Target recognition method, device and equipment based on HRRP spatial feature extraction
By performing HRRP spatial feature extraction and sparse image processing on radar signals, combining multi-scale convolutional networks and multi-layer perceptrons, the poor recognition performance caused by the missing data in radar automatic target recognition is solved, and higher recognition is achieved Accuracy and robustness.
Patent Information
- Application Number
- CN202411171203.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-26
AI Technical Summary
The existing radar automatic target recognition method has poor recognition performance when data is missing, the traditional method has low recognition rate, and deep learning methods rely on data integrity and cannot effectively process missing data in radar signals.
Using the HRRP spatial feature extraction method, the original one-dimensional distance image data is preprocessed and converted into sparse two-dimensional images, and feature extraction and target recognition are used for multi-scale convolutional network and multi-layer perceptron classification network. , build a target recognition network to improve recognition accuracy and robustness.
In the case of data loss, the recognition accuracy and robustness of the radar target recognition algorithm are improved, and the spatiotemporal characteristics between HRRP sequences with missing can be effectively extracted and processed, reducing the impact of data loss on the recognition results.
Smart Images

Figure CN118690254B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of radar signal technology, and in particular to a target recognition method, device and equipment for extracting HRRP spatial features. Background Art
[0002] Radar Auto Target Recognition (RATR) plays an important role in the field of aviation safety. How to quickly, accurately and robustly identify air and space targets is the main task of RATR. In the field of radar signals, one-dimensional high-resolution range profile (HRRP) data is usually obtained through broadband radar. The data contains a lot of separable information about the size, physical configuration and scattering distribution of the target, so HRRP is often used in RATR tasks. For HRRP data, how to fully mine the separable information about the target in HRRP data to improve the performance of the target recognition algorithm is an urgent problem to be solved in the RATR field at this stage.
[0003] Existing methods are mainly divided into two categories: traditional signal processing methods and deep learning methods. Traditional signal processing methods use template matching methods to extract and match statistical and mathematical features in radar signals to achieve target recognition. However, the recognition rate of traditional methods is low. In recent years, RATR methods based on deep learning have gradually emerged. These methods are based on a large number of data samples and extract target features in HRRP data through end-to-end models, ultimately achieving higher recognition performance. However, these methods perform poorly in the case of missing data. Summary of the invention
[0004] Based on this, it is necessary to provide a target recognition method, device and equipment based on HRRP spatial feature extraction that can improve the recognition accuracy and robustness of the recognition algorithm in order to address the above technical problems.
[0005] A target recognition method based on HRRP spatial feature extraction, the method comprising:
[0006] Acquire original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0007] Converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0008] Using a multi-scale convolutional network to extract features from the sparse two-dimensional image, and obtaining multiple feature maps of different scales;
[0009] A classification network based on a multi-layer perceptron is used to perform target recognition based on multiple feature maps of different scales to obtain target recognition results.
[0010] In one embodiment, the preprocessing of the original one-dimensional range image data includes:
[0011] Performing amplitude processing on the original one-dimensional range image data to obtain real one-dimensional range image data;
[0012] The real one-dimensional range image data is normalized by modulo 2 norm normalization to obtain a normalized one-dimensional range image sequence.
[0013] A center of gravity alignment method is used to perform a cyclic shift operation on the normalized one-dimensional range image data to obtain the preprocessed one-dimensional range image data.
[0014] In one embodiment, the real one-dimensional range image data is normalized by modulo 2 norm normalization, which is expressed as:
[0015] ;
[0016] In the above formula, represents the normalized one-dimensional range image sequence, represents real one-dimensional distance image data, Represents the total number of distance cells.
[0017] In one embodiment, a centroid alignment method is used to perform a cyclic shift operation on the normalized one-dimensional range image data, and the centroid of the sequence can be calculated using the following formula:
[0018] ;
[0019] In the above formula, Express Modulo value, represents the total number of distance units, represents the centroid of the normalized one-dimensional range image sequence.
[0020] In one embodiment, the multi-scale convolutional network includes three layers of feature extraction layers connected in sequence, and the three layers of feature extraction layers respectively extract three feature maps of different scales;
[0021] In each of the feature extraction layers, three convolution kernels of different sizes are used to extract features of the input data respectively, and then the extracted features are weighted and fused, and then the output data of the feature extraction layer of this layer is obtained through a maximum pooling layer;
[0022] The input data is the sparse two-dimensional image or the output data of the previous feature extraction layer, and the output data is used as the input data of the next feature extraction layer and the feature map extracted by the feature extraction layer.
[0023] In one embodiment, in the classification network:
[0024] After flattening multiple feature maps of different scales into one-dimensional vectors, they are concatenated to obtain fused features;
[0025] A multi-layer perceptron is used to predict the target category based on the fusion features to obtain the target recognition result.
[0026] In one embodiment, the method further comprises: constructing a target recognition network according to the multi-scale convolutional network and the classification network, inputting the sparse two-dimensional image into the target recognition network, and obtaining the target recognition result;
[0027] When training the target recognition network, a cross entropy loss function is used to update the adjustable parameters in the multi-scale convolutional network and the classification network.
[0028] The present application also provides a target recognition device based on HRRP spatial feature extraction, the device comprising:
[0029] A preprocessing module, used for acquiring original one-dimensional range image data, and preprocessing the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0030] A sparse two-dimensional image acquisition module is used to convert each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splice each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0031] A multi-scale feature extraction module, used to extract features from the sparse two-dimensional image using a multi-scale convolutional network to obtain multiple feature maps of different scales;
[0032] The target recognition module is used to use a classification network based on a multi-layer perceptron to perform target recognition according to multiple feature maps of different scales to obtain target recognition results.
[0033] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0034] Acquire original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0035] Converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0036] Using a multi-scale convolutional network to extract features from the sparse two-dimensional image, and obtaining multiple feature maps of different scales;
[0037] A classification network based on a multi-layer perceptron is used to perform target recognition based on multiple feature maps of different scales to obtain target recognition results.
[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:
[0039] Acquire original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0040] Converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0041] Using a multi-scale convolutional network to extract features from the sparse two-dimensional image, and obtaining multiple feature maps of different scales;
[0042] A classification network based on a multi-layer perceptron is used to perform target recognition based on multiple feature maps of different scales to obtain target recognition results.
[0043] The above-mentioned target recognition method, device and equipment based on HRRP spatial feature extraction preprocesses the original one-dimensional range image data, converts each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, splices each line graph in a sequential arrangement to obtain a sparse two-dimensional image, uses a multi-scale convolutional network to extract features from the sparse two-dimensional image to obtain multiple feature graphs of different scales, and uses a classification network based on a multi-layer perceptron to perform target recognition based on multiple feature graphs of different scales to obtain a target recognition result. The use of this method can improve the recognition accuracy and robustness of the recognition algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 1 is a flow chart of a target recognition method based on HRRP spatial feature extraction in one embodiment;
[0045] Figure 2 In one embodiment, some schematic diagrams of different sequence number input examples are given, wherein: Figure 2 (a) The number of sequences shown is 2×2. Figure 2(b) The number of 3×3 sequences shown, Figure 2 (c) The number of 4×4 sequences shown, Figure 2 (d) The number of 5×5 sequences shown, Figure 2 (e) The number of 6×6 sequences shown, Figure 2 (f) The number of 7×7 sequences displayed;
[0046] Figure 3 A schematic diagram of the overall process of target recognition network training in one embodiment;
[0047] Figure 4 A schematic diagram of mapping a one-dimensional distance dimension into a sparse image in one embodiment to avoid the influence of missing data on the convolution operation;
[0048] Figure 5 FIG. 1 is a schematic diagram of the effect of feature clustering extracted by different methods in an embodiment when the data is missing by 20%, wherein: Figure 5 (a) Schematic diagram showing the effect of the PCA-SVW method. Figure 5 (b) Schematic diagram showing the effect of the RTRBM method. Figure 5 (c) Schematic diagram showing the effect of the SPTC method. Figure 5 (d) Schematic diagram showing the effect of the VGM-RNN method. Figure 5 (e) Schematic diagram showing the effect of the Transformed method. Figure 5 (f) Schematic diagram showing the effect of the TSFA-Net method. Figure 5 (g) Schematic diagram showing the effect of the method;
[0049] Figure 6 It is a structural block diagram of a target recognition device based on HRRP spatial feature extraction in one embodiment;
[0050] Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0051] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application 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 application and are not used to limit the present application.
[0052] In view of the problems in the prior art, both the traditional method and the deep learning method have defects. For example, the traditional method has a low recognition rate, and when the deep learning method is used, its recognition accuracy depends on the completeness of the input data. In the actual background, radar echo data is usually missing, which leads to the unsatisfactory target recognition results of the deep learning method in this case. Figure 1 As shown, a target recognition method based on HRRP spatial feature extraction is provided, comprising the following steps:
[0053] Step S100 , obtaining original one-dimensional range image data, and preprocessing the original one-dimensional range image data to obtain preprocessed one-dimensional range image data.
[0054] Step S110 , converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each line graph in a sequential arrangement to obtain a sparse two-dimensional image.
[0055] Step S120, using a multi-scale convolutional network to extract features from the sparse two-dimensional image to obtain multiple feature maps of different scales.
[0056] Step S130, using a classification network based on a multi-layer perceptron to perform target recognition according to a plurality of feature maps of different scales to obtain a target recognition result.
[0057] In this embodiment, feature mining and extraction are performed on HRRP data with missing data, and the temporal correlation between HRRP sequences is modeled through signal imaging technology. Based on this, a multi-scale convolutional network is designed to extract the spatiotemporal features between HRRP sequences with missing data, so that the model can obtain the ability to resist the interference of missing data. Through the above method, the model can finally obtain well-characterized separable data features, which can improve the recognition accuracy and robustness of the recognition algorithm.
[0058] In step S100 and step S110, multiple one-dimensional HRRP sequences are converted into two-dimensional sparse images and the two-dimensional images are spliced, so that the subsequent feature extraction task can effectively extract the temporal and spatial correlation features of the multiple HRRP sequences and classify them.
[0059] Taking into account the problems of azimuth sensitivity, amplitude sensitivity and translation sensitivity of radar echo data, which will have a great impact on the model's extraction of robust target features, therefore, in step S100, the original radar echo data needs to be normalized and aligned for data preprocessing to reduce the impact of echo data sensitivity.
[0060] In this embodiment, the original one-dimensional range image data is preprocessed, including: performing amplitude processing on the original one-dimensional range image data to obtain real one-dimensional range image data, performing amplitude normalization processing on the real one-dimensional range image data using modulo 2 norm normalization to obtain a normalized one-dimensional range image sequence, and then performing a cyclic shift operation on the normalized one-dimensional range image data using a center of gravity alignment method to obtain preprocessed one-dimensional range image data.
[0061] Specifically, assume that the obtained one-dimensional distance image data is Indicates that Indicates The sub-echo of the range unit, Represents the total number of distance units. Since the data type of the one-dimensional range image data is in complex form, the input one-dimensional range image data is processed by amplitude processing to obtain the real one-dimensional range image data, which is expressed as:
[0062] (1)
[0063] Furthermore, in order to reduce the sensitivity of radar echo data to amplitude, the real one-dimensional range image data is normalized by modulo 2 norm. The amplitude normalization process is expressed as:
[0064] (2)
[0065] Furthermore, in order to reduce the translation sensitivity of the one-dimensional range image data, the center of gravity alignment method is used to Perform a circular shift operation so that its center is located at the center of the range window. Center of gravity It can be calculated by the following formula:
[0066] (3)
[0067] In formula (3), Express The modulus value, that is, the amplitude.
[0068] In step S110, the one-dimensional range image sequence after the centroid correction in step S100 is converted into an image form. At this time, the pre-processed one-dimensional range image data is represented as a collection of multiple HRRP sequences, and each sequence is converted into a corresponding polyline image, which is represented as:
[0069] (4)
[0070] In formula (4), Represents the operation of converting the input data into a line chart without coordinates. Represents the input sequence set, i.e. the preprocessed one-dimensional distance image data, Represents a collection of transformed subgraphs.
[0071] Furthermore, the transformed sub-graphs are arranged in sequence, that is, they are spliced into a graph containing Zhang Zitu's large picture:
[0072] (5)
[0073] In formula (5), Represented by A large picture composed of a sequence of Represents the operation of splicing sub-images into a large image. In this way, HRRP is transformed from a one-dimensional image to a sparse two-dimensional image.
[0074] Furthermore, when arranging in the natural order of the sequence, within a given image size, the sequence may be arranged in the form of 2×2, 3×3, 4×4 or 5×5. Preferably, the sequence is arranged in the form of 5×5.
[0075] The sparse two-dimensional images alleviate the impact of missing data on network feature extraction, and the image representation method enables the use of deep networks for image processing to extract features and classify them. Figure 2 Some examples of different sequence number inputs are given in Figure 2 (a) The number of sequences shown is 2×2. Figure 2 (b) The number of 3×3 sequences shown, Figure 2 (c) The number of 4×4 sequences shown, Figure 2 (d) The number of 5×5 sequences shown, Figure 2 (e) The number of 6×6 sequences shown, Figure 2 (f) The number of 7×7 sequences shown.
[0076] In order to effectively extract the generated HRRP composite image, i.e., the sparse two-dimensional image, in step S120, a multi-scale convolutional network is used to extract its features. In the multi-scale convolutional network, the spatial features of the local drastic changes and the temporal changes of each sequence in the HRRP composite image are extracted through different convolution receptive fields. At the same time, by utilizing the translation invariance property of the convolution and the previous visualization of the one-dimensional distance image data, the convolution operation of the existing sequence will not be affected even in the case of some data missing. Among them, the use of large convolution kernels enables the network to achieve better recognition results for non-ideal data such as existing data when there is a certain degree of distance deviation.
[0077] In this embodiment, the multi-scale convolutional network includes three layers of feature extraction layers connected in sequence, and three feature maps of different scales are extracted by the three feature extraction layers respectively. In each feature extraction layer, three convolution kernels of different sizes are used to extract the features of the input data respectively, and then the extracted features are weighted and fused, and then the output data of the feature extraction layer of this layer is obtained through the maximum pooling layer. Among them, the input data is a sparse two-dimensional image or the output data of the previous feature extraction layer, and the output data is used as the input data of the next feature extraction layer and the feature map extracted by the feature extraction layer of this layer.
[0078] Specifically, the generated HRRP composite image is firstly subjected to feature extraction through the first layer of convolution kernels with kernel sizes of 3, 5, and 7, and the features extracted by each convolution kernel are then weighted fused:
[0079] (6)
[0080] In formula (6), Indicates that the convolution kernel size is The convolution operation, It means that the first layer corresponds to the convolution kernel size of The weight coefficient of It represents the feature map of the first layer fusion output of three different convolution kernels.
[0081] Next, in this layer of feature map Before being input to the next layer, a maximum pooling operation with a kernel size of 2 is performed:
[0082] (7)
[0083] In formula (7), Represents the feature map to be input into the next layer. Represents a max pooling operation with a kernel size of 2.
[0084] Through the subsequent extraction of the two feature extraction layers, three feature maps of different sizes can be obtained, which are expressed as:
[0085] (8)
[0086] In the obtained feature maps, shallow feature maps can effectively capture drastic changes in the sequence, and deep feature maps can effectively reflect the changing trend of the sequence.
[0087] In step S130, the spatiotemporal features extracted by the multi-scale convolutional network are fused and input into the fully connected layer to achieve target classification.
[0088] In this embodiment, in the classification network: after flattening multiple feature maps of different scales into one-dimensional vectors, a splicing operation is performed to obtain fused features, and a multi-layer perceptron is used to predict the target category based on the fused features to obtain a target recognition result.
[0089] Specifically, in the classification network, the extracted feature maps of different scales are first flattened into one-dimensional vectors for subsequent feature concatenation and fusion operations:
[0090] (9)
[0091] In formula (9), Represents the feature map The flattened one-dimensional vector is weighted fused by splicing:
[0092] (10)
[0093] In formula (10), Represents the fusion features finally used for classification kernel calculation loss, represents the weight corresponding to the flattened one-dimensional vector, Represents a feature concatenation operation.
[0094] Next, the fusion features By inputting a multi-layer perceptron to classify the target, it is expressed as:
[0095] (11)
[0096] In formula (11), represents the classification result, and MLP represents the multi-layer perceptron.
[0097] In this embodiment, a target recognition network is constructed based on a multi-scale convolutional network and a classification network, and a sparse two-dimensional image is input into the target recognition network to obtain a target recognition result.
[0098] Furthermore, when training the target recognition network, the cross entropy loss function is used to update the adjustable parameters in the multi-scale convolutional network and the classification network. The cross entropy loss function is expressed as:
[0099] (12)
[0100] In formula (12), represents the cross entropy loss, represents the number of targets, represents the network parameters in the multi-scale convolutional network, Represents the network parameters in the classification network. In the parameter back propagation process, and Will be based on Update parameters, Indicates the corresponding The true value label of .
[0101] In one embodiment, when training the target recognition network, it includes first mapping n×n sequences (i.e., a set of HRRP sequences) to a sparse image of a fixed size of 228×228. There may be missing sequences in the n×n sequences. For the missing sequences, they are directly given in the form of blank images, and finally the sub-images are spliced in order to construct a sparse image. Secondly, the obtained sparse image is input into a three-layer multi-scale convolutional network for feature extraction. In each layer, the feature maps obtained under different convolution kernel operations are weighted fused, and the features between different layers are weighted spliced by flattening and splicing. Finally, the obtained spatiotemporal correlation feature vector is input into the downstream classification head to obtain the classification result and calculate the loss function, and the parameters of the multi-scale convolutional network and the classification head are updated according to the loss function.
[0102] In one embodiment, the process of training the target recognition network is as follows: Figure 3 As shown, Figure 3 (a) is a schematic diagram of the overall framework of the training process. Figure 3 (b) describes in detail how to convert a one-dimensional image data sequence into a two-dimensional coefficient image, which is denoted as the SLM module. Figure 3 The structure of the multi-scale convolutional network is shown in (c) and is denoted as the FEM module.
[0103] In this paper, various experimental results are also given to experimentally prove the effectiveness of the proposed method.
[0104] As shown in the following table, given an image size of 228×228, the network recognition performance under different numbers of input sequences is measured by accuracy, precision, recall, and F1 score. The results show that when the sequence input is 5×5, all indicators are better than other input sequence numbers.
[0105] Table 1 Recognition results of different numbers of input sequences
[0106]
[0107] like Figure 4 As shown in the figure, it shows how this method processes HRRP data by converting the sequence into an image and utilizing the image processing method, and avoids the influence of missing data on the convolution operation results. Through this conversion, the missing data will not affect the existing data convolution.
[0108] As shown in the table below, by comparing the performance with other methods when the data missing rate is 20% to 80%, the results show that this method can perform well in the case of missing data, and performs better than traditional methods and some existing methods in the case of missing data.
[0109] Table 2 Recognition accuracy and parameter quantity of different methods in the case of missing data
[0110]
[0111] like Figure 5 As shown in FIG. 1 , the clustering results of the features extracted by this method and other methods are shown. The results show that the features extracted by this method have better separability.
[0112] As shown in the following table, the recognition performance of this method under non-ideal conditions, that is, different probabilities of range deviation, is demonstrated. This experiment simulates the case where 20% of the data is missing and the existing data has a probability of 10% to 90% range deviation and the deviation distance ranges from 10 range units to 40 range units. From this result, it can be seen that STI-MKCNN can resist the impact of range deviation to a certain extent, and also achieves the best performance compared with the comparison method.
[0113] Table 3 Performance comparison in the case of range offset
[0114]
[0115] In the above-mentioned target recognition method based on HRRP spatial feature extraction, n×n HRRP sequences are first mapped to a sparse image containing these sequences. This operation avoids the influence of missing data between sequences on the convolution operation; then a multi-scale convolution network is constructed based on different convolution kernels to extract features from the generated sparse image. Feature extraction under different receptive fields is achieved through different network depths and convolution kernels of different sizes. The extracted information includes the local drastic changes and global contour information of each subsequence; finally, the features extracted by the multi-scale convolution module are fused and the classification results are obtained. This method processes multiple HRRP sequences to achieve end-to-end target recognition integration, which can effectively extract local and global features of multiple HRRP sequences, thereby improving the recognition and classification accuracy of HRRP sequences, and ensuring a high recognition rate when a certain proportion of data is missing.
[0116] This method can convert HRRP sequence signals into sparse images for processing, avoiding the impact of data missing on other data convolution operations. At the same time, it can obtain features with good characterization from the missing sequences, achieve high recognition accuracy, and maintain high recognition accuracy under the condition of data missing. It can effectively solve the problem of data missing and distance offset in actual recognition tasks.
[0117] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0118] In one embodiment, Figure 6 As shown, a target recognition device based on HRRP spatial feature extraction is provided, comprising: a preprocessing module 200, a sparse two-dimensional image acquisition module 210, a multi-scale feature extraction module 220 and a target recognition module 230, wherein:
[0119] A preprocessing module 200 is used to obtain original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0120] The sparse two-dimensional image obtaining module 210 is used to convert each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splice each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0121] A multi-scale feature extraction module 220, used to extract features from the sparse two-dimensional image using a multi-scale convolutional network to obtain a plurality of feature maps of different scales;
[0122] The target recognition module 230 is used to perform target recognition according to multiple feature maps of different scales using a classification network based on a multi-layer perceptron to obtain a target recognition result.
[0123] For the specific definition of the target recognition device based on HRRP spatial feature extraction, please refer to the definition of the target recognition method based on HRRP spatial feature extraction above, which will not be repeated here. The various modules in the above-mentioned target recognition device based on HRRP spatial feature extraction can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0124] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a target recognition method based on HRRP spatial feature extraction is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.
[0125] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0126] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:
[0127] Acquire original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0128] Converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0129] Using a multi-scale convolutional network to extract features from the sparse two-dimensional image, and obtaining multiple feature maps of different scales;
[0130] A classification network based on a multi-layer perceptron is used to perform target recognition based on multiple feature maps of different scales to obtain target recognition results.
[0131] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:
[0132] Acquire original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data;
[0133] Converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image;
[0134] Using a multi-scale convolutional network to extract features from the sparse two-dimensional image, and obtaining multiple feature maps of different scales;
[0135] A classification network based on a multi-layer perceptron is used to perform target recognition based on multiple feature maps of different scales to obtain target recognition results.
[0136] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0137] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0138] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A target recognition method based on HRRP spatial feature extraction, It is characterized in that The method comprises: Acquire original one-dimensional range image data, and preprocess the original one-dimensional range image data to obtain preprocessed one-dimensional range image data; Converting each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splicing each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image; Using a multi-scale convolutional network to extract features from the sparse two-dimensional image, and obtaining multiple feature maps of different scales; A classification network based on a multi-layer perceptron is used to perform target recognition based on multiple feature maps of different scales to obtain target recognition results.
2. The target recognition method according to claim 1, It is characterized in that The preprocessing of the original one-dimensional range image data comprises: Performing amplitude processing on the original one-dimensional range image data to obtain real one-dimensional range image data; The real one-dimensional range image data is normalized by modulo 2 norm normalization to obtain the normalized one-dimensional range image data; A center of gravity alignment method is used to perform a cyclic shift operation on the normalized one-dimensional range image data to obtain the preprocessed one-dimensional range image data.
3. The target recognition method according to claim 2, It is characterized in that The real one-dimensional range image data is normalized by modulo 2 norm normalization, which is expressed as: In the above formula, represents the normalized one-dimensional range image data, represents real one-dimensional distance image data, Represents the total number of distance cells.
4. The target recognition method according to claim 2, It is characterized in that The centroid alignment method is used to perform a cyclic shift operation on the normalized one-dimensional range image data, wherein the centroid of the one-dimensional range image data is calculated using the following formula: In the above formula, Express Modulo value, represents the total number of distance units, represents the centroid of the normalized one-dimensional range image data.
5. The target recognition method according to any one of claims 1 to 4, It is characterized in that The multi-scale convolutional network includes three layers of feature extraction layers connected in sequence, and the three layers of feature extraction layers respectively extract three feature maps of different scales; In each of the feature extraction layers, three convolution kernels of different sizes are used to extract features of the input data respectively, and then the extracted features are weighted and fused, and then the output data of the feature extraction layer of this layer is obtained through a maximum pooling layer; The input data is the sparse two-dimensional image or the output data of the previous feature extraction layer, and the output data is used as the input data of the next feature extraction layer and the feature map extracted by the feature extraction layer.
6. The target recognition method according to claim 5, It is characterized in that In the multi-layer perceptron-based classification network: After flattening multiple feature maps of different scales into one-dimensional vectors, they are concatenated to obtain fused features; A multi-layer perceptron is used to predict the target category based on the fusion features to obtain the target recognition result.
7. The target recognition method according to claim 6, It is characterized in that The method further includes: constructing a target recognition network according to the multi-scale convolutional network and a classification network based on a multi-layer perceptron, inputting the sparse two-dimensional image into the target recognition network, and obtaining the target recognition result; When training the target recognition network, a cross entropy loss function is used to update the adjustable parameters in the multi-scale convolutional network and the classification network.
8. A target recognition device based on HRRP spatial feature extraction, It is characterized in that The device comprises: A preprocessing module, used for acquiring original one-dimensional range image data, and preprocessing the original one-dimensional range image data to obtain preprocessed one-dimensional range image data; A sparse two-dimensional image acquisition module is used to convert each one-dimensional range image sequence of the preprocessed one-dimensional range image data into a line graph without coordinates, and splice each of the line graphs in a sequential arrangement to obtain a sparse two-dimensional image; A multi-scale feature extraction module, used to extract features from the sparse two-dimensional image using a multi-scale convolutional network to obtain multiple feature maps of different scales; The target recognition module is used to use a classification network based on a multi-layer perceptron to perform target recognition according to multiple feature maps of different scales to obtain target recognition results.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program. It is characterized in that When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.
