Intelligent target recognition method based on multi-modal characteristic fusion

Abstract

The application discloses a kind of multi-modal characteristic fusion's intelligent target identification method, on the basis of constructing typical aerial target high-resolution range image, infrared image sample library, respectively using attention bidirectional gate recurrent unit model from high-resolution range image time series sample, using light convolutional neural network model from infrared image sample learning extraction aerial target multi-modal deep feature representation;Using Gaussian distribution limited Boltzmann machine realizes aerial target multi-modal feature abstract fusion, removes redundancy, forms more discernibility, representative multi-modal feature joint representation;Using kernel ultra-limit learning machine with strong generalization ability under small sample condition, fast training speed and high classification accuracy as classifier, realize aerial target type identification.The application uses heterogeneous deep learning model to abstract extraction and fusion aerial target feature information in high-resolution range image, infrared image for target identification, improve recognition precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing and target recognition, and particularly relates to an intelligent target recognition method based on multi-modal characteristic fusion. BACKGROUND

[0002] The prior art mainly studies target recognition methods based on radar high resolution range profile (HRRP) or synthetic aperture radar (SAR) images and target recognition methods based on infrared images, which reflect some characteristics of the target and are not comprehensive enough, thus easily causing inaccurate target interpretation and recognition. SUMMARY

[0003] In order to solve or partially solve the problems in the related art, the present application provides an intelligent target recognition method based on multi-modal characteristic fusion, which uses a heterogeneous deep learning model to abstractly extract and fuse high resolution range profile (HRRP) and infrared image target feature information for target recognition, thereby improving the recognition accuracy.

[0004] The present application provides an intelligent target recognition method based on multi-modal characteristic fusion, comprising:

[0005] Obtaining multi-modal feature data samples of a target to be recognized in the air; wherein the multi-modal feature data samples comprise HRRP samples and infrared image samples;

[0006] Constructing and training a heterogeneous deep learning network for HRRP and infrared image target feature learning;

[0007] Training a Gaussian distribution restricted Boltzmann machine for HRRP and infrared image target feature fusion learning;

[0008] Training a kernel extreme learning machine classifier based on HRRP and infrared image deep fusion features.

[0009] Optionally, the obtaining of the multi-modal characteristic data samples of the target to be recognized in the air specifically comprises:

[0010] Obtaining HRRP samples of the target to be recognized in the air and performing normalization preprocessing thereon; and / or

[0011] Obtaining infrared image samples of the target to be recognized in the air and performing normalization preprocessing thereon.

[0012] Optionally, the obtaining of the HRRP samples of the target to be recognized in the air and the normalization preprocessing thereon specifically comprises:

[0013] The HRRP sample set of the aerial target to be identified is obtained by experimental measurement or electromagnetic calculation software simulation. Each HRRP sample is a 1×N array centered on the aerial target to be identified.

[0014] For each HRRP sample, the 2-norm normalization method is first used to eliminate the scattering amplitude sensitivity, and then the centroid alignment method is used to eliminate the translation sensitivity of the HRRP, resulting in an amplitude and position normalized HRRP sample. k is the sample number;

[0015] Define a one-dimensional sliding window with a width of d and a sliding step size of d. s By using a sliding window method, the amplitude and location-normalized HRRP samples are truncated to generate a set of sequence data with a time dimension of T and a data dimension of d at each time step, denoted as X. k =[x1,x2,...,x T ], where k is the sample number, and the sequence value at time i is

[0016] For each HRRP sample in the aerial target HRRP sample set, amplitude and position are normalized, and serialization extraction is performed to form a normalized HRRP time series dataset. From this dataset, p% of the samples are randomly selected as the training sample set, where p is an integer in [50, 80], and the remaining samples are used as the test sample set.

[0017] Optionally, the step of acquiring infrared image samples of the aerial target to be identified and performing normalization preprocessing specifically includes:

[0018] Infrared image samples of the aerial target to be identified are obtained by capturing infrared video of the aerial target, experimental measurement, or simulation by infrared image calculation software. Each image sample is an infrared image slice with a size of M×M pixels centered on the target, and the observation orientation is the same as or similar to that of the HRRP sample.

[0019] Each infrared image sample is normalized according to Equation (1) and its target type is marked to form a standardized infrared image sample library of aerial targets to be identified.

[0020]

[0021] In the formula, A represents the original infrared image sample numbered k; k To Image samples after amplitude normalization; μ k for The mean of all pixel values for The variance of all pixel values; the sample serial number catalog in the database is consistent with the corresponding HRRP sample serial number according to the observation direction similarity;

[0022] From the infrared image sample library of the air target to be identified, the sample with the same sample number as the HRRP training sample is extracted as the training sample set, and the remaining samples are used as the test sample set.

[0023] Optionally, the construction and training of the heterogeneous deep learning network for HRRP and infrared image air target feature learning specifically includes:

[0024] The attention bidirectional gated recurrent unit (BGRU) model for HRRP air target feature learning is constructed and trained.

[0025] The convolutional neural network (CNN) model for infrared image air target feature learning is constructed and trained.

[0026] Optionally, the construction and training of the attention bidirectional gated recurrent unit model for HRRP air target feature learning specifically includes:

[0027] A bidirectional gated recurrent unit network (BGRU) composed of an input layer, a gated recurrent unit (GRU) forward propagation layer, a GRU backward propagation layer, and a hidden layer output is constructed.

[0028] Each HRRP time series sample is simultaneously input to the GRU forward propagation layer and the GRU backward propagation layer.

[0029] The hidden layer outputs of the GRU forward propagation layer and the GRU backward propagation layer are combined to obtain the BGRU hidden layer output.

[0030] The weight coefficients representing the importance of the hidden layer nodes mapped by different distance units in the HRRP to target recognition are obtained through the attention mechanism, and then the BGRU hidden layer output is weighted and summed using the weight coefficients to obtain the HRRP time series feature representation.

[0031] The HRRP time series feature representation vector is input to the Softmax classifier, and based on the Softmax classification result, the attention bidirectional gated recurrent unit model parameters and the attention weight coefficient solving parameters are updated using the backpropagation method (BPTT) along the time dimension.

[0032] Through continuous iteration until the model converges, an attention bidirectional gated recurrent unit model for HRRP air target feature learning is obtained, marked as BGRU HRRP , and the network parameters of the model are solidified and stored.

[0033] Optionally, the convolutional neural network model for infrared image air target feature learning is constructed and trained, and specifically includes the following steps.

[0034] A CNN model composed of three groups of convolutional structures and one group of fully connected structures is constructed, wherein each group of convolutional structures is composed of a convolutional layer, a batch normalization layer and a ReLu activation layer; and the group of fully connected structures is composed of a fully connected layer, a ReLu activation layer and a Dropout layer.

[0035] Each infrared image training sample is input into the CNN model for training, and the output vector of the fully connected structure is input into a Softmax classifier.

[0036] Based on the Softmax classification result, the parameters of the CNN model are updated by using a back propagation method.

[0037] Through continuous iteration until the model converges, a trained convolutional neural network model for infrared image air target feature learning is obtained, marked as CNN Infra , and the convolutional neural network model parameters are solidified and stored.

[0038] Optionally, the Gaussian distribution restricted Boltzmann machine for air target feature fusion learning of HRRP and infrared image is trained, and specifically includes the following steps.

[0039] Each HRRP time series sample is input into the trained BGRU HRRP model, and the output vector of the HRRP time series feature representation layer obtained is marked as k represents the sample number.

[0040] Each infrared image training sample is input into the trained CNN Infra model, and the output vector of the fully connected structure layer obtained is marked as k represents the sample number.

[0041] The top layer output feature vector of the attention BGRU model for HRRP air target feature learning is combined with the top layer output feature vector of the CNN model for infrared image air target feature learning according to formula (2) to form a multi-modal feature vector X k .

[0042]

[0043] The multi-modal feature vector X corresponding to each group of HRRP, infrared image training sample k , is input into a 1-layer Gaussian distribution restricted Boltzmann machine (GRBM) for training, the parameters of the GRBM are saved, and the output vector thereof is input into a Softmax classifier; based on the Softmax classification result, the weight parameters of the GRBM are fine-tuned by using a back propagation method; in this way, through continuous iteration, until the GRBM converges, a trained multi-modal feature fusion learning model is obtained, and the parameters of the GRBM are solidified and saved.

[0044] Optionally, the training kernel extreme learning machine classifier based on HRRP and infrared image deep fusion features specifically comprises the following steps.

[0045] The multi-modal feature vector X corresponding to each group of HRRP, infrared image training sample k , is input into the trained 1-layer GRBM, the output multi-modal deep fusion feature vector thereof is input into a kernel extreme learning machine (KELM) classifier, the KELM classifier is trained, a trained KELM classifier is obtained, and the parameters of the KELM are solidified and saved.

[0046] Optionally, the method further comprises the following steps.

[0047] The target recognition test specifically comprises the following steps.

[0048] Each group of HRRP, infrared image test sample of an air target, containing a group of HRRP time sequence samples and an infrared image test sample, is input into the trained multi-modal feature fusion air target recognition model, the type of the air target is determined, and until the type of all test samples of the air target is recognized.

[0049] The technical scheme provided by the application can have the following beneficial effects.

[0050] The multi-modal composite detection system represented by radar / infrared composite detection adopts radar and infrared two kinds of sensor devices to detect target information in the same detection process, and realizes more accurate target detection and recognition through fusion and complementation of radar echo, infrared image and corresponding transform domain feature information.

[0051] The designed heterogeneous deep learning recognition network of the application comprehensively considers the time sequence characteristics of high-resolution range profile (HRRP) data of an aerial target and the internal target structure information, and the characteristics of a small target in an infrared image, and can realize automatic abstraction learning and fusion of multi-dimensional features of the aerial target from the HRRP and the infrared image for recognition, and improve the target classification and recognition precision.

[0052] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory and are not restrictive of the application. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the patent embodiments of the application, the drawings needed for the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0054] Figure 1 The flow chart of the intelligent target recognition method of multi-modal characteristic fusion in the embodiment of the application. DETAILED DESCRIPTION

[0055] The embodiments of the application will be described in detail with reference to the accompanying drawings. Although the embodiments of the application are shown in the drawings, it should be understood that the application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the application more thorough and complete, and to fully convey the scope of the application to those skilled in the art.

[0056] The terms used in the application are only for the purpose of describing specific embodiments and are not intended to limit the application. The singular forms "said" and "the" used in the application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0057] It should be understood that, although the terms "first", "second", "third", etc. can be used herein to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information without departing from the scope of the present application. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0058] The technical solutions of the embodiments of the present application will be described in detail below with reference to the drawings.

[0059] The embodiment provides a multi-modal characteristic fusion intelligent target recognition method. On the basis of constructing a typical air target HRRP and an infrared image sample library, an attention BGRU model is used to learn and extract multi-modal deep feature representations of air targets from HRRP time sequence samples, and a lightweight CNN model is used to learn and extract multi-modal deep feature representations of air targets from infrared image samples. Further, a one-layer Gaussian distribution restricted Boltzmann machine (GRBM) is used to realize multi-modal feature abstraction and fusion of air targets, remove redundancy, and form a more recognizable and representative multi-modal feature joint representation. Finally, a kernel extreme learning machine (KELM) with strong generalization ability, fast training speed and high classification precision under a small sample condition is used as a classifier to realize air target type recognition.

[0060] Please refer to Figure 1 , and the specific implementation process will be given below in combination with an air target recognition example, and the specific implementation process is as follows.

[0061] Step 1: Obtain multi-modal characteristic data samples of an air target to be recognized.

[0062] Step 1a: Obtain HRRP samples of an air target to be recognized and perform normalization preprocessing.

[0063] The HRRP sample set of three typical air targets, i.e., a helicopter, a passenger plane and a transport plane, is obtained by using an electromagnetic calculation software simulation method. Each HRRP sample is a 1x302 array with a target as the center, and there are 900 samples for each target. The azimuth angle range is [0°, 180°] with an interval of 1°, and the incident angle range is [-60°, 60°] with an interval of 15°.

[0064] For each HRRP sample, first, the 2-norm normalization method is used to eliminate the scattering amplitude sensitivity, and then the centroid alignment method is used to eliminate the translation sensitivity of the HRRP, to obtain an amplitude and position normalized HRRP sample, denoted as: k is the sample number.

[0065] Set a one-dimensional sliding window with a width of d = 22 and a sliding step size of d. s =7, using a sliding window method to truncate the amplitude and location-normalized HRRP samples, generating a set of sequence data with a time dimension of T=40 and a data dimension of d at each time point, denoted as: X k =[x1,x2,...,x T ], where k is the sample number, and the sequence value at time i is

[0066] For each HRRP sample in the three typical aerial target HRRP sample sets, amplitude and position are normalized, and serialization extraction is performed to form a normalized HRRP time series dataset. 70% of the samples in this dataset are randomly selected as the training sample set, and the remaining 30% of the samples are used as the test sample set.

[0067] Step 1b: Acquire infrared image samples of the aerial targets to be identified and perform normalization preprocessing.

[0068] Infrared image simulation software was used to obtain three typical aerial infrared image samples of helicopters, passenger planes and early warning aircraft at observation azimuths similar to those of HRRP samples. Each image sample is a single-channel infrared image slice with a size of 128×128 pixels centered on the ship, with 900 samples for each target.

[0069] Each infrared image sample is normalized according to Equation (1) and its target type is marked to form a standardized infrared image sample library of aerial targets to be identified.

[0070]

[0071] In the formula, A represents the original infrared image sample numbered k; k To Image samples after amplitude normalization; μ k for The mean of all pixel values for The variance of all pixel values. The sample numbers in this database are cataloged according to the similarity of the observed orientation with the corresponding HRRP sample numbers.

[0072] From the infrared image sample library of aerial targets to be identified, samples with the same number as the HRRP training sample are extracted as the training sample set, and the remaining samples are used as the test sample set.

[0073] Step 2: Construct and train a heterogeneous deep learning network for HRRP and infrared image aerial target feature learning.

[0074] Step 2a: Construct and train the attention Bi-directional Gated Recurrent Unit (BGRU) model for HRRP air target feature learning.

[0075] An attention Bi-directional Gated Recurrent Unit (BGRU) network was constructed, which consisted of an input layer, a Gated Recurrent Unit (GRU) forward propagation layer, a GRU backward propagation layer, an attention mechanism layer, and a feature representation layer. The number of input layer neurons was 22, the number of hidden neurons in the GRU forward propagation layer and the GRU backward propagation layer was 40, and the dimensions of the attention mechanism layer and the feature representation layer were 40.

[0076] The initial learning rate was set to 0.01, the adaptive learning rate adjustment (AdaDelta) algorithm was used to update the model parameters, the decay rate was set to 0.97, and the maximum number of training iterations was 3000. Each HRRP time series sample X k = [x1, x2,..., x T ] was input into the attention BGRU model for training. The HRRP time series feature representation vector output by the feature representation layer of the attention BGRU model was input into the Softmax classifier, which had 40 input nodes and 3 output nodes. Based on the Softmax classification results, the BPTT method was used to fine-tune the BGRU model parameters and the attention weight coefficient solving parameters. In this way, through continuous iteration, the trained HRRP air target feature learning attention BGRU model was obtained, which was marked as BGRU HRRP , and the network parameters of the model were saved.

[0077] Step 2b: Construct and train the Convolutional Neural Network (CNN) model for infrared image air target feature learning.

[0078] A CNN model composed of three groups of convolutional structures and one group of fully connected structures is constructed. Each group of convolutional structures is composed of a convolutional layer, a batch normalization layer and a ReLu activation layer. The channel number of the first convolutional layer is C1=32, and the convolution kernel size is N1xN1=23x23. The channel number of the second convolutional layer is C2=64, and the convolution kernel size is N2xN2=15x15. The channel number of the third convolutional layer is C3=128, and the convolution kernel size is N3xN3=7x7. The group of fully connected structures is composed of a fully connected layer, a ReLu activation layer and a Dropout layer. The number of neurons in the fully connected layer is M2=256.

[0079] The learning rate is set to 0.01, and the maximum number of iterations is 3000. Each infrared image training sample is input into the CNN model for training. The output vector of the fully connected structure is input into the Softmax classifier. The input node number of the Softmax classifier is 256, and the output node number is 3. Based on the Softmax classification result, the parameters of the CNN model are updated by the back propagation method. In this way, through continuous iteration, the trained infrared image air target feature learning CNN model is obtained until the model converges, which is marked as CNN Infra , and the CNN model parameters are saved.

[0080] Step 3: Train the Gaussian distribution restricted Boltzmann machine (GRBM) for HRRP and infrared image air target feature fusion learning.

[0081] Each HRRP time series sample is input into the BGRU HRRP model trained in step 2a, and the output vector of the obtained HRRP time series feature representation layer is marked as k represents the sample number.

[0082] Each infrared image training sample is input into the CNN Infra model trained in step 2b, and the output vector of the obtained fully connected structure layer is marked as k represents the sample number.

[0083] The top layer output feature vector of the attention BGRU model for HRRP air target feature learning is combined with the top layer output feature vector of the infrared image air feature learning CNN model according to formula (2) to form a multi-modal feature vector X k with a dimension of 296.

[0084]

[0085] The learning rate is set to 0.01, and the maximum number of iterations is 1000; the multimodal feature vector X corresponding to each group of aerial target HRRP and infrared image training samples is... k The parameters of the GRBM are input into a Gaussian Distribution Restricted Boltzmann Machine (GRBM) with 128 neurons in one layer for training. The parameters of the GRBM are saved, and its output vector is input into a Softmax classifier with 128 input nodes and 3 output nodes. Based on the Softmax classification results, the weight parameters of the GRBM are fine-tuned using the backpropagation method. This process is repeated until the GRBM converges, resulting in a well-trained multimodal feature fusion learning model. The parameters of the GRBM are then saved.

[0086] Step 4: Train a kernel extreme learning machine classifier (KELM) based on HRRP and deep fusion features of infrared images.

[0087] The multimodal feature vector X corresponding to each set of HRRP and infrared image training samples k The input is fed into the 1-layer GRBM trained in step 3, and its output multimodal deep fusion feature vector is input into the Kernel Extreme Learning Machine (KELM) classifier. The KELM classifier has 128 input nodes, 3 output nodes, and a maximum number of iterations of 1000. By training the KELM classifier, the trained KELM classifier is obtained, and the parameters of the KELM are saved.

[0088] Step 5: Target recognition test.

[0089] 270 sets of multimodal test samples, each set containing: a set of HRRP time series samples with similar observation azimuth and a set of infrared test samples, are input into the multimodal feature fusion aerial target recognition model trained in steps 2 to 4 to determine the type of aerial target until the target type recognition of all test samples is completed.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1.A method for intelligent target recognition based on multi-modal feature fusion, characterized in that, The method comprises the following steps: obtaining a multi-modal characteristic data sample of an air target to be identified, wherein the multi-modal characteristic data sample comprises a high-resolution range image sample and an infrared image sample; constructing and training a heterogeneous deep learning network for air target feature learning of the high-resolution range image and the infrared image; training a Gaussian distribution restricted Boltzmann machine for air target feature fusion learning of the high-resolution range image and the infrared image; training a kernel extreme learning machine classifier based on deep fusion features of the high-resolution range image and the infrared image; the training of the Gaussian distribution restricted Boltzmann machine for air target feature fusion learning of the high-resolution range image and the infrared image comprises: each high-resolution range image time series sample is input to the trained BGRU HRRP In the model, the output vector of the obtained high-resolution range image time series feature representation layer is marked as , denotes the sample number; inputting each infrared image training sample into the trained CNN Infra In the model, the output vector of the obtained fully connected structure layer is marked as , denotes a sample number; The top layer output feature vector of the attention BGRU model of high-resolution range image aerial target feature learning The top layer output feature vector of the infrared image aerial feature learning CNN model Combining according to formula (2) to form a multi-modal feature vector ; （2） The multimodal feature vectors corresponding to each set of high-resolution range images and infrared image training samples of aerial targets will be used. The parameters of the GRBM are input into a single-layer Gaussian distributed restricted Boltzmann machine (GRBM) for training. The parameters of the GRBM are saved, and its output vector is input into a Softmax classifier. Based on the Softmax classification results, the weight parameters of the GRBM are fine-tuned using the backpropagation method. This process is repeated until the GRBM converges, resulting in a well-trained multimodal feature fusion learning model. The parameters of the GRBM are then saved. the training of the kernel extreme learning machine classifier based on deep fusion features of the high-resolution range image and the infrared image comprises: The multi-modal feature vector corresponding to each set of high-resolution range image, infrared image training sample , input into the trained 1 layer GRBM, input the multi-modal deep fusion feature vector output by it into the kernel extreme learning machine classifier, train the KELM classifier, obtain the trained KELM classifier, and solidify and save the parameters of the KELM. 2.The intelligent target recognition method of multi-modal characteristic fusion according to claim 1, characterized in that, the obtaining of the multi-modal characteristic data sample of the air target to be identified comprises: obtaining a high-resolution range image sample of the air target to be identified and performing normalization preprocessing thereon; and / or obtaining an infrared image sample of the air target to be identified and performing normalization preprocessing thereon. 3.The intelligent target recognition method of multi-modal feature fusion according to claim 2, characterized in that, the obtaining of the high-resolution range image sample of the air target to be identified and the normalization preprocessing thereon comprises: A high-resolution range image sample set of the to-be-identified air target is acquired, each high-resolution range image sample is centered on the to-be-identified air target, and the size is array; For each high-resolution range image sample, first, the 2-norm normalization method is used to eliminate the scattering amplitude sensitivity, and then the centroid alignment method is used to eliminate the translation sensitivity of the high-resolution range image, to obtain the amplitude and position normalized high-resolution range image sample, , is the sample number; Set the width of the one-dimensional sliding window as d , the sliding step as , and the high-resolution range image samples normalized in amplitude and position are intercepted by the sliding window to generate a sequence of data with the time dimension of and the data dimension of d at each time, denoted as , where is the sample number, and the sequence value at the i th time is ; Each high-resolution range image sample in the high-resolution range image sample set of the aerial target is subjected to amplitude and position normalization, serialized extraction preprocessing, and a normalized high-resolution range image time series data set is formed; a proportion of samples in the data set are randomly extracted as a training sample set, , , an integer, and the remaining samples are used as a test sample set. 4.The intelligent target recognition method of multi-modal feature fusion according to claim 2, characterized in that, the obtaining of the infrared image sample of the air target to be identified and the normalization preprocessing thereon comprises: Obtaining infrared image samples of the air target to be identified, each image sample is centered on the target and has a size of An infrared image slice of the pixel, and the observation direction is the same as that of the high-resolution range image sample; performing amplitude normalization on each infrared image sample according to formula (1) and marking a target type of the infrared image sample, so as to form a normalized infrared image sample library of the air target to be identified; （1） In the formula, Indicates the number is Original infrared image samples; To Image samples after amplitude normalization; The mean of all pixel values for The variance of all pixel values; the sample numbers in this sample library are cataloged according to the similarity of the observation orientation and are consistent with the corresponding high-resolution distance image sample numbers; extracting, from the infrared image sample library of the air target to be identified, a sample with the same sample number as the high-resolution range image training sample as a training sample set, and taking the remaining samples as a test sample set. 5.The intelligent target recognition method of multi-modal feature fusion according to claim 1, characterized in that, the construction and training of the heterogeneous deep learning network for air target feature learning of the high-resolution range image and the infrared image comprises: constructing and training an attention bidirectional gated recurrent unit model for air target feature learning of the high-resolution range image; constructing and training a convolutional neural network model for air target feature learning of the infrared image. 6.The intelligent target recognition method of multi-modal feature fusion according to claim 5, characterized in that, the construction and training of the attention bidirectional gated recurrent unit model for air target feature learning of the high-resolution range image comprises: constructing a bidirectional gated recurrent unit network composed of an input layer, a gated recurrent unit (GRU) forward propagation layer, a GRU backward propagation layer, and a hidden layer output; inputting each high-resolution range image time series sample into the GRU forward propagation layer and the GRU backward propagation layer at the same time; combining the hidden layer outputs of the GRU forward propagation layer and the GRU backward propagation layer to obtain a BGRU hidden layer output; obtaining a weight coefficient representing the importance of a hidden layer node mapped by a different distance unit in the high-resolution range image to target identification through an attention mechanism, and then performing weighted summation on the BGRU hidden layer output by using the weight coefficient to obtain a high-resolution range image time series feature representation; inputting the high-resolution range image time series feature representation vector into a Softmax classifier, and performing fine-tuning and updating on the parameters of the attention bidirectional gated recurrent unit model and the attention weight value coefficient based on the Softmax classification result by using a back propagation method along a time dimension. Through continuous iteration until the model converges, a trained high-resolution range image aerial target feature learning attention bidirectional gated recurrent unit model is obtained, marked as BGRU HRRP , and the network parameters of the model are solidified and stored. 7.The intelligent target recognition method of multi-modal feature fusion according to claim 5, characterized in that, The convolutional neural network model for learning features of infrared image aerial targets is constructed and trained, and specifically includes the following steps. A CNN model composed of three groups of convolutional structures and one group of fully connected structures is constructed, wherein each group of convolutional structures is composed of a convolutional layer, a batch normalization layer and a ReLu activation layer; and the group of fully connected structures is composed of a fully connected layer, a ReLu activation layer and a Dropout layer. Each infrared image training sample is input into the CNN model for training, and the output vector of the fully connected structure is input into a Softmax classifier. Based on the Softmax classification result, the parameters of the CNN model are updated by using a back propagation method. Through continuous iteration until the model converges, a trained infrared image aerial target feature learning convolutional neural network model is obtained, marked as CNN Infra , solidifies and saves the convolutional neural network model parameters. 8.The intelligent target recognition method of multi-modal feature fusion according to claim 7, characterized in that, Further comprising: Target recognition testing, specifically including: Each group of high-resolution range images and infrared image test samples of aerial targets, including a group of high-resolution range image time sequence samples and an infrared image test sample, are input into the trained multi-modal characteristic fusion aerial target recognition model respectively to determine the type of the aerial target, until the type of all test samples is recognized.

Images