A radar imaging intelligent target recognition benchmark algorithm fast construction method

Abstract

The application provides a radar imaging intelligent target recognition benchmark algorithm rapid construction method, which uses four parts of a programming environment management module, a running environment management module, a benchmark algorithm management module and a process control module to rapidly construct a radar imaging intelligent target recognition benchmark algorithm.The application realizes rapid construction of the radar imaging intelligent target recognition benchmark algorithm under a complex background environment and a use scene; the modules proposed by the latest technology are written or adaptively modified according to the specification format of the neural network module, the calling of the intelligent target recognition benchmark algorithm model is realized, and the algorithm performance is improved; the function modules, the width, the depth and other structures of the intelligent target recognition network are modified based on the module modification of the benchmark algorithm construction process, the model parameter quantity can be reduced and the recognition speed can be improved under the premise of ensuring the accuracy of the radar imaging intelligent target recognition; and the target recognition benchmark algorithm rapid construction under multiple operating systems, multiple frameworks and multiple development environments is realized.

Description

Technical Field

[0001] This invention relates to the field of measurement and testing technology, and specifically to a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging. Background Technology

[0002] Today's society is an information society built on high technology, where the collection, transmission, processing, and application of information permeate the entire process of social operation. In urban planning and land resource management, synthetic aperture radar (SAR) and inverse synthetic aperture radar (ISAR) systems can acquire high-resolution remote sensing images around the clock. Therefore, rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging is of great significance for near real-time processing of acquired remote sensing images.

[0003] SAR image target recognition primarily targets large targets such as buildings, bridges, and airports, as well as small targets such as aircraft, engineering vehicles, and passenger cars, for accurate classification, identification, and localization. It is a crucial part of SAR image interpretation. ISAR is another radar imaging technology developed based on SAR, designed to solve the imaging problem of moving targets. It can obtain high-resolution remote sensing images of moving targets, such as aircraft, ships, and automobiles.

[0004] SAR / ISAR target recognition technology can be applied to various platforms such as satellites, drones, and weather balloons. Satellite platforms, due to their high operating orbits, offer a wider imaging swath compared to other platforms, and their imaging area is not restricted by airspace sovereignty. Acquiring SAR remote sensing images is stable, secure, and efficient, thus satellite-borne SAR / ISAR target recognition can play an important role in urban planning and land resource management. With the continuous development of deep learning technology and its enhanced feature extraction capabilities, researching intelligent target recognition in SAR / ISAR imaging is one of the directions in the development of artificial intelligence.

[0005] In the past decade, due to the rapid development of VLSI, high-speed integrated circuits, neural networks, quantum computing technology, multi-sensor integration technology, artificial intelligence technology, deep learning, big data technology, etc., research on target recognition algorithms based on deep learning has also increased both domestically and internationally. In terms of theoretical research, facing numerous target recognition algorithms based on deep learning, they can be roughly divided into three categories: (1) methods based on region sampling, target classification, and target position correction; (2) deep convolutional target recognition algorithms based on regression; (3) combination methods of the first two categories, with typical representatives being SSD (SingleShot MultiBox Detector), YOLO series, Retina-Net, and other algorithm models. Due to the wide variety of benchmark algorithms for intelligent target recognition in radar imaging, how to quickly construct and use them has become an urgent problem to be solved.

[0006] Therefore, a method is needed to quickly construct a benchmark algorithm for intelligent target recognition in radar imaging. Summary of the Invention

[0007] This invention aims to solve the problem of rapid construction by providing a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging. Based on the requirements of intelligent target recognition in radar imaging, the method uses four parts—programming environment management module, runtime environment management module, benchmark algorithm management module, and process control module—to rapidly construct a benchmark algorithm for intelligent target recognition in radar imaging to identify targets in SAR / ISAR images. This overcomes the shortcomings of traditional recognition methods, such as the mismatch between requirements and target recognition algorithms and the low construction speed of the benchmark algorithm.

[0008] This invention provides a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging, comprising the following steps:

[0009] S1. Based on the background environment and target complexity of the SAR / ISAR image recognition scene, the user selects the background environment and target of the recognition scene in the process control module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system. The process control module calls the SAR / ISAR image sample database to generate a specific SAR / ISAR experimental test dataset for the intelligent target recognition algorithm.

[0010] S2. Based on the target recognition requirements of a specific SAR / ISAR experimental test dataset, users can select at least one intelligent target recognition benchmark algorithm from the benchmark algorithm management module of the radar imaging intelligent target recognition benchmark algorithm quick construction system, or build and develop a new intelligent target recognition benchmark algorithm themselves. They can also import, delete, modify, build, and iteratively upgrade the selected algorithm model according to their usage requirements.

[0011] S3, the runtime environment management module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system provides the algorithm runtime environment for the selected intelligent target recognition benchmark algorithm. The algorithm runtime environment includes a runtime programming environment and a deep learning framework.

[0012] S4. The programming environment management module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system provides an algorithm integration development environment for the intelligent target recognition benchmark algorithm selected in step S2. The algorithm integration development environment includes a programming environment and a programming platform. The programming environment management module calls the corresponding programming platform according to the programming environment required to build the intelligent target recognition benchmark algorithm.

[0013] S5. The process control module configures parameters and controls the running process of the intelligent target recognition benchmark algorithm selected in step S2. After the selected intelligent target recognition benchmark algorithm completes the programming environment and running environment configuration, the process control module configures the algorithm training parameters and SAR / ISAR test dataset parameters according to the instructions, and runs the intelligent target recognition benchmark algorithm after calling a specific SAR / ISAR experimental test dataset. The process control module and the benchmark algorithm management module can optimize the parameters of the intelligent target recognition benchmark algorithm and add the modified new model to the benchmark algorithm library by the model management function of the benchmark algorithm management module. A rapid construction method for radar imaging intelligent target recognition benchmark algorithm is completed.

[0014] The present invention provides a method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm. In a preferred embodiment, in step S1, the background environment of the identified scene includes: desert, ocean, mountain, plain, and port.

[0015] The SAR / ISAR image sample database contains no fewer than 5,000 target images, no fewer than 5 background environments, no fewer than 10 target types, and no fewer than 6 image resolutions.

[0016] The target types include any one or more of the following: aircraft, bulldozers, vehicles, buildings, oil tankers, transport ships, fishing boats, ports, trucks, and islands / reefs.

[0017] Image resolutions include: 0.3m, 0.5m, 0.8m, 1m, 1.2m, and 3m;

[0018] A specific SAR / ISAR experimental test dataset includes a training set, a validation set, and a test set.

[0019] The rapid construction method for a radar imaging intelligent target recognition benchmark algorithm according to the present invention, in a preferred embodiment, includes step S2 as follows:

[0020] S21. Based on the target recognition requirements of a specific SAR / ISAR experimental test dataset, use the built-in intelligent target recognition basic algorithm in the benchmark algorithm library of the benchmark algorithm management module, proceed to step S22, and the user builds and develops the benchmark algorithm model, proceed to step S23.

[0021] S22. The user selects one or more intelligent target recognition benchmark algorithms from the benchmark algorithm library and proceeds to step S24.

[0022] S23. The benchmark algorithm management module contains a built-in neural network module and the latest module written or adapted according to the standard format of the neural network module. The built-in neural network module and the latest module are combined using YAML documents, and the input and output parameters are marked. By parsing the YAML documents, the model architecture, depth and width are modified and developed to obtain a new intelligent target recognition benchmark algorithm, which is then added to the benchmark algorithm library by the model management function of the benchmark algorithm management module. The new intelligent target recognition benchmark algorithm participates in the rapid construction of radar imaging intelligent target recognition benchmark algorithms.

[0023] S24. Users can import, delete, modify, and iteratively upgrade intelligent target recognition benchmark algorithms through the benchmark algorithm management module.

[0024] The rapid construction method for radar imaging intelligent target recognition benchmark algorithm of the present invention, as a preferred embodiment, in step S23, the built-in neural network module in the benchmark algorithm management module includes: Conv, C3, SPPF, Concat, Detect, Upsample, and Focus;

[0025] The latest modules include the CBAM module, DRM module, and CFN module.

[0026] In the preferred embodiment of the radar imaging intelligent target recognition benchmark algorithm construction method described in this invention, in step S23, the built-in neural network module is the neural network module of the YOLOv5 target recognition algorithm. A newly written CBAM module is added to the YOLOv5 YAML model document and inserted into the Neck part of the neural network module. The CBAM module as a whole adopts a cascaded channel attention module and a spatial attention module, and introduces the CBAM attention mechanism to enhance the spatial information of the features.

[0027] The present invention provides a method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm. In a preferred embodiment, step S3 includes running programming environments such as Anaconda, MSVC, MinGW, and Matlab; and deep learning frameworks such as TensorFlow, PaddlePaddle, PyTorch, and TensorRT.

[0028] Anaconda is a cross-platform Python environment management tool that includes runtime environments for six typical target recognition algorithms. It allows users to create, configure, and manage runtime environments for target recognition algorithms composed of different deep learning frameworks and Python libraries via commands.

[0029] The MSVC runtime environment is part of the Visual Studio integrated development environment. By selecting the corresponding MSVC runtime environment and Visual Studio through commands, the C / C++ inference and deployment development of typical image target recognition algorithms on the Windows platform can be implemented.

[0030] MinGw is a cross-platform C / C++ environment management tool. MinGw includes Windows-specific header files and a collection of import libraries using the GNU toolset. By selecting the runtime environment through commands, it enables the development, deployment, and inference of typical image target recognition algorithms in C / C++ on multiple system platforms.

[0031] The Deep Learning Toolbox is a toolbox within the Matlab development environment. It allows users to design, implement, and run target recognition algorithms by selecting the deep neural network framework corresponding to the Matlab version they are running on using commands.

[0032] The rapid construction method for radar imaging intelligent target recognition benchmark algorithm described in this invention, as a preferred embodiment, includes the following programming environments in step S4: Python, C++, and Matlab.

[0033] Programming platforms include: PyCharm, Microsoft Visual Studio, VS Code, CLion, and MATLAB.

[0034] The present invention provides a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging. As a preferred approach, PyCharm is used as the Python integrated development environment, and the target recognition algorithm is developed based on the PyTorch, TensorFlow, and PaddlePaddle deep learning frameworks.

[0035] Microsoft Visual Studio is a Windows-based integrated development environment for C / C++.

[0036] VS Code and CLion are cross-platform compatible, supporting Windows, macOS, and Linux operating systems. VS Code supports multiple programming languages, while CLion is a cross-platform IDE designed in C and C++. CLion uses the TensorRT inference framework to infer and deploy object recognition algorithms. MATLAB uses its built-in deep learning toolbox to develop and implement deep learning tasks within the MATLAB environment.

[0037] The present invention provides a method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm. In a preferred embodiment, the radar imaging intelligent target recognition benchmark algorithm rapid construction system includes a programming environment management module, a runtime environment management module, a benchmark algorithm management module, and a process control module electrically connected to the programming environment management module, the runtime environment management module, and the benchmark algorithm management module.

[0038] The benchmark algorithm management module includes a benchmark algorithm library and a benchmark algorithm management module. The benchmark algorithm library contains at least one intelligent target recognition benchmark algorithm. The benchmark algorithm library can also support the construction of self-developed benchmark recognition algorithm models. The benchmark algorithm management module manages the selected algorithm models according to the user's needs.

[0039] The programming environment management module provides an integrated development environment for intelligent target recognition benchmark algorithms based on the user's selected runtime environment;

[0040] The process control module constructs a specific SAR / ISAR experimental test dataset that meets the application scenario requirements according to the radar imaging type, background environment, and target recognition requirements selected by the user. It configures the parameters and controls the running process of the intelligent target recognition algorithm, configures the algorithm training parameters and SAR / ISAR test dataset parameters according to the instructions, and runs the intelligent target recognition benchmark algorithm.

[0041] The present invention provides a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging. As a preferred embodiment, the benchmark algorithm management module includes the following intelligent target recognition benchmark algorithms: Fast R-CNN, Faster R-CNN, YOLOv3, YOLOv5, YOLOv7, and YOLOv8.

[0042] Fast R-CNN is an end-to-end target recognition model that uses the Rol pooling layer to extract all Rol features in a single forward propagation.

[0043] Faster R-CNN is a target recognition model that uses a proposed region network structure instead of the candidate region operation in Fast R-CNN.

[0044] YOLOv3 is a target recognition model that improves the overall accuracy of SAR images and the recognition accuracy of small targets by incorporating multi-scale prediction. The basic network of YOLOv3 is Darknet-53.

[0045] YOLOv5 is a target recognition model that uses an FPN+PAN structure, efficiently utilizes feature information at different scales, and is easy to deploy and use on various platforms.

[0046] YOLOv7 is a target recognition model that uses a coarse-to-fine guided label assignment method, reduces the number of parameters, and improves recognition speed.

[0047] YOLOv8 is a target recognition model that supports image classification, detection, and segmentation tasks using a more lightweight network structure and a new loss function.

[0048] This invention focuses on solving the challenge of rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging. This invention has a certain degree of versatility in various radar imaging target recognition fields. Addressing issues such as multiple operating systems, multiple programming environments, complex target recognition scenes, and rapidly changing background environments encountered in practical applications, it rapidly modifies the network structure of the target recognition algorithm, enabling convenient and rapid construction of a benchmark algorithm for intelligent target recognition in radar imaging to identify the categories of SAR / ISAR imaging targets.

[0049] The technical solution of this invention is to rapidly construct a radar imaging intelligent target recognition benchmark algorithm using four parts: a programming environment management module, a runtime environment management module, a benchmark algorithm management module, and a process control module. The specific steps are as follows:

[0050] (1) Based on the background environment and target complexity of the SAR / ISAR image recognition scene, select the corresponding desert, ocean, mountain, plain, port and other background environments and recognition targets in the process control module to generate a specific SAR / ISAR experimental test dataset for the intelligent target recognition algorithm.

[0051] The SAR / ISAR image sample database contains no fewer than 5,000 target images, no fewer than 5 background environments, and no fewer than 10 target types. The image resolutions in the database are 0.3m, 0.5m, 0.8m, 1m, 1.2m, and 3m. This module can construct datasets that meet the needs of different application scenarios based on the different target recognition requirements of different radar imaging under different background environments.

[0052] (2) Based on the target recognition requirements of a specific SAR / ISAR experimental test dataset, select one or more intelligent target recognition benchmark algorithms from the benchmark algorithm management module, or build your own benchmark algorithm model. Manage the selected algorithm model according to your needs, such as importing, deleting, modifying, or building your own. A schematic diagram of the benchmark algorithm management is shown below. Figure 2 As shown.

[0053] The benchmark algorithm management module provides six typical image target recognition algorithms: Fast R-CNN, Faster R-CNN, YOLOv3, YOLOv5, YOLOv7, and YOLOv8. In addition to the built-in typical algorithms, additional typical image target recognition algorithms can be added according to usage requirements. Fast R-CNN is an end-to-end target recognition model that uses a Rol pooling layer to extract all Rol features in a single forward propagation, improving processing efficiency. Faster R-CNN uses a Proposal Region Network (RPN) structure instead of the candidate region operation in Fast R-CNN, improving detection speed. YOLOv3's base network is Darknet-53, which improves the overall accuracy of SAR image recognition and the recognition accuracy of small targets by incorporating multi-scale prediction. YOLOv5 uses an FPN+PAN structure to efficiently utilize feature information at different scales and is easy to deploy on various platforms. YOLOv7 proposes a coarse-to-fine guided label assignment method, reducing the number of parameters and improving recognition speed. YOLOv8 uses a more lightweight network structure and a new loss function, supporting tasks such as image classification, detection, and segmentation.

[0054] In addition to providing typical target recognition algorithms, the benchmark algorithm management module also supports operations such as importing, deleting, modifying, and independently building and iteratively upgrading algorithm models.

[0055] The architectures of typical image object recognition benchmark algorithms, such as YOLOv5 and YOLOv7, are written and stored in YAML format. Thanks to YAML's user-friendly data serialization format, which represents data in easily readable text, an increasing number of network models are built using YAML documents. Compared to other markup languages ​​such as XML or JSON, YAML is more concise, easier to read, and can interact with multiple programming languages.

[0056] The target recognition benchmark algorithm model consists of multiple neural network modules combined in different ways. Common neural network modules include Conv, C3, SPPF, Concat, Detect, Upsample, Focus, etc. The latest modules proposed in existing technologies are written or adapted according to the standardized format of neural network modules, such as CBAM, DRM, CFN, etc. The modules are combined in a YAML document, specifying the input and output parameters of each module. By parsing the YAML document, modifications and development of the model architecture, depth, and width can be quickly achieved, constructing new target recognition algorithm models. New models, whether modified or independently developed, are added to the benchmark algorithm library by the model management function of the benchmark algorithm management module, participating in the rapid construction of intelligent target recognition benchmark algorithms for radar imaging.

[0057] (3) The runtime environment management module provides a suitable runtime environment for the selected intelligent target recognition benchmark algorithm. Since different intelligent target recognition benchmark algorithms are developed based on different vendors or different versions of deep learning frameworks, the appropriate runtime environment is selected in the runtime environment management module according to the environment requirements of the selected benchmark algorithm.

[0058] The runtime environment management module provides programming environments with different configurations, including Anaconda, MSVC, MinGW, and Matlab. Mainstream open-source deep learning frameworks include TensorFlow, PaddlePaddle, PyTorch, and TensorRT. Therefore, various typical object recognition algorithms require this module to manage and configure different programming environments to support the use of different programming languages ​​or feature packages.

[0059] Anaconda is an open-source, cross-platform Python environment management tool that includes runtime environments for six typical object recognition algorithms. It allows for easy creation, configuration, and management of runtime environments for object recognition algorithms composed of different deep learning frameworks and Python libraries via commands. The MSVC runtime environment is part of the Visual Studio integrated development environment. By selecting the appropriate MSVC runtime environment and Visual Studio via commands, typical image object recognition algorithms in C / C++ can be inferred, deployed, and developed on the Windows platform. MinGW is an open-source, cross-platform C / C++ environment management tool that includes Windows-specific header files and a collection of import libraries using the GNU toolset. By selecting this runtime environment via commands, typical image object recognition algorithms in C / C++ can be inferred, deployed, and developed on multiple system platforms. The Deep Learning Toolbox is a toolbox within the Matlab development environment. By selecting the corresponding version of Matlab's deep neural network framework via commands, object recognition algorithms can be designed, implemented, and run. This document outlines the process of selecting a deep learning runtime environment for benchmark object recognition algorithms.

[0060] (4) The programming environment management module provides a suitable algorithm integration and development environment for the selected intelligent target recognition benchmark algorithm. Due to differences in system environment and development habits, the corresponding system and version of the intelligent target recognition benchmark algorithm integration and development environment should be selected in the programming environment management module according to the selected operating environment.

[0061] The programming environment management module mainly consists of different programming platforms. Typical algorithms are developed and run on different programming platforms based on different programming languages. Common programming environments include Python, C++, and Matlab, while programming platforms include PyCharm, Microsoft Visual Studio, VS Code, CLion, and MATLAB. The module calls the corresponding programming platform based on the programming environment required to build the target recognition benchmark algorithm, thereby managing the algorithm's programming environment.

[0062] PyCharm is a Python integrated development environment (IDE) for developing object recognition algorithms based on deep learning frameworks such as PyTorch, TensorFlow, and PaddlePaddle. Microsoft Visual Studio is a Windows-based IDE primarily for C / C++. VS Code and CLion are cross-platform compatible, supporting multiple operating systems including Windows, macOS, and Linux. VS Code supports multiple programming languages, while CLion is a cross-platform IDE specifically designed for C and C++ development, using the TensorRT inference framework for inference deployment of object recognition algorithms. MATLAB utilizes its built-in Deep Learning Toolbox to develop and implement deep learning tasks within the MATLAB environment.

[0063] Steps (3) and (4) both involve building a training model based on a YAML file.

[0064] (5) The process control module configures parameters and controls the operation flow of the intelligent target recognition algorithm. After the selected intelligent target recognition benchmark algorithm completes the programming and operation environment configuration, the process control module configures the algorithm training parameters and SAR / ISAR test dataset parameters according to the instructions, and runs the benchmark algorithm. The process control module and the benchmark algorithm management module can optimize the parameters of existing models. The modified new models are added to the benchmark algorithm library by the model management function, realizing the rapid construction of radar imaging intelligent target recognition benchmark algorithms, which is of great significance for the near real-time processing of remote sensing images.

[0065] The present invention has the following advantages:

[0066] (1) This invention enables the rapid construction of a radar imaging intelligent target recognition benchmark algorithm in complex background environments and usage scenarios;

[0067] (2) The modules proposed by the latest technology are written or adapted according to the standard format of neural network modules, which realizes the calling of the benchmark algorithm model for intelligent target recognition and the improvement of algorithm performance;

[0068] (3) Modifying the functional modules, width, depth and other structures of the intelligent target recognition network based on the construction process of the benchmark algorithm can reduce the number of model parameters and improve the recognition speed while ensuring the accuracy of intelligent target recognition in radar imaging.

[0069] (4) It enables the rapid construction of target recognition benchmark algorithms in multiple operating systems, multiple frameworks, and multiple development environments. Attached Figure Description

[0070] Figure 1 A block diagram illustrating the modular composition of a rapid construction method for a radar imaging intelligent target recognition benchmark algorithm;

[0071] Figure 2 A schematic diagram of the benchmark algorithm management process for a rapid construction method of a benchmark algorithm for intelligent target recognition in radar imaging;

[0072] Figure 3 A flowchart illustrating a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging;

[0073] Figure 4a A ship image from a SAR test dataset used to construct a benchmark algorithm for intelligent target recognition in radar imaging;

[0074] Figure 4b b-ship image of SAR test dataset for a rapid construction method of a benchmark algorithm for intelligent target recognition in radar imaging;

[0075] Figure 4c A ship image from a SAR test dataset for a rapid construction method of a benchmark algorithm for intelligent target recognition in radar imaging;

[0076] Figure 4d A d-ship image of a SAR test dataset for a rapid construction method of a benchmark algorithm for intelligent target recognition in radar imaging;

[0077] Figure 5 The network structure diagram of the YOLOv5 intelligent target recognition benchmark algorithm, which is a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging;

[0078] Figure 6a The first recognition result of ship images on a SAR test dataset is shown as a rapid construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0079] Figure 6b The second recognition effect of ship images on the SAR test dataset is shown as a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0080] Figure 6cA third recognition effect diagram of ship images on a SAR test dataset for a rapid construction method of a benchmark algorithm for intelligent target recognition in radar imaging;

[0081] Figure 6d The fourth recognition effect of ship images on the SAR test dataset is shown as a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0082] Figure 7a The first ship image recognition result is shown on the SAR test dataset, which is a fast construction method for a benchmark algorithm of intelligent target recognition in radar imaging.

[0083] Figure 7b The image shows the recognition effect of a second ship image on a SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0084] Figure 7c The image shows the recognition effect of a third ship image on a SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0085] Figure 7d The fourth ship image recognition effect diagram of the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging;

[0086] Figure 7e The image shows the recognition effect of the fifth ship image on the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0087] Figure 7f The image shows the recognition effect of the sixth ship image on the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0088] Figure 7g The image shows the recognition effect of the seventh ship in the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0089] Figure 7h The image shows the recognition effect of the eighth ship image on the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0090] Figure 7i The image shows the recognition effect of the ninth ship image on the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0091] Figure 7j The image shows the recognition effect of the tenth ship image on the SAR test dataset, which is a fast construction method for a benchmark algorithm for intelligent target recognition in radar imaging.

[0092] Figure 8 A flowchart illustrating the development and modification of the YOLOv5 network architecture in a radar imaging intelligent target recognition benchmark algorithm, which is part of a rapid construction method for an intelligent target recognition benchmark algorithm.

[0093] Figure 9 A schematic diagram of the YAML model architecture file of YOLOv5, a benchmark algorithm for intelligent target recognition in radar imaging, which is a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging.

[0094] Figure 10 The network structure diagram of YOLOv5, an improved intelligent target recognition benchmark algorithm, is shown as an example of a rapid construction method for an intelligent target recognition benchmark algorithm in radar imaging.

[0095] Figure 11 This image shows the recognition results of a subset of vehicle images from a SAR test dataset, representing a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging. Detailed Implementation

[0096] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0097] Example 1

[0098] like Figures 1-3 As shown in Figure 8, a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging includes the following steps:

[0099] (1) Figure 1 This is a diagram showing the module composition of this embodiment. Figure 2 This is a schematic diagram of the baseline algorithm management process in this embodiment. Figure 3 This is a flowchart of the embodiment, illustrating the rapid construction process of the radar imaging intelligent target recognition benchmark algorithm based on four parts: programming environment management module, runtime environment management module, benchmark algorithm management module, and process control module.

[0100] (2) Using the process control module, a SAR test dataset for a specific scene to be identified by the intelligent target recognition algorithm is generated according to the radar imaging test requirements. Partial images of the dataset are shown below. Figures 4a-4d As shown. The SAR test dataset for this scenario contains 3000 target images, identifying a port scene and five target types: transport ships, fishing boats, trucks, and islands / reefs. The image resolutions in the database are 1m and 1.2m, etc. The training, validation, and test sets have sizes of 2000, 500, and 500 images, respectively.

[0101] (3) Based on the target recognition requirements of the aforementioned SAR test dataset, the YOLOv5 intelligent target recognition benchmark algorithm was selected from six typical image target recognition algorithms in the benchmark algorithm management module. YOLOv5 employs an FPN+PAN structure to efficiently utilize feature information at different scales and is easy to deploy and use on various platforms. The YOLOv5 network structure diagram is shown below. Figure 5 As shown.

[0102] (4) The runtime environment management module provides different configurations of algorithm runtime environments. Since different intelligent target recognition benchmark algorithms are developed based on different vendors or different versions of deep learning frameworks, according to the environment requirements of the selected YOLOv5 benchmark algorithm, select Anaconda in the runtime environment management module to configure the PyTorch 1.12.0 deep learning framework and runtime environment for the YOLOv5 benchmark algorithm.

[0103] (5) The programming environment management module mainly consists of different programming platforms. Typical algorithms are developed and run on different programming platforms according to different programming languages. Based on the selected YOLOv5+PyTorch1.12.0 runtime environment, select the PyCharm2022.1 integrated development environment in the programming environment management module.

[0104] The current benchmark algorithm for intelligent target recognition in radar imaging is YOLOv5, the programming environment is PyCharm2022.1, and the runtime environment is Python 3.9, CUDA 11.6, Torch 1.12.0, TorchVision 0.13.0, and an RTX 3080 graphics card.

[0105] (6) After completing the programming and runtime environment configuration, the process control module configures the algorithm parameters and SAR test dataset parameters according to the instructions, and runs the benchmark algorithm. After the YOLOv5 network has undergone 300 iterations of training, the optimal network weight file is saved. The Detect network is used in conjunction with the optimal network weight file to perform inference on the test set and validation set to obtain the location, category, and recognition accuracy of the radar imaging target. The recognition effect of the YOLOv5 network on the above radar imaging targets is as follows: Figures 6a-6d As shown in the figures above, the YOLOv5 network demonstrates excellent target recognition performance in radar imaging and can be used in radar imaging target recognition systems.

[0106] The process control module and the benchmark algorithm management module can fine-tune the parameters of the existing model, setting the depth_multiple of the YOLOv5 network to 0.30 and the width_multiple to 0.20. This is done to reduce the number of model parameters and improve the recognition speed while maintaining the accuracy of radar imaging target recognition, thus lowering the requirements for the hardware environment. The modified new model is added to the benchmark algorithm library by the model management function, enabling the rapid construction of benchmark algorithms for intelligent target recognition in radar imaging.

[0107] Example 2

[0108] like Figures 1-3 As shown in Figure 8, a method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging includes the following steps:

[0109] (1) Using the process control module, a SAR test dataset for a specific scene to be identified by the intelligent target recognition algorithm is generated according to the test requirements of radar imaging. Part of the dataset includes images such as... Figures 7a-7j As shown. The SAR test dataset for this scenario contains 4000 target images, identifies a plain scene, and identifies five target types: aircraft, bulldozers, vehicles, buildings, and oil tankers. The image resolution in the database ranges from 0.8m to 5m. The training set, validation set, and test set sizes are 3000, 500, and 500 images, respectively.

[0110] (2) Based on the target recognition requirements of the aforementioned SAR test dataset, YOLOv5 was selected from the six typical image target recognition algorithms in the benchmark algorithm management module, and a smart target recognition benchmark algorithm was constructed by modifying this algorithm. A newly written CBAM module was added to the YOLOv5 YAML model document and inserted into the network's Neck section. The CBAM module as a whole adopts a cascaded channel attention module and a spatial attention module, introducing the CBAM attention mechanism to enhance the spatial information of the features. The flowchart of the YOLOv5 modification is as follows: Figure 8 As shown. The model architecture of the YOLOv5 algorithm is written and saved in YAML format, such as... Figure 9 As shown, by parsing the YOLOv5 YAML document, modifications to the model and the development of new object recognition algorithm models can be quickly achieved. The improved YOLOv5 network structure diagram is shown below. Figure 10 As shown.

[0111] (3) The runtime environment management module provides different configurations of algorithm runtime environments. Since different intelligent target recognition benchmark algorithms are developed based on different vendors or different versions of deep learning frameworks, according to the environment requirements of the selected improved YOLOv5 benchmark algorithm, Anaconda is selected in the runtime environment management module to configure the PyTorch 1.12.0 deep learning framework and runtime environment for the improved YOLOv5 benchmark algorithm.

[0112] (4) The programming environment management module mainly consists of different programming platforms. Typical algorithms are developed and run on different programming platforms according to different programming languages. Based on the selected improved YOLOv5+PyTorch1.12.0 runtime environment, the PyCharm2022.1 integrated development environment is selected in the programming environment management module.

[0113] The current benchmark algorithm for intelligent target recognition in radar imaging is an improved YOLOv5, with PyCharm 2022.1 as the programming environment and Python 3.9, CUDA 11.6, Torch 1.12.0, TorchVision 0.13.0, and an RTX 3080 graphics card as the runtime environment.

[0114] (5) After completing the programming and runtime environment configuration, the process control module configures the algorithm parameters and SAR test dataset parameters according to the instructions, and runs the benchmark algorithm. After the improved YOLOv5 network has undergone 300 iterations of training, the optimal network weight file is saved. The Detect network is used in conjunction with the optimal network weight file to perform inference on the test set and validation set to obtain the location, category, and recognition accuracy of the radar imaging target. The recognition effect of the improved YOLOv5 network on the above radar imaging targets is as follows: Figure 11 As shown in the figures above, the improved YOLOv5 network demonstrates excellent target recognition performance in radar imaging and can be used in radar imaging target recognition systems.

[0115] The process control module and the benchmark algorithm management module can fine-tune the parameters of the existing model. For example, the depth_multiple and width_multiple of the improved YOLOv5 network are set to 0.25 and 0.20, respectively. This is done to reduce the number of model parameters and improve recognition speed while maintaining the accuracy of radar imaging target recognition, thus lowering the requirements for the hardware environment. The modified new model is added to the benchmark algorithm library through the model management function, enabling the rapid construction of benchmark algorithms for intelligent target recognition in radar imaging.

[0116] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for rapidly constructing a benchmark algorithm for intelligent target recognition in radar imaging, characterized in that: Includes the following steps: S1. The user selects the background environment and target of the recognition scene in the process control module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system according to the background environment and target complexity of the SAR / ISAR image recognition scene. The process control module calls the SAR / ISAR image sample database to generate a specific SAR / ISAR experimental test dataset for the intelligent target recognition algorithm. S2. Based on the target recognition requirements of the specific SAR / ISAR experimental test dataset, the user selects at least one intelligent target recognition benchmark algorithm from the benchmark algorithm management module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system or builds and develops a new intelligent target recognition benchmark algorithm on their own, and imports, deletes, modifies, builds, and iteratively upgrades the selected algorithm model according to the usage requirements. S3. The runtime environment management module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system provides an algorithm runtime environment for the selected intelligent target recognition benchmark algorithm. The algorithm runtime environment includes a runtime programming environment and a deep learning framework. S4. The programming environment management module of the radar imaging intelligent target recognition benchmark algorithm rapid construction system provides an algorithm integration development environment for the intelligent target recognition benchmark algorithm selected in step S2. The algorithm integration development environment includes a programming environment and a programming platform. The programming environment management module calls the corresponding programming platform according to the programming environment required to construct the intelligent target recognition benchmark algorithm. S5. The process control module configures parameters and controls the running process of the intelligent target recognition benchmark algorithm selected in step S2. After the selected intelligent target recognition benchmark algorithm completes the programming environment and runtime environment configuration, the process control module configures the algorithm training parameters and SAR / ISAR test dataset parameters according to the instructions, and runs the intelligent target recognition benchmark algorithm after calling the specific SAR / ISAR experimental test dataset; The process control module and the benchmark algorithm management module can optimize the parameters of the intelligent target recognition benchmark algorithm and add the modified new model to the benchmark algorithm library by the model management function of the benchmark algorithm management module, thus completing a method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm; The radar imaging intelligent target recognition benchmark algorithm rapid construction system includes a programming environment management module, a runtime environment management module, a benchmark algorithm management module, and a process control module electrically connected to the programming environment management module, the runtime environment management module, and the benchmark algorithm management module. The benchmark algorithm management module includes a benchmark algorithm library and an algorithm management module. The benchmark algorithm library contains at least one intelligent target recognition benchmark algorithm. The benchmark algorithm library can also support the construction of self-developed benchmark recognition algorithm models. The algorithm management module manages the selected algorithm models according to user needs. The programming environment management module provides an integrated development environment for intelligent target recognition benchmark algorithms based on the user's selected operating environment. The process control module constructs a specific SAR / ISAR experimental test dataset that meets the application scenario requirements according to the radar imaging type, background environment and target recognition requirements selected by the user, configures the parameters and controls the running process of the intelligent target recognition algorithm, configures the algorithm training parameters and SAR / ISAR test dataset parameters according to the instructions, and runs the intelligent target recognition benchmark algorithm. The intelligent target recognition benchmark algorithms built into the benchmark algorithm management module include: Fast R-CNN, Faster R-CNN, YOLOv3, YOLOv5, YOLOv7, and YOLOv8; Fast R-CNN is an end-to-end target recognition model that uses RoI pooling layers to extract all RoI features in a single forward propagation. Faster R-CNN is a target recognition model that uses a proposed region network structure instead of the candidate region operation in Fast R-CNN. YOLOv3 is a target recognition model that improves the overall accuracy of SAR images and the recognition accuracy of small targets by incorporating multi-scale prediction. The basic network of YOLOv3 is Darknet-53. YOLOv5 is a target recognition model that uses an FPN+PAN structure, efficiently utilizes feature information at different scales, and is easy to deploy and use on various platforms. YOLOv7 is a target recognition model that uses a coarse-to-fine guided label assignment method, reduces the number of parameters, and improves recognition speed. YOLOv8 is a target recognition model that supports image classification, detection, and segmentation tasks using a more lightweight network structure and a new loss function.

2. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 1, characterized in that: In step S1, the background environment of the identified scene includes: desert, ocean, mountains, plains, and port; The SAR / ISAR image sample database contains no fewer than 5,000 target images, no fewer than 5 background environments, no fewer than 10 target types, and no fewer than 6 image resolutions. The target types include any one or more of the following: aircraft, bulldozers, vehicles, buildings, oil tankers, transport ships, fishing boats, ports, trucks, and islands / reefs. The image resolutions include: 0.3m, 0.5m, 0.8m, 1m, 1.2m, and 3m; The specific SAR / ISAR experimental test dataset includes a training set, a validation set, and a test set.

3. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 1, characterized in that: Step S2 includes: S21. Based on the target recognition requirements of the specific SAR / ISAR experimental test dataset, use the built-in intelligent target recognition basic algorithm in the benchmark algorithm library of the benchmark algorithm management module, and proceed to step S22. The user builds and develops the benchmark algorithm model and proceeds to step S23. S22. The user selects one or more intelligent target recognition benchmark algorithms from the benchmark algorithm library and proceeds to step S24. S23. The benchmark algorithm management module includes a built-in neural network module and a latest module written or adapted according to the standard format of the neural network module. The built-in neural network module and the latest module are combined using a YAML document, and the input and output parameters are marked. By parsing the YAML document, the model architecture, depth and width are modified and developed to obtain a new intelligent target recognition benchmark algorithm, which is then added to the benchmark algorithm library by the model management function of the benchmark algorithm management module. The new intelligent target recognition benchmark algorithm participates in the rapid construction of radar imaging intelligent target recognition benchmark algorithms. S24. Users can import, delete, modify, and iteratively upgrade intelligent target recognition benchmark algorithms through the benchmark algorithm management module.

4. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 3, characterized in that: In step S23, the built-in neural network module in the benchmark algorithm management module includes: Conv, C3, SPPF, Concat, Detect, Upsample, and Focus; The latest modules include the CBAM module, DRM module, and CFN module.

5. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 3, characterized in that: In step S23, the built-in neural network module is the neural network module of the YOLOv5 target recognition algorithm. A newly written CBAM module is added to the YOLOv5 YAML model document and inserted into the Neck part of the neural network module. The CBAM module as a whole adopts a cascaded channel attention module and a spatial attention module, and introduces the CBAM attention mechanism to enhance the spatial information of the features.

6. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 1, characterized in that: In step S3, the programming environment includes Anaconda, MSVC, MinGW, and Matlab; the deep learning framework includes TensorFlow, PaddlePaddle, PyTorch, and TensorRT. Anaconda is a cross-platform Python environment management tool that includes runtime environments for six typical target recognition algorithms. It allows users to create, configure, and manage runtime environments for target recognition algorithms composed of different deep learning frameworks and Python libraries via commands. The MSVC runtime environment is part of the Visual Studio integrated development environment. By selecting the corresponding MSVC runtime environment and Visual Studio through commands, the C / C++ inference and deployment development of typical image target recognition algorithms on the Windows platform can be implemented. MinGw is a cross-platform C / C++ environment management tool. MinGw includes Windows-specific header files and a collection of import libraries using the GNU toolset. By selecting the runtime environment through commands, it enables the development, deployment, and inference of typical image target recognition algorithms in C / C++ on multiple system platforms. The Deep Learning Toolbox is a toolbox within the Matlab development environment. It allows users to design, implement, and run target recognition algorithms by selecting the deep neural network framework corresponding to the Matlab version they are running on using commands.

7. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 1, characterized in that: In step S4, the programming environment includes: Python, C++, and Matlab; The programming platforms include: PyCharm, Microsoft Visual Studio, VS Code, Clion, and MATLAB.

8. The method for rapidly constructing a radar imaging intelligent target recognition benchmark algorithm according to claim 7, characterized in that: PyCharm is a Python integrated development environment and is used to develop target recognition algorithms based on the PyTorch, TensorFlow, and PaddlePaddle deep learning frameworks. Microsoft Visual Studio is a Windows-based integrated development environment for C / C++. VS Code and CLion are cross-platform compatible, supporting Windows, macOS, and Linux operating systems. VS Code supports multiple programming languages, while CLion is a cross-platform IDE designed in C and C++. CLion uses the TensorRT inference framework to infer and deploy object recognition algorithms. MATLAB uses its built-in deep learning toolbox to develop and implement deep learning tasks within the MATLAB environment.

Images