[0100] In natural image classification, tiny feature maps retain only a small amount of spatial information. It causes a lot of information loss in the classification of complex natural scenes or images. Therefore, the classification and recognition of natural images usually adopts measures to increase the resolution of the output feature map. The image characteristics shown by TFI are opposite to those of natural images. TFI has few subtle features and high noise energy. The key idea of ​​the embodiment of the present invention is to suppress noise interference in the output feature map. Although traditional models have many successful experiments in ordinary image classification, the inventors analyze that 7*7 or 28*28 feature maps are not suitable for TFI classification. Therefore, the 7*7 output feature map contains a large amount of noise energy, which is extremely unfavorable for the model to extract the target features of interest. Worse, after the global average pooling operation, the feature information may be masked by the noise information. If the present invention improves the resolution of the output feature map, the gain obtained is much less than the loss, and the decline in model accuracy will be aggravated.

[0101] Therefore, the present invention proposes a hypothesis: whether it is possible to break the typical configuration of the output feature map resolution and compress it. Reducing the output feature map resolution can indeed suppress the noise signal, but there is an adverse side effect: the effective feature signal will also be suppressed. Therefore, the gains and losses need to be carefully considered. In other words, the radar signal TFI has obvious surface features, similar to spatial structural connections, while only a small part of the TFI is the effective signal area. Therefore, the classification accuracy of TFI mainly depends on the interference of noise because the features of TFI are simple and few. In contrast, ordinary images usually have multiple objects, and their identities and relative configurations are important for classification. Therefore, ordinary images have more features than TFI, especially subtle features, and the classification accuracy of ordinary images mainly depends on the number of key features.

[0102] In view of the comparative analysis of TFI and ordinary images, the inventors believe that the gains from compressing the output feature map resolution far outweigh the losses. In addition, if a smaller feature map is used, the global average pool will be more likely to process feature information, and then the key features will dominate the output feature map. For example, if a 3×3 feature map is used, only 9 / 49 times the information needs to be processed compared to the original feature map, and the problem of the "curse of dimensionality" will be greatly alleviated. And the tiny output feature map will include most of the key features.

[0103] Example 1

[0104] This embodiment involves testing compressed residual networks of different depths with a basic block structure as the basic component unit. In order to form a performance comparison, the original residual networks are named: ResNet-10, ResNet-18, ResNet-26, ResNet-34, and the numbers represent the model depth. The compressed residual network provided by the embodiment of the present invention has two models according to each depth, namely: a model with an output feature map resolution of 5*5, named ResNet-10-1, ResNet-18-1, ResNet-26-1, ResNet-34-1 according to the depth of the model; a model with an output feature map resolution of 3*3, named ResNet-10-2, ResNet-18-2, ResNet-26-2, ResNet-34-2 according to the depth of the model.

[0105] from Figure 6It can be seen that the classification accuracy of the ResNet model increases with the increase of the model depth. The experimental results are very consistent with the results of the residual network applied to the field of image recognition. The scientific nature of the ResNet architecture of different depths designed in the embodiment of the present invention is verified. Importantly, the compressed model has better performance than the corresponding original model. For example, the accuracy of CRN-10-1 and CRN-10-2 is 3.07% and 3.88% higher than that of ResNet-10, respectively, the accuracy of CRN-18-1 and CRN-18-2 is 0.63% and 1.28% higher than that of ResNet-18, respectively, the accuracy of CRN-26-1 and CRN-26-2 is 0.34% and 0.65% higher than that of ResNet-26, respectively, and the accuracy of CRN-34-1 and CRN-34-2 is 0.49% and 0.5% higher than that of ResNet-34, respectively. The classification accuracy of CRN-10-2 is better than ResNet-18, but the complexity of CRN-10-2 is much lower than ResNet-18. This verifies that the compression method provided by the embodiment of the present invention can be well applied to ResNet models of different depths. In addition, under the conditions of similar depth and complexity, it is difficult for traditional models to achieve the high level of accuracy of the compressed model.

[0106] In order to further verify the stability and robustness of the compression model described in the embodiment of the present invention, four models are selected to identify radar signals under 14 different signal-to-noise ratio conditions. The four models are: CRN-18-1, CRN-18-2, ResNet-18 and DRN-18. It is worth noting that DRN-18 was designed in 2017. Its design concept is opposite to that of this article, and measures are taken to expand the resolution of the output feature map. The output feature map resolution of DRN-18 is 28*28. Apart from this, its parameter quantity and complexity are the same as those of ResNet-18.

[0107] from Figure 7 It can be seen that under different signal-to-noise ratio conditions, both CRN models are greater than or equal to the corresponding ResNet model and DRN model, despite the same depth and volume. It is worth noting that from the trends of the four curves, it can be seen that when the signal-to-noise ratio is lower than -10dB, the accuracy of DRN-18 and ResNet-18 decreases faster than the two CRN models. In other words, the lower the signal-to-noise ratio, the more obvious the recognition performance advantage of CRN. For example, when the signal-to-noise ratio is -10dB, the accuracy of CRN-18-1 and CRN-18-2 is 0.63% and 1.28% higher than the corresponding ResNet-18, respectively, and when the signal-to-noise ratio is -16dB, the accuracy gap increases to 2.00% and 5.14%, respectively.

[0108] Comparing the recognition of different signals by the four models, it can be seen that under different signal-to-noise ratio conditions, the classification accuracy of CRN-18-1 and CRN-18-2 for BFSK, BPSK, CW, FRANK, LFM, NLFM, P1, P3 and QPSK is greater than or equal to the corresponding ResNet-18 and DRN-18. However, for the classification of COSTAS, QFSK and P2, the classification accuracy of CRN-18-1, CRN-18-2 and ResNet-18 is similar. By comparison, the CRN model has no obvious advantage.

[0109] For signals that are easier to classify, CRN-18-1, CRN-18-2, and ResNet-18 all have similar performance, with higher accuracy at higher SNRs. For example, when the SNR is above -10dB, the classification accuracy of the three models is close to or equal to 100% for the classification of BPSK, CW, COSTAS, LFM, NLFM, P4, QFSK, and QPSK. However, when the SNR is extremely low, the accuracy of ResNet-18 drops faster than CRN-18-1 and CRN-18-2. For example, for the classification of BPSK and CW, when the SNR is -16dB, the accuracy of CRN-18-2 is 5.5% and 4% higher than ResNet-18, respectively.

[0110] For signals that are relatively difficult to classify, the classification accuracy of all models is relatively low under various conditions of signal-to-noise ratio. For example, when the signal-to-noise ratio is higher than -2dB, the classification accuracy of CRN-18-1, CRN-18-2, and ResNet-18 for FMCW, P1, P2, P3, and P4 is close to 100%. Importantly, when the signal-to-noise ratio gradually decreases, the performance gap is further amplified compared to signals that are easy to classify. For example, for the classification of FMCW and P1, the accuracy of CRN-18-2 is 23% and 11% higher than that of ResNet-18, respectively.

[0111] In addition, under various SNR conditions, DRN-18's classification performance for all signals is lower than that of the other three models. Worse, DRN-18 has particularly low classification accuracy for easily confused signals, such as P1, P2, and P4. Especially under low SNR conditions, DRN-18 cannot classify easily confused signals. It can be seen that enlarging the output feature map resolution has an adverse effect on the performance of the model. At the same time, it verifies the rationality of our method of compressing the output feature map resolution from the side.

[0112] The above results show that under extremely low signal-to-noise ratio conditions, the compression model provided by the embodiment of the present invention has obvious advantages over the non-compression model, especially for the classification of easily confused signals. Through the analysis and comparison of the above experimental results, the following conclusions can be drawn: ① Noise energy will destroy the characteristics of TFI to a certain extent, which is harmful to classification, and the intensity of noise energy is inversely proportional to the classification accuracy; ② The design of the resolution of the compressed output feature map can suppress the adverse effects of noise to a certain extent, which is conducive to improving the classification accuracy.

[0113] The CRN model is more suitable for a class of objects with relatively few features or containing high noise energy. However, it should be noted that the degree of compression of the output feature map resolution is related to the object characteristics. For example, if the noise energy in the image increases, a smaller output feature map resolution should be used. On the contrary, if the image has a large number of features and less noise, the DRN model should be used for classification.

[0114] Example 2

[0115] This embodiment involves testing a compressed residual network with a bottleneck structure as a basic unit. The control group model is ResNet-50. The experimental group is two types of compression models designed using the compression method of the embodiment of the present invention, namely: a model with an output feature map resolution of 5*5, named ResNet-50-1; a model with an output feature map resolution of 3*3, named ResNet-50-2.

[0116] Figure 8 The comparison results of the three models are shown. Obviously, all compressed models have better performance than the corresponding original models. The average recognition rates of ResNet-50-2 and ResNet-50-1 models are 0.76% and 0.45% higher than those of ResNet-50 models, respectively.

[0117] Example 3

[0118] This embodiment involves testing GoogleNet with the inception structure as the basic unit. The control group model is GoogleNet, designed by Google in 2015. The experimental group is two types of compression models designed using the compression method of this article, namely: a model with an output feature map resolution of 5*5, named GoogleNet-1; a model with an output feature map resolution of 3*3, named GoogleNet-2.

[0119] Figure 8The comparison results of the three models are shown. Obviously, all compressed models have better performance than the corresponding original models. The GoogleNet-2 and GoogleNet-1 models have an average recognition rate that is 4.75% and 1.39% higher than the GoogleNet model, respectively.

[0120] Example 4

[0121] This embodiment involves testing MobileNetV2 with an inverted bottleneck structure as the basic unit. The control group model is MobileNetV2, designed by Google in 2018. The experimental group is two types of compression models designed using the compression method in this article, namely: a model with an output feature map resolution of 5*5, named MobileNetV2-1; a model with an output feature map resolution of 3*3, named MobileNetV2-2.

[0122] Figure 8 The comparison results of the three models are shown. Obviously, all compressed models have better performance than the corresponding original models. The MobileNetV2-2 and MobileNetV2-1 models have an average recognition rate that is 1.21% and 0.49% higher than the MobileNetV2 model, respectively.

[0123] The embodiment of the present invention relates to a radar emitter signal recognition device based on a compressed convolutional neural network, comprising at least one processor and a memory connected to the at least one processor in communication. The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above-mentioned radar emitter signal recognition method based on a compressed convolutional neural network.

[0124] Among them, the memory and the processor are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripheral devices, voltage regulators and power management circuits through interfaces, which are well known in the art. The interface provides an interface between the bus and the transceiver, such as a communication interface and a user interface. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium through an antenna, and further, the antenna also receives data and transmits the data to the processor. The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management and other control functions. The memory can be used to store data used by the processor when performing operations.

[0125] The first embodiment of the present invention relates to a computer-readable storage medium storing a computer program, and the computer program implements the above method embodiment when executed by a processor.

[0126] Those skilled in the art can understand from the above description that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions for making a device (which can be a single-chip microcomputer, chip, etc.) or a processor (processor) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes but is not limited to various media that can store program codes, such as USB flash drives, mobile hard disks, magnetic storage devices, and optical storage devices.

[0127] It can be seen from the above description that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the accompanying drawings. Any equivalent or modification that is completed without departing from the spirit disclosed in the present invention falls within the scope of protection of the present invention.

[0128] In the description of the above embodiments, specific features, structures, materials or characteristics may be combined in a suitable manner in any one or more embodiments or examples.

[0129] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A compressed convolutional neural network, characterized in that: The compressed convolutional neural network is a residual network, the residual network includes a plurality of residual blocks, and each of the residual blocks includes two sequentially connected convolutional layers; The compressed convolutional network is generated by the following steps: compressing the last residual block, where before compression, the output feature maps of the two convolutional layers of the last residual block have the original resolution; For the last residual block, the input feature map input to the penultimate convolutional layer is not padded; Set the convolution step size of the penultimate convolutional layer; A convolution operation is performed on the input feature map using an original convolution kernel of the penultimate convolution layer to obtain an output feature map, where the output feature map has a compressed resolution smaller than the original resolution.

2. The compressed convolutional neural network according to claim 1, characterized in that The original resolution is 7*7, the convolution step is 1, the original convolution kernel is 3*3, and the compressed resolution is 5*5; or, The original resolution is 7*7, the convolution step is 2, the original convolution kernel is 3*3, and the compressed resolution is 3*3.

3. A compressed convolutional neural network, characterized in that: The convolutional neural network is ResNet50, and the ResNet50 includes a plurality of bottleneck structures, each of which includes three sequentially connected convolutional layers; The compressed convolutional neural network is generated by the following steps: compressing the last bottleneck structure, where before compression, the output feature maps of the three convolutional layers of the last bottleneck structure have the original resolution; For the last bottleneck structure, the input feature map input to the penultimate convolutional layer is not padded; Set the convolution step size of the penultimate convolutional layer; A convolution operation is performed on the input feature map using an original convolution kernel of the penultimate convolution layer to obtain an output feature map, where the output feature map has a compressed resolution smaller than the original resolution.

4. The compressed convolutional neural network according to claim 3, characterized in that The original convolution kernels of the three convolution layers from top to bottom are 1*1, 3*3 and 1*1 respectively, the original resolution is 7*7, the convolution step is 1, and the compressed resolution is 5*5; or, the original resolution is 7*7, the convolution step is 2, and the compressed resolution is 3*3.

5. A compressed convolutional neural network, characterized in that: The compressed convolutional neural network is GoogleNet, which is composed of inception as a basic unit; each inception includes three parallel convolutional layers and one maximum pooling layer; the original convolution kernels of the three convolutional layers are 5*5, 3*3 and 1*1 respectively, and the size of the maximum pooling layer is 3*3; before compression, the output feature maps of the three convolutional layers and one maximum pooling layer have the same original resolution; the compressed convolutional neural network is generated by the following steps: Filling the input feature map of the convolution layer with the original convolution kernel as 1*1, setting the first convolution step size, performing the convolution operation, and obtaining the first output feature map; Fill the input feature map of the convolution layer with the original convolution kernel of 5*5, set the second convolution step size, perform the convolution operation, and obtain the second output feature map; The input feature map of the convolution layer whose original convolution kernel is 3*3 is not filled, a third convolution step size is set, a convolution operation is performed, and a third output feature map is obtained; The input feature map of the maximum pooling layer with a size of 3*3 is not filled, the first pooling step size is set, and the pooling operation is performed to obtain the fourth output feature map; concatenating the first output feature map, the second output feature map, the third output feature map, and the fourth output feature map to obtain an output feature map, wherein the output feature map has a first compressed resolution that is smaller than the original resolution; or, The compressed convolutional neural network is generated by the following steps: The input feature map of the convolution layer whose original convolution kernel is 1*1 is not filled, a fourth convolution step size is set, a convolution operation is performed, and a fifth output feature map is obtained; The input feature map of the convolution layer whose original convolution kernel is 5*5 is not filled, the fifth convolution step size is set, and the convolution operation is performed to obtain the sixth output feature map; The input feature map of the convolution layer whose original convolution kernel is 3*3 is not filled, the sixth convolution step size is set, and the convolution operation is performed to obtain the seventh output feature map; The input feature map of the maximum pooling layer with a size of 3*3 is not filled, the second pooling step size is set, and the pooling operation is performed to obtain the eighth output feature map; The fifth output feature map, the sixth output feature map, the seventh output feature map and the eighth output feature map are concatenated to obtain an output feature map having a second compressed resolution that is smaller than the original resolution.

6. The compressed convolutional neural network according to claim 5, characterized in that The first convolution step size is 2, the second convolution step size is 1, the third convolution step size is 1, the first pooling step size is 1, and the first compression resolution is 5*5; The fourth convolution step size is 3, the fifth convolution step size is 1, the sixth convolution step size is 2, and the second pooling step size is 2. At this time, the first compression resolution is 3*3.

7. The compressed convolutional neural network according to claim 5, characterized in that Filling the input feature map of the convolution layer whose original convolution kernel is 1*1 includes: filling the input feature map with a value of 0 for one pixel; Filling the input feature map of the convolution layer whose original convolution kernel is 5*5 includes: filling the input feature map with a value of 0 for one pixel.

8. A compressed convolutional neural network, characterized in that The compressed convolutional neural network is MobileNetV2, and the MobileNetV2 is composed of a linear bottleneck structure with an inverted residual as a basic unit; one of the linear bottleneck structures includes three convolutional layers connected in sequence; The compressed convolutional neural network is generated by the following steps: compressing the last linear bottleneck structure, where before compression, the output feature maps of the three convolutional layers of the last bottleneck structure have the original resolution; For the last linear bottleneck structure, the input feature map input to the penultimate convolutional layer is not padded; Set the convolution step size of the penultimate convolutional layer; A convolution operation is performed on the input feature map using an original convolution kernel of the penultimate convolution layer to obtain an output feature map, where the output feature map has a compressed resolution smaller than the original resolution.

9. The compressed convolutional neural network according to claim 8, characterized in that The original convolution kernels of the three convolution layers from top to bottom are 1*1, 3*3 and 1*1 respectively, the original resolution is 7*7, the convolution step is 1, and the compressed resolution is 5*5; or, the original resolution is 7*7, the convolution step is 2, and the compressed resolution is 3*3.

10. A radar radiation source signal recognition method, characterized in that: The steps include: Receiving a radar radiation source signal, and preprocessing the radar radiation source signal to obtain a preprocessed image; The preprocessed image is used as an input feature map, and the compressed convolutional neural network described in any one of claims 1 to 9 is used for recognition to obtain the category of the input image.

Images