ISAR target identification method based on double-branch scattering feature fusion model
Through the dual-branch scattering feature fusion model, combining the ISAR encoding network and complex echo encoding network, the image domain and echo domain characteristics are combined, and the problems of low signal-to-noise ratio and noise sensitivity in ISAR target recognition are solved, and the ISAR target component is realized Improved precise identification and segmentation performance.
Patent Information
- Application Number
- CN202410924453.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-22
AI Technical Summary
The existing ISAR target recognition method is difficult to accurately segment the target components under low signal-to-noise ratio and noise sensitivity, and the existing multimodal feature fusion method has a large difference between the image domain and the signal domain, so it has not made full use of ISAR. Rich information about data.
Using the dual-branch scattering feature fusion model, the high-dimensional scattering features of the image domain and the echo domain are extracted by constructing the ISAR encoding network and the complex echo encoding network, and the high-dimensional scattering features of the image domain and the echo domain are processed through the statistical quantization module to fuse the image domain and the echo domain. Characteristics of wave domain to improve segmentation performance.
The precise identification of ISAR target components is achieved under the condition of few samples, which improves segmentation performance and identification accuracy, effectively suppresses noise interference, and makes full use of the characteristic information of ISAR data.
Smart Images

Figure CN119006993A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target recognition, and in particular to an ISAR target recognition method based on a double-branch scattering feature fusion model. Background Art
[0002] Inverse Synthetic Aperture Radar (ISAR) has the characteristics of all-day, all-weather, and not restricted by weather conditions. It is an important means to obtain geometric structure information of non-cooperative targets. The continuous development of imaging technology has promoted the improvement of ISAR imaging resolution. The precise interpretation technology of radar targets has also attracted more and more attention from researchers. Deep neural networks have shown great advantages in nonlinear modeling and feature mining, and have developed rapidly in the fields of space environment detection and medical image segmentation. The semantic segmentation algorithm can extract the target from the image, generate accurate local feature boundaries, and can segment the image into multiple independent regions. It is suitable for accurately extracting local components of the target from ISAR images.
[0003] Most of the target component recognition methods based on deep learning semantic segmentation are designed for optical images. ISAR imaging and optical imaging are different at the mechanism level, and the visual representation of target ISAR images and optical images is significantly different. Semantic segmentation methods based on optical images include clustering, threshold judgment and edge detection. The above methods are based on traditional feature designs such as pixel grayscale and problem gradient, and are difficult to adapt to ISAR images with low signal-to-noise ratio and noise sensitivity. Different from optical imaging, target ISAR images often have problems such as low signal-to-noise ratio, blurred edges, and severe sidelobe effects of some strong scattering points, which poses a severe challenge to the task of accurate component segmentation.
[0004] The target ISAR image is composed of discontinuous scattering points. The intensity information of the scattering points is affected by many factors such as the target shape, material and radar bandwidth. In addition, the change of the relative observation angle of the radar to the target may also cause a large difference in the target ISAR image.
[0005] Existing few-shot semantic segmentation methods have achieved different degrees of performance improvement by improving the matching strategy between support samples and query samples. However, the relatively scarce data conditions can only represent limited target differences, and intra-class diversity and inter-class differences are still important factors restricting the performance improvement of few-shot semantic segmentation algorithms.
[0006] Feature fusion methods that introduce more semantic information into low-level features or embed more spatial information into high-level features can help improve segmentation performance. There is a lack of available ISAR data for targets, and the reliability of segmentation components that rely solely on single-domain scattering pixel information is low. Fusion of multi-domain features helps obtain segmentation boundaries with higher confidence. Most multimodal feature fusion methods focus on images and images, and images and text. Due to the complexity of complex-valued echoes, there are few studies on multimodal feature fusion methods in image domain and signal domain. In the image domain, the recognition algorithm based on ISAR images only uses amplitude information and loses phase information. In the signal domain, the amplitude / phase in the ISAR complex-valued echo is important information for characterizing target features and contains richer information. There are large differences between the features of different modes. The key to designing a fusion model is to give full play to the advantages of fusion while not introducing negative information. Summary of the invention
[0007] The technical problem to be solved by the present invention is how to improve the recognition performance of an ISAR system; in view of this, the present invention provides an ISAR target recognition method based on a dual-branch scattering feature fusion model.
[0008] The technical solution adopted by the present invention is an ISAR target recognition method based on a dual-branch scattering feature fusion model, comprising:
[0009] Step S1, preprocessing the target ISAR data;
[0010] Step S2, dividing the current ISAR data into a training data set and a test data set according to a preset ratio, and training the current training data set;
[0011] Step S3, constructing a dual-branch scattering feature fusion model to extract high-dimensional scattering features in the ISAR image domain;
[0012] Step S4, under the experimental setting of few samples N way-K shot, input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model to obtain the corresponding high-dimensional feature tensor, and fuse the high-dimensional feature tensors of the dual branches to determine the target component segmentation prediction result;
[0013] Step S5, performing model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight;
[0014] Step S6, loading the model weights, and using the test data set to obtain a few-sample ISAR target component recognition result.
[0015] In one embodiment, step S1 includes:
[0016] Motion compensation is performed on the defocused ISAR image, and filtering and gamma transformation are used to improve the contrast of weak components in the image, making the detailed information in the ISAR image richer.
[0017] In one embodiment, step S2 includes:
[0018] Divide the training data set and the test data set into a ratio of 7:3, and divide the training data set into support samples and query samples;
[0019] Based on the training data set and the test data set, the mapping of the support samples and the query samples corresponding to the ISAR images and the complex-valued ISAR echoes of the corresponding targets and the mask labels of the support samples to the query samples is learned.
[0020] In one embodiment, step S3 includes:
[0021] A dual-branch ISAR coding network and a complex-valued echo coding network are constructed. The ISAR coding network is used to generate high-dimensional scattering features from ISAR images, and the complex-valued echo coding network is used to generate high-dimensional scattering features from complex-valued echoes.
[0022] In one embodiment, step S4 includes:
[0023] Input the target ISAR image into the ISAR encoding network, process the input data using a convolutional residual structure, increase the number of channels to three and obtain high-level semantic information;
[0024] The high-level semantic information is processed by downsampling and convolution to generate multi-scale scattering features in the image domain, including supporting features and query features;
[0025] Determine the similarity between the supporting features and the query features, and generate a similarity feature mapping matrix;
[0026] Processing the mapping matrix using a 1×1 convolution layer and a ReLU activation function to improve the coupling between the support sample and the query sample, and obtaining a strongly correlated scattering feature tensor corresponding to different receptive fields;
[0027] The spatial resolution of the scattering feature tensor is improved by upsampling, and the feature tensors of different scales of the three channels are adjusted to the same dimension and then spliced in the channel dimension to obtain a high-dimensional scattering feature derived from the ISAR image;
[0028] Input the original ISAR echo into the complex-valued echo coding network and perform preprocessing;
[0029] The echo domain input data is processed using the residual structure to generate echo domain multi-scale scattering feature tensors corresponding to different receptive fields;
[0030] Three SQM modules are used to process the multi-scale scattering features in the echo domain to generate three scales of strongly correlated scattering feature tensors;
[0031] After adjusting the feature tensors of different scales of the three channels to the same dimension, they are concatenated in the channel dimension to obtain high-dimensional scattering features derived from complex-valued echoes;
[0032] The high-dimensional feature tensors of the two branches are fused to determine the target component segmentation prediction result.
[0033] In one embodiment, step S5 includes:
[0034] Under the supervision of a pixel-by-pixel cross entropy loss function, the high-dimensional feature tensor and the component segmentation prediction result are used as inputs to determine a distance metric between the two types of inputs;
[0035] Determine a loss function to optimize training gradient based on the distance metric;
[0036] The model is trained based on the training gradient until the loss function converges or meets the preset requirements, and the final model weights are obtained.
[0037] Another aspect of the present invention further provides an ISAR target recognition device based on a dual-branch scattering feature fusion model, comprising:
[0038] A preprocessing unit configured to preprocess the target ISAR image;
[0039] A division unit is configured to divide the current ISAR image into a training data set and a test data set according to a preset ratio, and train the current training data set;
[0040] A construction unit is configured to construct a dual-branch scattering feature fusion model to extract high-resolution scattering texture features in the ISAR image domain and scattering quantization features in the complex-valued echo domain;
[0041] The computing unit is configured to input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model under the experimental setting of a few samples N way-K shot, obtain the corresponding high-dimensional feature tensor, and fuse the high-resolution scattering texture feature with the scattering quantization feature in the complex-valued echo domain to determine the target component segmentation prediction result;
[0042] A training unit configured to perform model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight;
[0043] The output unit is configured to load the model weights and obtain a few-sample ISAR target component recognition result using the test data set.
[0044] Another aspect of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the ISAR target recognition method based on the dual-branch scattering feature fusion model as described in any one of the above items.
[0045] Another aspect of the present invention further provides a computer storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the ISAR target recognition method based on the dual-branch scattering feature fusion model as described in any of the above items are implemented.
[0046] Compared with the prior art, the present invention has at least the following advantages:
[0047] Based on the current situation that the ISAR images of low signal-to-noise ratio targets have blurred edges and large differences in scattering intensity, which affect the recognition performance of the ISAR system, this paper adopts semantic segmentation to solve the ISAR target component recognition problem, and proposes a few-sample semantic segmentation framework based on a dual-branch scattering feature fusion model, which realizes accurate recognition of target components under few-sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 A schematic flow chart of an ISAR target recognition method based on a dual-branch scattering feature fusion model according to an embodiment of the present invention;
[0049] Figure 2 A logical schematic diagram of an ISAR target recognition method based on a dual-branch scattering feature fusion model according to an embodiment of the present invention;
[0050] Figure 3 A schematic diagram of a component segmentation scheme according to an embodiment of the present invention;
[0051] Figure 4 A schematic diagram of an ISAR target imaging result according to an embodiment of the present invention;
[0052] Figure 5 A schematic diagram of a convolution residual structure according to an embodiment of the present invention;
[0053] Figure 6 A schematic diagram of a dual-branch scattering feature encoding network according to an embodiment of the present invention;
[0054] Figure 7 is a structural diagram of an SQM according to an embodiment of the present invention;
[0055] Figure 8 A schematic diagram of a decoding network structure according to an embodiment of the present invention;
[0056] Fig. 9 is a schematic diagram of component segmentation results according to an embodiment of the present invention;
[0057] Fig.10 It is a schematic diagram of the structure of an ISAR target recognition device based on a dual-branch scattering feature fusion model according to an embodiment of the present invention;
[0058] Fig.11 The figure is a schematic diagram of the composition of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose, the present invention is described in detail below in conjunction with the accompanying drawings and preferred embodiments.
[0060] Unless otherwise defined, all terms (including technical terms and scientific terms) used in this article have the same meaning as commonly understood by ordinary technicians in the field to which this application belongs. It should also be understood that terms (such as terms defined in commonly used dictionaries) should be interpreted as having the same meaning as their meaning in the context of the relevant technology, and will not be interpreted in an idealized or overly formal sense unless explicitly defined in this article.
[0061] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0062] For ease of understanding, necessary explanations are given below on the relevant technologies involved in the present invention.
[0063] The embodiment of the present invention provides an ISAR target recognition method based on a dual-branch scattering feature fusion model, such as Figure 1 As shown, including:
[0064] Step S1, preprocessing the target ISAR data;
[0065] Step S2, dividing the current ISAR data into a training data set and a test data set according to a preset ratio, and training the current training data set;
[0066] Step S3, constructing a dual-branch scattering feature fusion model to extract high-dimensional scattering features in the ISAR image domain;
[0067] Step S4, under the experimental setting of few samples N way-K shot, input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model to obtain the corresponding high-dimensional feature tensor, and fuse the high-dimensional feature tensors of the dual branches to determine the target component segmentation prediction result;
[0068] Step S5, performing model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight;
[0069] Step S6, loading the model weights, and using the test data set to obtain a few-sample ISAR target component recognition result.
[0070] It should be noted that the target recognition method proposed in this article may specifically refer to a method for recognizing ISAR target components.
[0071] Can be combined Figure 2 as well as Figure 3 , the method provided in this embodiment will be described in detail step by step below.
[0072] Step S1, preprocessing the target ISAR data.
[0073] In this embodiment, the target ISAR imaging results often show low signal-to-noise ratio and blurred edges. The sidelobe effect of some strong scattering points affects the scattering information representation of the components in the ISAR image. The above factors may have a significant impact on the accuracy of component recognition. Therefore, it is necessary to first perform motion compensation on the defocused ISAR image, and at the same time use filtering and gamma transformation to improve the contrast of weak components in the image, so that the detailed information in the ISAR image is richer.
[0074] Step S2, dividing the current ISAR data into a training data set and a test data set according to a preset ratio, and training the current training data set.
[0075] For example, the few-sample ISAR target component recognition architecture proposed in this embodiment can divide the data into a training set and a test set D according to a preset ratio of 7:3. train ,D test , The training set contains 350 pairs of samples, and the test set contains 245 pairs of samples, all of which are composed of ISAR images and complex-valued echoes of five types of targets. Figure 4 , adopts the few-shot episodic learning paradigm, that is, the training data is divided into support samples and query samples. Support samples are represented by superscript s, and query samples are represented by superscript q. The training process consists of multiple iterations, each of which contains the support set S = (I s ,Echo s ,M s ) and the query set Q = (I q ,E cho q ,M q ). represents the ISAR image, represents the complex-valued ISAR echo of the corresponding target, is the mask label. In each training episode, the component segmentation model needs to learn (I q ,I s ,E cho q ,E cho s ,M s ) to M q 's mapping.
[0076] Step S3, constructing a dual-branch scattering feature fusion model to extract high-dimensional scattering features in the ISAR image domain.
[0077] In this embodiment, a dual-branch ISAR coding network and a complex-valued echo coding network are constructed. The ISAR coding network is used to generate high-dimensional scattering features from ISAR images, and the complex-valued echo coding network is used to generate high-dimensional scattering features from complex-valued echoes.
[0078] It should be noted that the feature embedding network is used to process the target ISAR image. Through multi-layer convolutional neural network operations, a high-dimensional scattering feature tensor F with both shallow texture information and high-level semantic information is obtained. s ,F q As a pixel-level classification task, semantic segmentation loses more detailed texture information when downsampling. Therefore, when designing the encoding network, avoid using a network structure with more downsampling or deeper layers.
[0079] Step S4, under the experimental setting of few-sample Nway-Kshot, the ISAR image and the corresponding ISAR echo data are input into the dual-branch scattering feature fusion model to obtain the corresponding high-dimensional feature tensor, and the high-dimensional feature tensors of the dual branches are fused to determine the target component segmentation prediction result.
[0080] (1) ISAR image coding branch:
[0081] The preprocessed target ISAR image is input into the image encoding branch. Using the convolution residual structure ( Figure 5The input data is processed by using a multi-scale scattering feature layer (as shown in Figure 2) to increase the number of channels while obtaining high-level semantic information. Following the structure of the feature pyramid, multi-scale scattering features in the image domain are generated after downsampling and convolution. The similarity between the support features and the query features is calculated to generate a similarity feature mapping matrix. The mapping matrix is then processed using a 1×1 convolution layer and a ReLU activation function to increase the coupling between the support samples and the query samples. Strongly correlated scattering feature tensors corresponding to different receptive fields are obtained. Finally, upsampling is used to improve the spatial resolution of the features. After adjusting the feature tensors of three different scales to the same dimension, they are spliced in the channel dimension to obtain high-dimensional scattering features from ISAR images.
[0082] (2) Complex-valued echo coding branch:
[0083] The input of the complex-valued echo encoding branch is the original ISAR echo, which is a linear frequency modulation signal in the form of a complex exponential. After the ISAR echo is processed by compensation and imaging algorithms, it can be used to generate radar images. Figure 6 The complex-valued echo coding branch first designs a complex-valued echo preprocessing module, which is processed by 1×1 complex convolution, complex PRELU activation function, 3×3 complex convolution, and complex PRELU activation function, and then concatenates the feature tensors of the real and imaginary parts in the channel dimension.
[0084] After preprocessing, use Figure 5 The residual structure shown processes the echo domain scattering features and generates echo domain multi-scale scattering feature tensors corresponding to different receptive fields. Secondly, three SQMs (statistical quantification modules) are used to process the echo domain multi-scale scattering features respectively to generate three proportions of strongly correlated scattering feature tensors. The complex-valued echo retains the amplitude and phase characteristics more completely. Since they originate from the same ISAR complex-valued echo, and SQM can accurately quantify local scattering features while retaining global features, more contextual information is retained between the three types of scattering feature tensors. The strong correlation improves the segmentation robustness of the component identification scheme described in the present invention for large-scale sailboards, small-scale loads, and other types of components.
[0085] It should be noted that the selection of thresholds is overly dependent on expert experience and the algorithm has insufficient generalization ability. This paper proposes a statistical quantization module SQM. SQM generates scattering statistical quantization codes based on the scattering intensity statistics of different components at different levels, and accurately describes the scattering information of different intensities in the echo domain based on the hierarchical quantization idea. SQM integrates the scattering information measurement into the intelligent module, effectively extracts the high-resolution scattering robust features of the target, and improves the segmentation accuracy of the components. The specific SQM structure is as follows: Figure 7 .
[0086] Furthermore, let the input feature be E∈R C×H×W , C, H, W represent the number of channels, height and width respectively. The height and width of the feature map are related to the number of sampling points in the ISAR azimuth dimension and Doppler dimension respectively. The next step of the processing of this embodiment can refer to the following process:
[0087] 1) Process the scattering characteristic map from ISAR echo and establish each spatial point E i,j Similarity mapping with pooling factor g
[0088]
[0089] Among them, GAP(*) represents global average pooling. ||*||2 represents the calculation of the bi-norm.
[0090] 2) Hierarchical quantization. Assume that the average is divided into T levels, L amp =[L1,L2,...L t ,...,L T ]. amp Represents a set of quantization layers divided according to the spatial point scattering intensity information. After hierarchical quantization, the statistical coding feature A∈R is generated T×HW . That is, using T statistical metric functions to process Each space vector A mapping relationship is established based on the echo scattering intensity, which is classified into different levels L t .L t , The calculation formula is as follows:
[0091]
[0092] 3) Calculate the quantization operator. Adjust the dimension of statistical coding features. The radar's observation angle of the target is constantly changing. Changes in the observation angle will cause changes in the amplitude and phase parameters of the component echo. In order to improve the robustness of SPSF in identifying target components at different observation angles, SQM calculates the product of each spatial tensor in A and its neighborhood as a feature quantization operator that characterizes the scattering characteristics of the current tensor. For every pair of adjacent tensors A in A i,j ∈R T×1 ,A i,j+1 ∈R T×1 , calculate the product A i,* is a statistical descriptor of dimension T. When the scattering intensity information of the spatial tensor belongs to the level L t , the statistical descriptor of the corresponding position is set to 1, otherwise it is 0. Similarly, A*,j It is also a statistical descriptor of dimension T. When the tensor belongs to level L t' When , the descriptor corresponds to position 1. Since the ISAR echo scattering features are spatially continuous, SQM smoothes the embedded features to avoid the gradient of the training process being unable to be returned.
[0093] 4) Calculate the statistic F. Scattering quantification characteristics Used to characterize the scattering characteristics distribution of ISAR echoes. (Approximately the number of spatial tensors corresponding to each quantization level).
[0094]
[0095] Use pReLU activation function to process g, After adjusting the dimension, concatenate the tensors and get Represents all possible pairs of scattered quantization levels of adjacent spatial tensors in the feature map.
[0096]
[0097] The specific structure of the decoding network is as follows Figure 8 After being processed by the dual-branch scattering feature encoding network, the dimension of the high-dimensional scattering feature from the ISAR image is (B×C img ×H p ×W p ), the dimension of the high-dimensional scattering feature from ISAR echo is (B×C echo ×H p' ×W p' ). Where B represents the batch training data size. C represents the number of feature channels. p Represents the spatial scale of the feature tensor.
[0098] First, use decoding block-1 to process the image domain scattering features, reduce the number of feature channels, and improve the spatial resolution. The image domain scattering features processed by decoding block-1 are spliced with the scattering features in the echo domain. At this time, the image domain high-dimensional features and the echo domain high-dimensional features have the same spatial resolution, and the number of channels is 128. After feature splicing in the channel dimension, a scattering fusion feature map with a feature number of 256 is obtained. Secondly, use decoding block-2 to process the scattering fusion feature map. Decoding block-2 is used to adjust the resolution of the fused feature map to the same scale as the original ISAR image. At the same time, the number of channels is gradually compressed. The last 1×1 convolution layer adjusts the number of channels to the same number as the number of categories to be identified. After processing by the softmax function, the component segmentation result is generated.
[0099] In order to make the initial scatter prediction mask generated by the query sample closer to the real mask, cover more areas that are conducive to improving segmentation accuracy, and make the features generated by the query sample contain clearer division boundaries, the following pixel-by-pixel cross entropy loss function is used to optimize the component segmentation framework.
[0100]
[0101] Among them, L ce Indicates the calculation of pixel-by-pixel cross entropy loss. fusion Represents the high-dimensional scattering feature tensor generated by fusing the scattering features in the image domain and the echo domain. It represents the prediction result output by the component segmentation method of the present invention.
[0102] Step S5, performing model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine model weights.
[0103] Specifically, mean Intersection-over-Union (mIoU), the most common metric in the field of few-shot semantic segmentation, is used as the evaluation metric. At the same time, the component recognition algorithm proposed in the present invention is evaluated under 1-shot / 5-shot settings. Considering the differences between optical data and ISAR imaging principles, the present invention does not load a pre-trained model, and uses the ISAR dataset to train the component recognition model from scratch.
[0104] The initial learning rate is 1e-3, the momentum factor is set to 0.9, and the weight decay is set to 0.0001. During the algorithm training process, the poly strategy is used to optimize the learning rate setting. The learning rate in each iteration is the initial parameter multiplied by The present invention is implemented using Pytorch 1.11. The experimental platform is equipped with RTX3070GPU (cuda11.3) and Intel i7-10875 CPU. The experimental results are based on the accuracy without pre-training. Five types of random seeds are generated in each test process, and the mean of the five experimental results is retained in combination with the test data.
[0105] The total number of training rounds is 1500, and the batch size is set to 20. During the training process, the initial size of the target ISAR image and its mask is retained for input feature embedding network training, that is, the height and width of the input data are both 256 pixels, which is consistent with the number of sampling points in the ISAR dataset construction process. After the training is completed, the weight model of the segmentation algorithm is saved to test the algorithm segmentation accuracy.
[0106] Step S6, loading the model weights, and using the test data set to obtain a few-sample ISAR target component recognition result.
[0107] That is, load the weight model obtained in the algorithm training phase, input the constructed test data set into the few-sample component segmentation model, and output the target component segmentation result as Fig. 9 .
[0108] Compared with the prior art, the method provided by the embodiment of the present invention has at least the following advantages:
[0109] 1) Aiming at the current situation that the ISAR images of low signal-to-noise ratio targets have blurred edges and large differences in scattering intensity, which affect the recognition performance of the ISAR system, and the lack of data poses a greater challenge to the accurate identification of components, this paper uses semantic segmentation to solve the problem of ISAR target component recognition and proposes a few-sample semantic segmentation framework based on a dual-branch scattering feature fusion model, which achieves accurate recognition of target components under few-sample conditions.
[0110] 2) The dual-branch scattering feature encoding structure can fully exploit the scattering texture information in the image domain and the scattering quantization information in the echo domain; the branch scattering feature fusion model extracted by the present invention aggregates high-level semantic features and shallow scattering texture into accurate segmentation features. By fusing the fine-grained scattering texture features in the image domain with the scattering quantization information in the echo domain, a more accurate segmentation boundary is obtained.
[0111] Aiming at the problem of blurred edges of low signal-to-noise ratio ISAR images, the present invention designs a statistical quantization module to generate scattering quantization coding to mine the scattering characteristics of ISAR echoes, and the feature fusion structure effectively suppresses noise interference.
[0112] To address the problem of high complexity of complex domain echo data, the present invention constructs an echo feature coding network based on complex domain convolution, enhances the coupling between query features and supporting features, and further improves the matching accuracy of differential features.
[0113] The proposed multi-layer feature mining architecture fully retains the contextual information of the components and improves the correlation between features. The dual-branch scattering feature fusion model greatly improves the feature utilization of ISAR data. The fused features help improve the component recognition performance.
[0114] As an advanced solution for ISAR target component identification, the present invention can provide important data support for tasks such as target state analysis.
[0115] The second embodiment of the present invention corresponds to the first embodiment. This embodiment introduces an ISAR target recognition device based on a dual-branch scattering feature fusion model. Fig.10 , including the following components:
[0116] A preprocessing unit configured to preprocess the target ISAR image;
[0117] A division unit is configured to divide the current ISAR image into a training data set and a test data set according to a preset ratio, and train the current training data set;
[0118] A construction unit is configured to construct a dual-branch scattering feature fusion model to extract high-resolution scattering texture features in the ISAR image domain and scattering quantization features in the complex-valued echo domain;
[0119] The computing unit is configured to input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model under the experimental setting of a few samples N way-K shot, obtain the corresponding high-dimensional feature tensor, and fuse the high-resolution scattering texture feature with the scattering quantization feature in the complex-valued echo domain to determine the target component segmentation prediction result;
[0120] A training unit configured to perform model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight;
[0121] The output unit is configured to load the model weights and obtain a few-sample ISAR target component recognition result using the test data set.
[0122] A third embodiment of the present invention is an electronic device, referring to Fig.11 , can be understood as a physical device, including a processor and a memory storing instructions executable by the processor. When the instructions are executed by the processor, the following operations are performed:
[0123] Step S1, preprocessing the target ISAR data;
[0124] Step S2, dividing the current ISAR data into a training data set and a test data set according to a preset ratio, and training the current training data set;
[0125] Step S3, constructing a dual-branch scattering feature fusion model to extract high-dimensional scattering features in the ISAR image domain;
[0126] Step S4, under the experimental setting of few samples N way-K shot, input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model to obtain the corresponding high-dimensional feature tensor, and fuse the high-dimensional feature tensors of the dual branches to determine the target component segmentation prediction result;
[0127] Step S5, performing model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight;
[0128] Step S6, loading the model weights, and using the test data set to obtain a few-sample ISAR target component recognition result.
[0129] The fourth embodiment of the present invention, the process of the ISAR target recognition method based on the dual-branch scattering feature fusion model of this embodiment is the same as that of the first, second or third embodiment, the difference is that, in engineering implementation, this embodiment can be implemented by means of software plus a necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the method of the present invention can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), including a number of instructions for a device to execute the method described in the embodiment of the present invention.
[0130] Through the description of the specific implementation methods, a deeper and more specific understanding of the technical means and effects adopted by the present invention to achieve the predetermined purpose should be obtained. However, the accompanying drawings are only for reference and illustration purposes and are not intended to limit the present invention.
Claims
1. An ISAR target recognition method based on a dual-branch scattering feature fusion model, characterized in that: include: Step S1, preprocessing the target ISAR data; Step S2, dividing the current ISAR data into a training data set and a test data set according to a preset ratio, and training the current training data set; Step S3, constructing a dual-branch scattering feature fusion model to extract high-dimensional scattering features in the ISAR image domain; Step S4, under the experimental setting of few samples N way-K shot, input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model to obtain the corresponding high-dimensional feature tensor, and fuse the high-dimensional feature tensors of the dual branches to determine the target component segmentation prediction result; Step S5, performing model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight; Step S6, loading the model weights, and using the test data set to obtain a few-sample ISAR target component recognition result.
2. The ISAR target recognition method based on the dual-branch scattering feature fusion model according to claim 1 is characterized in that: The step S1 comprises: Motion compensation is performed on the defocused ISAR image, and filtering and gamma transformation are used to improve the contrast of weak components in the image, making the detailed information in the ISAR image richer.
3. The ISAR target recognition method based on the dual-branch scattering feature fusion model according to claim 2 is characterized in that: The step S2 comprises: Divide the training data set and the test data set into a ratio of 7:3, and divide the training data set into support samples and query samples; Based on the training data set and the test data set, the mapping of the support samples and the query samples corresponding to the ISAR images and the complex-valued ISAR echoes of the corresponding targets and the mask labels of the support samples to the query samples is learned.
4. The ISAR target recognition method based on the dual-branch scattering feature fusion model according to claim 3 is characterized in that: The step S3 comprises: A dual-branch ISAR coding network and a complex-valued echo coding network are constructed. The ISAR coding network is used to generate high-dimensional scattering features from ISAR images, and the complex-valued echo coding network is used to generate high-dimensional scattering features from complex-valued echoes.
5. The ISAR target recognition method based on the dual-branch scattering feature fusion model according to claim 4 is characterized in that: The step S4 comprises: Input the target ISAR image into the ISAR encoding network, process the input data using a convolutional residual structure, increase the number of channels to three and obtain high-level semantic information; The high-level semantic information is processed by downsampling and convolution to generate multi-scale scattering features in the image domain, including supporting features and query features; Determine the similarity between the supporting features and the query features, and generate a similarity feature mapping matrix; Processing the mapping matrix using a 1×1 convolution layer and a ReLU activation function to improve the coupling between the support sample and the query sample, and obtaining a strongly correlated scattering feature tensor corresponding to different receptive fields; The spatial resolution of the scattering feature tensor is improved by upsampling, and the feature tensors of different scales of the three channels are adjusted to the same dimension and then spliced in the channel dimension to obtain a high-dimensional scattering feature derived from the ISAR image; Input the original ISAR echo into the complex-valued echo coding network and perform preprocessing; The echo domain input data is processed using the residual structure to generate echo domain multi-scale scattering feature tensors corresponding to different receptive fields; Three SQM modules are used to process the multi-scale scattering features in the echo domain to generate three scales of strongly correlated scattering feature tensors; After adjusting the feature tensors of different scales of the three channels to the same dimension, they are concatenated in the channel dimension to obtain high-dimensional scattering features derived from complex-valued echoes; The high-dimensional feature tensors of the two branches are fused to determine the target component segmentation prediction result.
6. The ISAR target recognition method based on the dual-branch scattering feature fusion model according to claim 5 is characterized in that: The step S5 comprises: Under the supervision of a pixel-by-pixel cross entropy loss function, the high-dimensional feature tensor and the component segmentation prediction result are used as inputs to determine a distance metric between the two types of inputs; Determine a loss function to optimize training gradient based on the distance metric; The model is trained based on the training gradient until the loss function converges or meets the preset requirements, and the final model weights are obtained.
7. An ISAR target recognition device based on a dual-branch scattering feature fusion model, characterized in that: include: A preprocessing unit configured to preprocess the target ISAR image; A division unit is configured to divide the current ISAR image into a training data set and a test data set according to a preset ratio, and train the current training data set; A construction unit is configured to construct a dual-branch scattering feature fusion model to extract high-resolution scattering texture features in the ISAR image domain and scattering quantization features in the complex-valued echo domain; The computing unit is configured to input the ISAR image and the corresponding ISAR echo data into the dual-branch scattering feature fusion model under the experimental setting of a few samples N way-K shot, obtain the corresponding high-dimensional feature tensor, and fuse the high-resolution scattering texture feature with the scattering quantization feature in the complex-valued echo domain to determine the target component segmentation prediction result; A training unit, configured to perform model training using the target component segmentation prediction result and the high-dimensional feature tensor to determine a model weight; The output unit is configured to load the model weights and obtain a few-sample ISAR target component recognition result using the test data set.
8. An electronic device, characterized in that: The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the ISAR target recognition method based on the dual-branch scattering feature fusion model as described in any one of claims 1 to 6 are implemented.
9. A computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the ISAR target recognition method based on the dual-branch scattering feature fusion model as described in any one of claims 1 to 6.
Citation Information
Cited By
SAR target identification method based on echo domain scattering point collaborative learning
CN120539695A
