Method for training three-dimensional ISAR image recognition network model and electronic device

Abstract

The method comprises the following steps: taking three-dimensional coordinates of point clouds in a three-dimensional ISAR sample image as input, and taking scattering intensities of corresponding points in the point clouds as reference data; training a self-encoder in a deep auto-encoding network, so that the self-encoder has the ability to regress the scattering intensities by using three-dimensional coordinates of point clouds in a three-dimensional ISAR image to be identified; taking output of an encoding module in the self-encoder which has been trained as input of a multi-layer perceptron classifier in the deep auto-encoding network, and taking real categories of point cloud targets in the three-dimensional ISAR sample image as reference; training the multi-layer perceptron classifier, so that the multi-layer perceptron classifier obtains the ability to classify the point cloud targets by using a feature matrix extracted by the encoder; and obtaining a trained three-dimensional ISAR image recognition network model through the encoding module and the multi-layer perceptron classifier.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of radar communication, in particular to a training method of a three-dimensional ISAR image recognition network model and an electronic device. BACKGROUND

[0002] ISAR is an all-weather and all-time active microwave imaging sensor, which can image non-cooperative targets by using the bandwidth of the transmitted microwave and the relative motion between the target and the radar, and is of great significance for automatic target recognition of non-cooperative targets. However, the traditional two-dimensional ISAR image is mainly limited in three aspects in the recognition application: first, the two-dimensional ISAR image can be regarded as the two-dimensional projection of the real three-dimensional target scattering points on the image projection plane (IPP) of the effective angular velocity vector of the target, and for non-cooperative targets whose geometric shape and motion information are unknown, the direction of the IPP is usually unknown, which makes it difficult to interpret the two-dimensional ISAR image and brings great uncertainty to target recognition; second, the Doppler frequency of the two-dimensional single station ISAR image comes from the relative angular velocity of the target and the radar, and if the motion direction of the non-cooperative target coincides with the line connecting the radar and the target, the azimuth angle does not change or changes very little, which will lead to poor imaging effect or even no imaging; finally, the self-occlusion of the target will produce a shadow effect, resulting in the loss of part of the target information in the image. The above three points make it difficult for the two-dimensional single station ISAR image to provide sufficient information to support accurate target recognition in the task of non-cooperative target recognition.

[0003] Three-dimensional ISAR image reconstruction technology provides a solution to solve the limitations of two-dimensional ISAR images in non-cooperative target recognition. Combined with traditional ISAR imaging and interference principles, multiple sensors are used to map the main scattering centers of the target to a three-dimensional space to obtain point cloud data with three coordinates of distance, azimuth angle and elevation angle. Three-dimensional point cloud data contains more sufficient three-dimensional structure information of the target, which can overcome the problem of unknown IPP of non-cooperative targets. In addition, distributed reception in different directions can solve the problem of ISAR forward-looking imaging and to some extent, alleviate the self-occlusion in ISAR images. However, the current three-dimensional ISAR image recognition technology still faces challenges from the system and algorithm: in terms of hardware of the interference ISAR imaging system, the three-dimensional resolution of the three-dimensional ISAR image comes from the large bandwidth in the range direction and the large equivalent aperture caused by the sparse distribution of the antenna array in the angle dimension. For a pure electronic system, the bandwidth of radar signal generation, distribution, reception and processing is severely limited, and the coaxial loss of the connection between the antenna arrays also limits the equivalent aperture. These factors lead to a bottleneck in the resolution of the three-dimensional ISAR image, the local features of the target are blurred, and the difficulty of target recognition is increased; in terms of target recognition software, the current mainstream point cloud target recognition algorithm is based on deep learning technology, and the effectiveness of the algorithm is based on a large amount of training data. At present, there is no complete three-dimensional ISAR image database of non-cooperative targets to support the training of the algorithm. In addition, each point in the three-dimensional ISAR image contains not only three-dimensional coordinate information but also scattering intensity information closely related to the local features of the target. The existing point cloud target recognition network based on deep learning is mostly designed for laser point cloud and only uses three-dimensional coordinate information of the point cloud for recognition. Therefore, how to introduce the scattering intensity information into the decision-making process of target recognition to further improve the accuracy of target recognition is a problem to be solved.

[0004] In summary, three-dimensional ISAR imaging is an effective means of non-cooperative target detection, but the target recognition of three-dimensional ISAR image faces the problems of low resolution, insufficient training database and lack of targeted recognition algorithm. SUMMARY

[0005] The embodiments of the present disclosure provide a training method of a three-dimensional ISAR image recognition network model and an electronic device.

[0006] In a first aspect, a training method of a three-dimensional ISAR image recognition network model is provided in the embodiments of the present disclosure, comprising:

[0007] The three-dimensional coordinates of the point cloud in the three-dimensional ISAR sample image are taken as input, and the scattering intensity of the corresponding points in the point cloud is taken as reference data. The autoencoder in the deep autoencoding network is trained so that the autoencoder has the ability to regress the scattering intensity by using the three-dimensional coordinates of the point cloud in the three-dimensional ISAR image to be recognized.

[0008] The output of the encoding module in the trained autoencoder is taken as the input of a multi-layer perceptron classifier in the deep autoencoding network, and the real category of the point cloud target in the three-dimensional ISAR sample image is taken as a reference to train the multi-layer perceptron classifier, so that the multi-layer perceptron classifier obtains the ability to classify the point cloud target by using the feature matrix extracted by the encoder.

[0009] By combining the encoding module in the autoencoder and the multi-layer perceptron classifier, a trained three-dimensional ISAR image recognition network model is obtained.

[0010] Further, the autoencoder comprises an encoding module and a decoding module.

[0011] The encoding module comprises a point cloud sampling module, a point cloud grouping module and a first feature extraction module.

[0012] The decoding module comprises an inverse interpolation module, a skip connection module and a second feature extraction module; wherein:

[0013] The point cloud sampling module samples a plurality of center points from the input point cloud.

[0014] The point cloud grouping module filters other points within a preset range from the center point from the input point cloud, and takes the center point and the corresponding other points as a point cloud subset.

[0015] The feature extraction is performed on each point cloud subset, and the features of the points in the point cloud subset are compressed into the corresponding features of the points based on the extracted features, as the compressed features of the point cloud subset.

[0016] The inverse interpolation module inversely interpolates the interpolation features corresponding to each point in the point cloud based on the compressed features corresponding to the point cloud subset.

[0017] The skip connection module replicates and splices the original features of each point in the point cloud to the interpolation features corresponding to the corresponding point in the point cloud, to obtain the target features carried by each point in the point cloud.

[0018] The second feature extraction module compresses the target features carried by each point in the point cloud to obtain the final features carried by each point in the point cloud.

[0019] Further, the method further comprises:

[0020] An original three-dimensional ISAR sample image is obtained; the original three-dimensional ISAR sample image comprises three-dimensional coordinates and scattering intensities of each point in the collected point cloud.

[0021] determine the center of gravity coordinates of the point cloud by using the scattering intensity and three-dimensional coordinates of each point in the point cloud;

[0022] align the center of gravity coordinates to the origin of the three-dimensional coordinate system of the three-dimensional ISAR sample image to obtain the three-dimensional ISAR sample image after point cloud registration.

[0023] Further, the method further comprises:

[0024] adopting a local anomaly factor method to remove small-scale noise points on the surface of the point cloud in the three-dimensional ISAR sample image; and / or,

[0025] adopting a statistical filtering method to remove outlier noise points.

[0026] Further, the method further comprises:

[0027] The microwave photon-enabled distributed array ISAR system performs three-dimensional ISAR imaging on targets appearing in a set region to obtain a three-dimensional original image;

[0028] Utilizing target motion synthetic aperture, a high-resolution two-dimensional ISAR image is obtained from the three-dimensional original image;

[0029] According to the real aperture formed by the antenna array perpendicular to the two-dimensional IPP in the distributed array ISAR system, the high-resolution two-dimensional ISAR image obtained for the same target is coherently fused to reconstruct a high-resolution three-dimensional ISAR sample image.

[0030] Further, the microwave photon-enabled distributed array ISAR system comprises a central control module, a transmitting module and a receiving module connected by optical fibers in a centralized manner;

[0031] The central control module comprises a clock, a photonic quadrature waveform generator, a photonic multi-channel de-chirp processor, an analog-to-digital converter and a digital signal processor;

[0032] The photonic quadrature waveform generator comprises a photonic digital-to-analog converter for generating a wideband intermediate frequency signal and a pulse shaper for up-converting the wideband intermediate frequency signal, the pulse shaper outputting a reference signal;

[0033] The photonic multi-channel de-chirp processor comprises a photodetector, a bandpass filter and a Mach-Zehnder modulator, the photodetector converts the reference signal into an electrical signal, which is filtered by the bandpass filter and input to the Mach-Zehnder modulator to obtain a de-chirped reference signal;

[0034] The photonic multi-channel de-chirp processor further comprises a wavelength division multiplexer, an erbium-doped fiber amplifier, a low noise amplifier, an analog-to-digital converter and a digital signal processor; the wavelength division multiplexer combines the echo signals received by the receiving module and inputs the combined echo signals to the Mach-Zehnder modulator for de-chirp processing, the de-chirp processed echo signals are amplified by the erbium-doped fiber amplifier and the low noise amplifier, and are sent to the analog-to-digital converter and the digital signal processor to complete digitization and operation of the signals;

[0035] The transmitting module comprises a photodetector, a band-pass filter, a low noise amplifier and a power amplifier; the optical signal generated by the photonic orthogonal waveform generator is converted into a radio frequency signal by the photodetector, and is transmitted after passing through the band-pass filter, the low noise amplifier and the power amplifier;

[0036] The receiving module comprises a low noise amplifier, a band-pass filter and a Mach-Zehnder modulator biased at a quadrature point; the received echo signal is amplified by the low noise amplifier and filtered by the band-pass filter, and is then sent to the Mach-Zehnder modulator biased at the quadrature point to modulate the optical carrier with a wavelength of λ n .

[0037] In a second aspect, a target recognition method for a three-dimensional ISAR image is provided in the embodiments of the present disclosure, and the method comprises:

[0038] Obtaining a three-dimensional ISAR image to be recognized;

[0039] Inputting the three-dimensional coordinates and scattering intensities of point clouds in the three-dimensional ISAR image to be recognized into a trained three-dimensional ISAR image recognition network model to obtain a recognition result of the point cloud targets in the three-dimensional ISAR image to be recognized; the three-dimensional ISAR image recognition network model is trained by the method of the first aspect.

[0040] In a third aspect, a training device for a three-dimensional ISAR image recognition network model is provided in the embodiments of the present disclosure, and the training device comprises:

[0041] A first training module is configured to input the three-dimensional coordinates of point clouds in a three-dimensional ISAR sample image as input, and input the scattering intensities of corresponding points in the point clouds as reference data, train an autoencoder in a deep autoencoding network, so that the autoencoder has the ability to regress the scattering intensities by using the three-dimensional coordinates of point clouds in a three-dimensional ISAR image to be recognized;

[0042] The second training module is configured to take the output of the encoding module in the trained auto-encoder as the input of a multi-layer perceptron classifier in the deep auto-encoding network, take the true category of the point cloud target in the three-dimensional ISAR sample image as a reference, and train the multi-layer perceptron classifier, so that the multi-layer perceptron classifier obtains the ability to classify the point cloud target by using the feature matrix extracted by the encoder.

[0043] The acquisition module is configured to obtain a trained three-dimensional ISAR image recognition network model by combining the encoding module in the auto-encoder and the multi-layer perceptron classifier.

[0044] The functions can be implemented by hardware, or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the functions described above.

[0045] In one possible design, the apparatus includes a memory and a processor, where the memory is configured to store one or more computer instructions supporting the apparatus to perform the corresponding method described above, and the processor is configured to execute the computer instructions stored in the memory. The apparatus can further include a communication interface configured to enable the apparatus to communicate with other devices or communication networks.

[0046] In a fourth aspect, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory, where the processor executes the computer program to implement the method in any of the aspects described above.

[0047] In a fifth aspect, a computer-readable storage medium is provided, which is configured to store computer instructions for the apparatus in any of the aspects described above, where the computer instructions are executed by a processor to implement the method in any of the aspects described above.

[0048] In a sixth aspect, a computer program product is provided, which includes computer instructions, where the computer instructions are executed by a processor to implement the method in any of the aspects described above.

[0049] The technical solutions provided by the embodiments of the present disclosure can have the following beneficial effects:

[0050] The present disclosure comprises a series of preprocessing steps and a scattering intensity regression deep auto-encoding recognition network designed for three-dimensional ISAR images. A distributed array ISAR empowered by microwave photonics technology obtains three-dimensional ISAR images with high resolution in three coordinate dimensions to construct a database, and a series of preprocessing steps are used to achieve alignment, noise elimination and stem extraction. In order to realize fast and accurate recognition of three-dimensional ISAR images, a deep auto-encoding regression network is used to associate the three-dimensional coordinates and scattering intensity of the point cloud. The trained auto-encoder is used to extract features from the point cloud, and the extracted high-level features containing the three-dimensional position and scattering intensity of the point cloud are used to classify the point cloud, which improves the target recognition accuracy.

[0051] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0052] Other features, objects and advantages of the present disclosure will become more apparent from the following detailed description of the non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings:

[0053] Figure 1 A flow chart showing a training method of a three-dimensional ISAR image recognition network model according to an embodiment of the present disclosure is shown;

[0054] Figure 2 A technical flowchart showing the implementation of a deep auto-encoding network based on scattering intensity regression to recognize three-dimensional ISAR images of non-cooperative aircraft targets according to an embodiment of the present disclosure is shown; Figure 3 A schematic diagram of a microwave photonically empowered distributed array ISAR system for three-dimensional ISAR imaging of non-cooperative targets according to an embodiment of the present disclosure is shown;

[0055] Figures 4(a)-4(d) An example diagram of a three-dimensional ISAR image database according to an embodiment of the present disclosure is shown;

[0056] Figures 5(a)-5(d) A schematic diagram showing the results after each preprocessing step according to an embodiment of the present disclosure is shown;

[0057] Figure 6 An example diagram of a scattering intensity regression deep auto-encoding network structure based on scattering intensity regression according to an embodiment of the present disclosure is shown;

[0058] Figure 7 A schematic diagram of an electronic device suitable for implementing a training method of a three-dimensional ISAR image recognition network model and / or a target recognition method of a three-dimensional ISAR image according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0059] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so as to be easily carried out by one of ordinary skill in the art. Also, portions unrelated to describing the exemplary embodiments are omitted in the drawings for the sake of clarity.

[0060] In the present disclosure, it should be understood that terms such as "include" or "have" or the like are intended to indicate that there are features, numbers, steps, actions, components, parts or combinations thereof disclosed in the specification, and do not exclude the presence or addition of one or more other features, numbers, steps, actions, components, parts or combinations thereof.

[0061] It should also be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0062] In order to use the deep learning point cloud target recognition method to quickly and accurately recognize the three-dimensional ISAR image, the present disclosure proposes a point cloud target recognition method based on scattering intensity regression deep auto-encoding network. The present disclosure includes a series of preprocessing steps and scattering intensity regression deep auto-encoding recognition network designed for three-dimensional ISAR images. The distributed array ISAR enabled by microwave photonics technology obtains three-dimensional ISAR images with high resolution in three coordinate dimensions to construct a database, and a series of preprocessing steps are used to achieve alignment, eliminate noise and extract the main part. In order to achieve fast and accurate recognition of three-dimensional ISAR images, a deep auto-encoding regression network is used to associate the three-dimensional coordinates and scattering intensity of the point cloud. The trained auto-encoder is used to extract features from the point cloud, and the extracted high-level features containing the three-dimensional position and scattering intensity of the point cloud are used to classify the point cloud.

[0063] The details of the embodiments of the present disclosure will be described in detail below through specific embodiments.

[0064] Figure 1 A flowchart showing a training method of a three-dimensional ISAR image recognition network model according to an embodiment of the present disclosure is shown.

[0065] As shown in Figure 1 The training method of the image recognition network model includes the following steps:

[0066] In step S101, the three-dimensional coordinates of the point cloud in the three-dimensional ISAR sample image are taken as input, and the scattering intensity of the corresponding points in the point cloud is taken as reference data. The auto-encoder in the deep auto-encoding network is trained so that the auto-encoder has the ability to regress the scattering intensity using the three-dimensional coordinates of the point cloud in the three-dimensional ISAR image to be recognized.

[0067] In step S102, the output of the encoding module in the trained autoencoder is taken as the input of a multilayer perceptron classifier in the deep autoencoding network, and the real category of the point cloud target in the three-dimensional ISAR sample image is taken as a reference to train the multilayer perceptron classifier, so that the multilayer perceptron classifier obtains the ability to classify the point cloud target by using the feature matrix extracted by the encoder.

[0068] In step S103, a trained three-dimensional ISAR image recognition network model is obtained by combining the encoding module in the autoencoder and the multilayer perceptron classifier.

[0069] The following will introduce an implementation manner of the present disclosure in detail through specific details.

[0070] In an optional implementation manner of the present embodiment, an original three-dimensional ISAR sample image is acquired; the original three-dimensional ISAR sample image includes three-dimensional coordinates and scattering intensity of each point in the collected point cloud.

[0071] The center of gravity coordinates of the point cloud are determined by using the scattering intensity and three-dimensional coordinates of each point in the point cloud.

[0072] The center of gravity coordinates are aligned to the origin of the three-dimensional coordinate system of the three-dimensional ISAR sample image, and the three-dimensional ISAR sample image after point cloud registration is obtained.

[0073] In an optional implementation manner of the present embodiment, the method further includes:

[0074] The local outlier factor method is adopted to remove small-scale noise points on the surface of the point cloud in the three-dimensional ISAR sample image; and / or,

[0075] The statistical filtering method is adopted to remove outlier noise points.

[0076] In the optional implementation manner, the point cloud preprocessing step is designed according to the characteristics of the three-dimensional ISAR point cloud: the acquired three-dimensional ISAR image data is an N*4 matrix, where N represents the number of points in the point cloud data, each point contains coordinate information of three dimensions of range, azimuth and elevation and scattering intensity information, and a total of 4-dimensional data. The point cloud matrix has disorder, that is, the order of points in the matrix does not affect the expression of the point cloud data.

[0077] The preprocessing of the point cloud data mainly includes two parts of point cloud registration and point cloud denoising:

[0078] 1、Point cloud registration, since all three-dimensional ISAR images in the database come from the same section on the same route, the direction of the aircraft target remains consistent here, so the attitude of each aircraft target in the three-dimensional ISAR point cloud obtained under the same imaging condition also remains consistent. When the point cloud is registered, the slight difference in attitude can be ignored, and the position can be directly aligned. The three-dimensional ISAR image is registered by using the scattering intensity weighted barycenter alignment method. The barycenter of the point cloud target is determined by using the scattering intensity and position coordinates of the point cloud, that is, the barycenter point is obtained by weighted average of all point coordinates, and then the barycenter coordinates are aligned to the origin:

[0079]

[0080] Wherein, x n , y n , z n and I n are the distance coordinates, Doppler angle coordinates, pitch angle coordinates and scattering intensity coordinates of each point. x G , y G and z G are the distance coordinates, Doppler angle coordinates and pitch angle coordinates of the barycenter of the point cloud. The barycenter coordinates aligned to the origin can be understood as translating the entire point cloud target in the three-dimensional coordinate system, so that the barycenter coordinates fall on the origin (0, 0, 0) of the three-dimensional coordinate system. That is, the coordinates of all point clouds are added to a vector from the barycenter point to the origin; the origin is still the origin of the three-dimensional coordinate system, but the entire point cloud target is translated to the vicinity of the origin. This is called point cloud registration, which moves the point cloud to the same spatial position.

[0081] In the model training process, the original three-dimensional ISAR sample image obtained can be obtained by the above-mentioned point cloud registration method to obtain a three-dimensional ISAR sample image for model training.

[0082] In the preprocessing step, the three-dimensional ISAR sample image after point cloud registration can also be subjected to the following point cloud denoising processing. The three-dimensional ISAR sample image after point cloud denoising can be used as input data in the model training process. It can be understood that after the model training is completed, the three-dimensional ISAR to-be-identified image which needs to be identified can also be subjected to the above-mentioned preprocessing step, that is, the point cloud registration and the point cloud denoising, and then input into the trained image recognition network model for target identification.

[0083] 2、Point cloud denoising, there are many interference noises in three-dimensional ISAR image, mainly divided into point cloud surface small scale noise and outlier point cloud cluster, using local outlier factor (Local Outlier Factor, LOF) and statistical filtering to remove respectively. LOF compares the density of any point p and its neighborhood points to determine whether the point is a noise point, the lower the density of point p, the more likely it is a noise, the high and low of p density is mainly judged by local outlier factor. For any point p, its local outlier factor can be expressed as:

[0084]

[0085] Where, N k (p) represents the kth distance neighborhood of point p. N k (p) is defined as in the set for any two points, point p and point o, if it satisfies in the area range with p point and o point distance d(p,o) as the radius, at least there are k points o' not including p point in the range, satisfy d(p,o')≤d(p,o), and at most there are k-1 points not including point o in the range, satisfy d(p,o')<d(p,o), then the distance d(p,o) between point p and point o is called the kth distance d k (p) of point p, N k (p) is the set of all points within the kth distance of point p. lrd k (p) represents the local reachable density of point p, the expression is:

[0086]

[0087] reach_dist k (p,o) is the kth reachable distance of point o to point p, the expression is:

[0088] reach_dist k (p,o)=max{d k (o),d(p,o)} (4)

[0089] Statistical filtering judges the abnormal points by the average distance between each point and its neighborhood points in point cloud data. First, assume that the average distance between each point and its k neighborhood points in point cloud data is normally distributed, and the distribution is determined by the mean value μ and the standard deviation σ in normal distribution. If the actual distance of a point o in point cloud data does not conform to this normal distribution rule, that is, if the average distance between a point and its neighborhood is outside the distance threshold d=μ±σ×std, where std is a hyperparameter for adjusting the threshold range, the smaller std is, the better the denoising effect is, then the point is considered as a noise point and is removed from the whole point cloud data set.

[0090] In an optional implementation of the embodiment, the autoencoder comprises an encoding module and a decoding module.

[0091] The encoding module comprises a point cloud sampling module, a point cloud grouping module and a first feature extraction module.

[0092] The decoding module comprises an inverse interpolation module, a skip connection module and a second feature extraction module.

[0093] The point cloud sampling module samples a plurality of center points from the input point cloud.

[0094] The point cloud grouping module filters other points in the input point cloud that are within a preset range from the center points, and takes the center points and the corresponding other points as a point cloud subset.

[0095] The feature extraction is performed on each point cloud subset, and the features of the points in the point cloud subset are compressed into corresponding features of the points as compressed features of the point cloud subset based on the extracted features.

[0096] The inverse interpolation module inversely interpolates the interpolation features corresponding to each point in the point cloud based on the compressed features corresponding to the point cloud subset.

[0097] The skip connection module copies and splices the original features of each point in the point cloud to the interpolation features corresponding to the corresponding points in the point cloud to obtain target features carried by the points in the point cloud.

[0098] The second feature extraction module compresses the target features carried by the points in the point cloud to obtain final features carried by the points in the point cloud.

[0099] The structures of the autoencoder and the classifier and the training details are introduced as follows:

[0100] A point cloud target recognition network based on a scattering intensity regression depth autoencoding network is constructed, and the framework comprises two main modules.

[0101] 1. An autoencoder based on scattering intensity regression, the encoding and decoding modules in the autoencoder adopt the point cloud data compression and restoration process in the classical point cloud segmentation network Pointnet++.

[0102] (1) The encoding module comprises three steps: a point cloud sampling module, a point cloud grouping module and a Pointnet feature extraction module.

[0103] The point cloud sampling module uses a farthest point sampling method (FPS) to select N' center points from N points of an input point cloud NxC by uniformly sampling, to obtain N'xC, where C is the number of feature channels of each point in the encoding process;

[0104] The point cloud grouping module takes the N' center points obtained in the point cloud sampling as the center, finds k points closest to the center points in a circular region with a preset length as the radius, to obtain N' point cloud subsets (that is, one center point corresponds to one point cloud subset), each point cloud subset contains k points, that is, N' (kxC);

[0105] Finally, the Pointnet feature extraction is performed on each point cloud subset, the traditional Pointnet architecture is adopted, the features of each point in the point cloud subset are convolved to achieve the purpose of expanding the channel to obtain the features, and finally the maximum value of the feature of each channel in the point cloud subset is taken, each point cloud subset kxC is compressed into a point 1xC', the coordinates of the point are the coordinates of the center point of the point cloud subset, and the feature of the point is the Pointnet operation result of the features of all points in the range of the center point of the point cloud subset, that is, the feature of the point is the compression result of the features of all points in the range of the center point of the point cloud subset. Finally, a feature matrix of N'xC' is obtained, where C' is the number of feature channels carried by the center point of the point cloud subset, so as to achieve the purpose of point cloud sampling and feature extraction.

[0106] (2) The decoding module includes an inverse interpolation module, a jump connection module and a feature extraction module, the inverse interpolation module uses the three-dimensional coordinates NxB corresponding to the center points NxB obtained by point cloud sampling and the three-dimensional coordinates N'x3 of the point cloud before sampling (N'>N), for each point p before sampling, find the k points closest to it in the point cloud after sampling, and use the features carried by the k points to predict the feature of p, the expression is:

[0107]

[0108] Where B is the number of feature channels of each point in the decoding process, x i ,...,x k are the k points closest to p in the point cloud subset after sampling, f i (j) is the jth-dimensional feature of the ith point, f (j) (p) is the feature of point p, d(p,x i ) 2 is the Euclidean distance between point p and point x i , ω i (p) is the weight coefficient of the feature of point x i , xi ..., x k The farther the point is from the center, the smaller the feature contribution of the point p is, and finally, the NxB feature matrix carried by the center point obtained by compression using the encoding module is used for reverse interpolation to obtain the NxB feature matrix carried by the pre-sampling point cloud.

[0109] The skip connection module learns from the idea of the classic U-Net image segmentation network, and copies and splices the feature NxC' corresponding to the sampling layer in the point cloud data compression to the point cloud feature NxB after reverse interpolation, to obtain an N'(C'+B) feature matrix.

[0110] The feature extraction uses one-dimensional convolution to compress the feature carried by each point, and compresses the N'(C'+B) feature matrix into NxB'. After compression and restoration of the point cloud data, the feature carried by each point is finally output.

[0111] The three-dimensional coordinates of the point cloud in the three-dimensional ISAR image are used as the input of the autoencoder, and the corresponding scattering intensity of the point cloud is used as the reference data output by the encoder, and the autoencoder is trained. The trained autoencoder has the ability to regress the scattering intensity of the ISAR point cloud three-dimensional coordinates;

[0112] 2. The classifier using the features extracted by the autoencoder for classification, the encoding module of the trained autoencoder outputs the compressed point cloud feature, each three-dimensional ISAR image is compressed into a 1x1024 feature matrix, and this matrix is a high-level feature obtained by compressing the three-dimensional position and scattering intensity information of all scattering points in the point cloud. The feature matrix is used as the input, and the type of the point cloud target (such as non-cooperative) is used as the output reference data, and the multi-layer perceptron classifier is trained. The trained multi-layer perceptron classifier has the ability to judge the type of the three-dimensional ISAR image using the feature matrix output by the encoding module. The final point cloud target recognition network is composed of the encoding module based on the autoencoder and the multi-layer perceptron classifier, and the input and output are the three-dimensional coordinates of the three-dimensional ISAR image point cloud and the target type.

[0113] In an optional implementation of the embodiment, the method further includes:

[0114] The microwave photon-enabled distributed array ISAR system performs three-dimensional ISAR imaging on the target appearing in the set region to obtain a three-dimensional original image;

[0115] A high-resolution two-dimensional ISAR image is obtained from the three-dimensional original image by using target motion synthetic aperture;

[0116] According to a real aperture formed by an antenna array perpendicular to a two-dimensional IPP in the distributed array ISAR system, the high-resolution two-dimensional ISAR images obtained for the same target are coherently fused to reconstruct a high-resolution three-dimensional ISAR sample image.

[0117] In an optional implementation of the embodiment, the microwave photon-enabled distributed array ISAR system comprises a central control module, a transmitting module and a receiving module connected through optical fibers in a centralized manner.

[0118] The central control module comprises a clock, a photon quadrature waveform generator, a photon multi-channel de-chirp processor, an analog-to-digital converter and a digital signal processor.

[0119] The photon quadrature waveform generator comprises a photon digital-to-analog converter for generating a wideband intermediate frequency signal and a pulse shaper for up-converting the wideband intermediate frequency signal, wherein the pulse shaper outputs a reference signal.

[0120] The photon multi-channel de-chirp processor comprises a photoelectric detector, a band-pass filter and a Mach-Zehnder modulator, wherein the photoelectric detector converts the reference signal into an electrical signal, the electrical signal is filtered by the band-pass filter and then input into the Mach-Zehnder modulator to obtain a de-chirped reference signal.

[0121] The photon multi-channel de-chirp processor further comprises a wavelength division multiplexer, an erbium-doped fiber amplifier, a low-noise amplifier, an analog-to-digital converter and a digital signal processor, wherein the wavelength division multiplexer combines echo signals received by the receiving module and then inputs the combined echo signals into the Mach-Zehnder modulator for de-chirp processing, the de-chirped echo signals are amplified by the erbium-doped fiber amplifier and the low-noise amplifier, and then are sent to the analog-to-digital converter and the digital signal processor for signal digitization and operation.

[0122] The transmitting module comprises a photoelectric detector, a band-pass filter, a low-noise amplifier and a power amplifier, wherein the optical signal generated by the photon quadrature waveform generator is converted into a radio frequency signal by the photoelectric detector, and then is transmitted after passing through the band-pass filter, the low-noise amplifier and the power amplifier.

[0123] The receiving module comprises a low-noise amplifier, a band-pass filter and a Mach-Zehnder modulator biased at a quadrature point, wherein the received echo signal is amplified by the low-noise amplifier and filtered by the band-pass filter, and then is input into the Mach-Zehnder modulator biased at the quadrature point to modulate an optical carrier with a wavelength of λ n

[0124] ​The acquisition process of the three-dimensional ISAR sample image and the imaging system of the three-dimensional ISAR sample image are the same as the acquisition process and the imaging system of the three-dimensional ISAR to-be-identified image, and specific details are introduced in the related examples of the target identification method below.

[0125] According to an embodiment of the target identification method of the three-dimensional ISAR image, the target identification method comprises the following steps: acquiring a three-dimensional ISAR to-be-identified image; inputting the three-dimensional coordinates and the scattering intensity of the point cloud in the three-dimensional ISAR to-be-identified image into a trained three-dimensional ISAR image identification network model to obtain the identification result of the point cloud target in the three-dimensional ISAR to-be-identified image; and the three-dimensional ISAR image identification network model is obtained by training the image identification network model.

[0126] The target identification method of the three-dimensional ISAR image according to the present disclosure is described in detail below through specific examples. The target identification method uses the deep auto-encoding network model based on scattering intensity regression trained above to identify the non-cooperative aircraft target three-dimensional ISAR image.

[0127] Figure 2 A flowchart of the target identification method of the non-cooperative aircraft target three-dimensional ISAR image using the deep auto-encoding network based on scattering intensity regression according to an embodiment of the present disclosure is shown. As shown in FIG. 2, the target identification method comprises the following steps: Figure 2

[0128] First, a microwave photon-enabled distributed array ISAR system is built, Figure 3 A schematic diagram of a microwave photon-enabled distributed array ISAR system for three-dimensional ISAR imaging of non-cooperative targets according to an embodiment of the present disclosure is shown. As shown in FIG. 3, the distributed array ISAR system comprises a center control module (CO), a plurality of transmitter modules (TX) and a plurality of receiver modules (RX), and the TX, RX and CO are connected through optical fibers in a centralized manner:

[0129] ​1、Central control module contains clock, photonic quadrature waveform generator, photonic multi-channel dechirp processor, analog-to-digital converter and digital signal processor. The clock provides a time reference of one pulse per second (1PPS) and a frequency reference of 10MHz for multi-channel radar signal generation and dechirp data acquisition; the photonic quadrature waveform generator contains a photonic digital-to-analog converter (PDAC) and a pulse shaper, which are used for broadband intermediate frequency signal generation and up-conversion, respectively, to obtain a reference signal; the photonic multi-channel dechirp processor, on the one hand, transmits the reference signal generated by the photonic quadrature waveform generator, which is converted into an electrical signal by a photodetector (PD), filtered by a bandpass filter (BPF), and finally input into a Mach-Zehnder modulator (MZM) as a dechirp reference signal, on the other hand, combines the echoes received by RX through a wavelength division multiplexer (WDM), and then inputs into the MZM for dechirp processing. The dechirped signal is amplified by an erbium-doped fiber amplifier (EDFA) and an LNA, and then sent to an analog-to-digital converter and a digital signal processor for digitization and operation of the signal.

[0130] 2、Remote transceivers (TX and RX) are only responsible for signal conversion, amplification and radiation or collection between photonics and electronics. In TX, the optical signal generated by the photonic quadrature waveform generator is converted into a radio frequency (RF) signal by a PD, then passed through a BPF, a low noise amplifier (LNA) and a power amplifier (PA) for emission; in RX, the received echo signal is amplified by a suitable LNA and filtered by a BPF, and then sent to a MZM biased at the quadrature point to modulate the optical carrier with a wavelength of λ n . In this way, the echo conversion between electronics and photonics is completed. The radiation and collection of RX and TX are both completed by horn antennas. The final transceiver array radar contains 4 TX and 4 RX.

[0131] Then, a microwave photon-enabled distributed array ISAR is used to create a three-dimensional ISAR image database of aircraft targets, and the training set and test set are divided:

[0132] A microwave photon-enabled distributed array ISAR is used to monitor a fixed passenger aircraft route, and a three-dimensional ISAR image of a non-cooperative passenger aircraft target is obtained.

[0133] The distributed array ISAR includes multiple receiving channels (e.g., receiving antennas), a single receiving channel utilizes target motion synthetic aperture to obtain a high-resolution two-dimensional ISAR image, and all two-dimensional ISAR images obtained by the receiving channels are coherently fused according to a real aperture formed by an antenna array perpendicular to the two-dimensional IPP to perform accurate three-dimensional reconstruction and obtain a high-resolution three-dimensional ISAR image.

[0134] The flight information of the route is queried and is corresponded with the radar data acquisition time to obtain a real passenger aircraft target type corresponding to each three-dimensional ISAR image, and the three-dimensional ISAR image is labeled according to the real passenger aircraft target type. Finally, a three-dimensional ISAR image database containing 4 passenger aircraft types and a total of 516 samples is formed, and the three-dimensional ISAR image database is divided into a training set and a test set according to a ratio of 4:1, and instances in the database are shown in FIG. 4, wherein four passenger aircraft types are shown; and the darker the color of the point, the greater the scattering intensity. Figures 4(a)-4(d)

[0135] The three-dimensional ISAR image is preprocessed by using the designed preprocessing step:

[0136] 1. Point cloud target registration, a gravity center alignment method is adopted, the gravity center of the point cloud is calculated according to the scattering intensity weighting, and the gravity centers of the point cloud targets (one sample image can contain one point cloud target) in all sample images are aligned to the origin;

[0137] 2. Point cloud target small-scale noise removal, a LOF method is used to calculate the local outlier factor of all points, and the threshold is adjusted to remove the scattered noise on the surface of the point cloud target while retaining the main body of the point cloud target;

[0138] 3. Point cloud target outlier cluster removal, a statistical filtering method is used to divide the point cloud clusters according to the statistical distribution, and remove the point cloud clusters far away from the main trunk of the aircraft;

[0139] 4. Normalization processing, the three coordinate dimensions and the scattering intensity of the point cloud target are normalized to avoid the influence of data value difference on recognition. An example of the preprocessed three-dimensional ISAR image is shown in FIG. 5. FIG. 5(a) shows the gravity center alignment result, FIG. 5(b) shows the LOF filtering result, FIG. 5(c) shows the statistical filtering result, and FIG. 5(d) shows the normalization result. Figures 5(a)-5(d)

[0140] A point cloud target recognition network based on a scattering intensity regression deep auto-encoding network is built, and the network structure is shown in FIG. 6. Figure 6

[0141] The training set is used to train the recognition network:

[0142] ​​​Firstly, the three-dimensional coordinates of the three-dimensional ISAR image are taken as input, the scattering intensity of the corresponding point is taken as reference, the deep auto-encoding network is trained to obtain the ability of using three-dimensional coordinates to regress scattering intensity.

[0143] Then, the output of the encoding module in the trained auto-encoder is taken as input of the multi-layer perceptron classifier, the category of the point cloud target is taken as reference, the multi-layer perceptron classifier is trained to obtain the ability of using the feature matrix extracted by the encoder to classify the point cloud target.

[0144] Finally, the encoding module in the auto-encoder and the multi-layer perceptron classifier are combined to obtain the three-dimensional ISAR point cloud recognition network.

[0145] The training strategy adopted in the training is as follows:

[0146] In order to avoid the network falling into local minimum value in the training, the Adam optimizer with momentum is adopted in the training of the two steps, and the learning rate is exponentially decayed periodically to help the network quickly reach a stable convergence state; in order to obtain a self-encoding network with better generalization ability, the self-encoding network is tested in real time by using the test data set, the performance is monitored, and the training is stopped before the network falls into overfitting, in addition, L2 regularization is also adopted to avoid the self-encoding network falling into overfitting; in order to enhance the stability and convergence speed of the multi-layer perceptron classifier training, the Mini-batch with a size of 32 is adopted, and batch normalization is performed between the outputs of each layer.

[0147] Finally, the trained point cloud recognition network is used to identify the target type in the test set data, and the average accuracy of 95.97% is obtained by replacing five random training set test set groupings. Compared with the accuracy of 77.68% and 83.04% respectively obtained by the classical methods pointne and pointnet++ in the three-dimensional ISAR image database, there is a significant improvement. The experimental results prove the effectiveness of the present disclosure in the three-dimensional ISAR image target recognition task.

[0148] (1) The present disclosure proposes a three-dimensional ISAR image target recognition method based on scattering intensity regression deep auto-encoding network. The three-dimensional coordinates and scattering intensity of the three-dimensional ISAR image are associated by using the deep auto-encoding network, the completed auto-encoder is used for feature extraction of the point cloud, and the high-level features containing the three-dimensional position and scattering intensity of the point cloud are used for classification of the point cloud, which improves the recognition accuracy compared with the traditional point cloud recognition method, and provides a new idea for processing various three-dimensional microwave point cloud data with scattering intensity information.

[0149] (2) The embodiment of the present disclosure first utilizes the microwave photonically enabled distributed array ISAR to create a high-resolution three-dimensional ISAR image database of an aircraft target, opens the door for the application of a deep learning method to three-dimensional ISAR image processing, and paves the way for further exploration of new applications of three-dimensional ISAR images.

[0150] The following is an apparatus embodiment of the present disclosure, which can be used to perform the method embodiments of the present disclosure.

[0151] The training device of the three-dimensional ISAR image recognition network model according to an embodiment of the present disclosure can be realized as part or all of an electronic device through software, hardware, or a combination of both. The training device of the three-dimensional ISAR image recognition network model includes:

[0152] The first training module is configured to take the three-dimensional coordinates of the point cloud in the three-dimensional ISAR sample image as input, and take the scattering intensity of the corresponding points in the point cloud as reference data, train the autoencoder in the deep autoencoding network, so that the autoencoder has the ability to regress the scattering intensity using the three-dimensional coordinates of the point cloud in the three-dimensional ISAR image to be recognized;

[0153] The second training module is configured to take the output of the encoding module in the trained autoencoder as the input of the multilayer perceptron classifier in the deep autoencoding network, and take the true category of the point cloud target in the three-dimensional ISAR sample image as a reference, train the multilayer perceptron classifier, so that the multilayer perceptron classifier obtains the ability to classify the point cloud target using the feature matrix extracted by the encoder.

[0154] The acquisition module is configured to obtain the trained three-dimensional ISAR image recognition network model by combining the encoding module in the autoencoder and the multilayer perceptron classifier.

[0155] The training device of the three-dimensional ISAR image recognition network model in the embodiment corresponds to the training method of the three-dimensional ISAR image recognition network model described above, and specific details can be referred to the description of the training method of the three-dimensional ISAR image recognition network model above, which will not be repeated here.

[0156] Figure 7 is a structural schematic diagram of an electronic device suitable for implementing the training method of the three-dimensional ISAR image recognition network model and / or the target recognition method of the three-dimensional ISAR image according to an embodiment of the present disclosure.

[0157] As Figure 7As shown, the electronic device 700 includes a processing unit 701, which can be implemented as a CPU, a GPU, a FPGA, a NPU, or the like. The processing unit 701 can perform various processes in the embodiments of any of the methods of the present disclosure according to a program stored in a read only memory (ROM) 702 or a program loaded into a random access memory (RAM) 703 from the storage section 708. Various programs and data required for the operation of the electronic device 700 are also stored in the RAM 703. The processing unit 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0158] The following components are connected to the I / O interface 705: an input section 706 including a keyboard, a mouse, etc.; an output section 707 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, a modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as necessary. A removable recording medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 710 as necessary, so that a computer program read therefrom is installed into the storage section 708 as necessary.

[0159] In particular, in accordance with embodiments of the present disclosure, the above with reference to any of the methods of the embodiments of the present disclosure can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program tangibly embodied on a machine readable medium, the computer program comprising program code for executing any of the methods of the embodiments of the present disclosure. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 709, and / or installed from the removable recording medium 711.

[0160] The flow and block diagrams in the drawings represent possible architectural, functional, and operational architectures of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block can represent a module, a segment, or a portion of code that comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0161] The units or modules described in the embodiments of the present disclosure can be implemented by software, or by hardware. The described units or modules can also be arranged in a processor, and the name of the unit or module does not constitute a limitation on the unit or module itself in some cases.

[0162] As another aspect, the present disclosure also provides a computer readable storage medium, which can be the computer readable storage medium included in the apparatus described in the above embodiments, or can exist separately from the apparatus and not be assembled into the apparatus. The computer readable storage medium stores one or more programs for execution by one or more processors to perform the methods described in the present disclosure.

[0163] The above description is merely preferred embodiments of the present disclosure and a description of principles of applied technologies. It should be understood by those skilled in the art that the scope of the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and also includes other technical solutions formed by the combinations of the above technical features or equivalent features without departing from the inventive concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features disclosed in the present disclosure (but not limited to) having similar functions.

Claims

1. A training method for a three-dimensional ISAR image recognition network model, characterized in that, include: The three-dimensional coordinates of the point cloud in the three-dimensional ISAR sample image are used as input, and the scattering intensity of the corresponding point in the point cloud is used as reference data to train the autoencoder in the deep autoencoder network, so that the autoencoder has the ability to regress the scattering intensity using the three-dimensional coordinates of the point cloud in the three-dimensional ISAR image to be identified. The output of the encoding module in the pre-trained autoencoder is used as the input of the multilayer perceptron classifier in the deep autoencoder network. The true types of point cloud targets in the 3D ISAR sample image are used as a reference to train the multilayer perceptron classifier, so that the multilayer perceptron classifier can acquire the ability to classify the point cloud targets using the feature matrix extracted by the encoder. By combining the encoding module in the autoencoder and the multilayer perceptron classifier, a trained 3D ISAR image recognition network model is obtained. The autoencoder includes an encoding module and a decoding module; The encoding module includes a point cloud sampling module, a point cloud grouping module, and a first feature extraction module; The decoding module includes an inverse interpolation module, a skip connection module, and a second feature extraction module; wherein: The point cloud sampling module samples multiple center points from the input point cloud; The point cloud grouping module filters out other points from the input point cloud whose distance from the center point is within a preset range, and groups the center point and the corresponding other points as a subset of the point cloud; Feature extraction is performed on each of the point cloud subsets, and based on the extracted features, the features of each point in the point cloud subset are compressed into the features corresponding to a single point, which are used as the compressed features of the point cloud subset. The reverse interpolation module obtains the interpolation feature corresponding to each point in the point cloud by reverse interpolation based on the compressed feature corresponding to the subset of the point cloud. The jump connection module copies and splices the original features of each point in the point cloud to the interpolated features corresponding to the corresponding points in the point cloud, thereby obtaining the target features carried by each point in the point cloud; The second feature extraction module compresses the target features carried by each point in the point cloud to obtain the final features carried by each point in the point cloud.

2. The method according to claim 1, characterized in that, The method further includes: Acquire the original three-dimensional ISAR sample image; the original three-dimensional ISAR sample image includes the three-dimensional coordinates and scattering intensity of each point in the acquired point cloud; The centroid coordinates of the point cloud are determined using the scattering intensity and three-dimensional coordinates of each point in the point cloud. Align the centroid coordinates to the origin of the three-dimensional coordinate system of the three-dimensional ISAR sample image to obtain the three-dimensional ISAR sample image after point cloud registration.

3. The method according to any one of claims 1-2, characterized in that, The method further includes: The local anomaly factor method is used to remove small-scale noise points from the point cloud surface in the 3D ISAR sample image; and / or, Outlier noise points are removed using statistical filtering.

4. The method according to any one of claims 1-2, characterized in that, The method further includes: A microwave photonics-enabled distributed array ISAR system performs three-dimensional ISAR imaging on targets appearing in a designated area to obtain a raw three-dimensional image. High-resolution two-dimensional ISAR images are obtained from the original three-dimensional image using target motion synthetic aperture; Based on the actual aperture formed by the antenna array perpendicular to the two-dimensional IPP in the distributed array ISAR system, the high-resolution two-dimensional ISAR images obtained for the same target are coherently fused to reconstruct high-resolution three-dimensional ISAR sample images.

5. The method according to claim 4, characterized in that, The microwave photonics-enabled distributed array ISAR system includes a central control module, a transmitting module, and a receiving module connected centrally via optical fiber. The central control module includes a clock, a photonic quadrature waveform generator, a photonic multi-channel dechirping processor, an analog-to-digital converter, and a digital signal processor; The photonic orthogonal waveform generator includes a photonic digital-to-analog converter for generating a broadband intermediate frequency signal and a pulse shaper for up-converting the broadband intermediate frequency signal, wherein the pulse shaper outputs a reference signal. The photonic multichannel dechirping processor includes a photodetector, a bandpass filter, and a Mach-Zehnder modulator. The photodetector converts the reference signal into an electrical signal, which is then filtered by the bandpass filter and input to the Mach-Zehnder modulator to obtain the dechirped reference signal. The photonic multichannel dechirping processor further includes a wavelength division multiplexer, an erbium-doped fiber amplifier, a low-noise amplifier, an analog-to-digital converter, and a digital signal processor. The wavelength division multiplexer combines the echo signals received by the receiving module and inputs them to the Mach-Zehnder modulator for dechirping processing. The dechirped echo signals are amplified by the erbium-doped fiber amplifier and the low-noise amplifier, and then sent to the analog-to-digital converter and the digital signal processor to complete the digitization and computation of the signals. The transmitting module includes a photodetector, a bandpass filter, a low-noise amplifier, and a power amplifier; the optical signal generated by the photonic orthogonal waveform generator is converted into a radio frequency signal by the photodetector, and then transmitted after passing through the bandpass filter, the low-noise amplifier, and the power amplifier; The receiving module includes a low-noise amplifier, a bandpass filter, and a Mach-Zehnder modulator biased at the quadrature point. The received echo signal is amplified by the low-noise amplifier and filtered by the bandpass filter before being fed into the Mach-Zehnder modulator biased at the quadrature point, with a modulation wavelength of... The optical carrier wave completes the echo conversion between electrons and photons.

6. A target recognition method for three-dimensional ISAR images, characterized in that, include: Acquire the 3D ISAR image to be identified; The three-dimensional coordinates and scattering intensity of the point cloud in the three-dimensional ISAR image to be identified are input into the trained three-dimensional ISAR image recognition network model to obtain the recognition result of the point cloud target in the three-dimensional ISAR image to be identified; the three-dimensional ISAR image recognition network model is trained using the method described in any one of claims 1-5.

7. A training apparatus for a three-dimensional ISAR image recognition network model that implements the method of claim 1, characterized in that, include: The first training module is configured to take the three-dimensional coordinates of the point cloud in the three-dimensional ISAR sample image as input and the scattering intensity of the corresponding point in the point cloud as reference data to train the autoencoder in the deep autoencoder network, so that the autoencoder has the ability to regress the scattering intensity using the three-dimensional coordinates of the point cloud in the three-dimensional ISAR image to be identified. The second training module is configured to use the output of the encoding module in the already trained autoencoder as the input of the multilayer perceptron classifier in the deep autoencoder network, and the true types of point cloud targets in the 3D ISAR sample image as a reference to train the multilayer perceptron classifier, so that the multilayer perceptron classifier can acquire the ability to classify the point cloud targets using the feature matrix extracted by the encoder. The acquisition module is configured to obtain a trained 3D ISAR image recognition network model by combining the encoding module in the autoencoder and the multilayer perceptron classifier.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method of any one of claims 1-6.

9. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the method described in any one of claims 1-6.

Images