Synthetic aperture radar image target recognition method and system based on lightweight neural network

Abstract

The application discloses a synthetic aperture radar (SAR) image target recognition method and system based on a lightweight neural network, and the method comprises the following steps: constructing a lightweight encoder-decoder network, wherein each layer of the encoder comprises two SCAF-DW residual blocks, and each layer of the decoder comprises one SCAF-DW residual block; acquiring a SAR amplitude image and inputting the SAR amplitude image into the lightweight encoder-decoder network; using an asymmetric topological encoder to extract a deep semantic feature map; using the decoder to fuse multi-scale features from the encoder through a parameter-free jump connection to generate a final feature map; and inputting the final feature map into a classification head to output a class label with the maximum probability. The application overcomes the problems of structural redundancy, large parameters and difficulty in efficient deployment of ResUNet architecture in the prior art when the ResUNet architecture is used for SAR target recognition.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing image processing technology, specifically to a method and system for target recognition in synthetic aperture radar images based on lightweight neural networks. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave remote sensing imaging technology, has become an indispensable core data source in Earth observation systems and national defense security fields due to its unique advantages of penetrating clouds and being unaffected by day / night cycles or inclement weather. With the rapid development of hardware technology, the spatial resolution of SAR images has entered the sub-meter era, making it possible to accurately identify and classify ground targets, and greatly promoting the research boom of SAR image intelligent interpretation technology based on deep learning.

[0003] Currently, deep learning-based encoder-decoder architectures (such as ResUNet) are widely used due to their powerful feature fusion capabilities. However, their structure has inherent design flaws, resulting in a large number of model parameters and computational overhead, making them difficult to apply directly to resource-constrained platforms. Specifically, ResUNet-like architectures exhibit structural redundancy when applied to SAR image target recognition; their skip connections and symmetric encoding / decoding design lead to significant computational and parameter costs. Furthermore, common post-processing compression methods (such as pruning and quantization) often struggle to effectively optimize these models while maintaining accuracy. Therefore, existing technologies cannot meet the requirements of lightweight deployment while maintaining recognition accuracy, thus hindering the application of SAR recognition technology in real-world scenarios. Summary of the Invention

[0004] To overcome the problems of structural redundancy, large parameters, and difficulty in efficient deployment of existing ResUNet-like architectures for SAR target recognition, this invention provides a lightweight neural network named LW-ResUNet and its method.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect is a synthetic aperture radar image target recognition method based on lightweight neural networks, which includes the following steps: S1. Construct a lightweight encoder-decoder network, where each layer of the encoder contains two SCAF-DW residual blocks and each layer of the decoder contains one SCAF-DW residual block; acquire SAR amplitude images and input them into the lightweight encoder-decoder network. S2. Use an encoder with asymmetric topology to extract deep semantic feature maps; S3. Use the decoder to fuse multi-scale features from the encoder through parameterless skip connections to generate the final feature map; S4. Input the final feature map into the classification head and output the class label with the highest probability.

[0006] Preferably, in S1, the SCAF-DW residual block includes: convolving the input points of the SAR amplitude image X to expand the channels, generating expanded features; and using depthwise convolutional spatial filtering to process the expanded features. Spatial features are obtained through depthwise separable convolution; lightweight adaptive feature calibration is then performed on the spatial features.

[0007] As a preferred option, lightweight adaptive feature calibration includes: Global average pooling is used to compress the input features to generate channel statistical descriptors; Channel statistical descriptors are used to generate channel weights via a fully connected layer with a channel compression ratio of r and a sigmoid function. Based on channel weights, use learnable scalars Weighted features are obtained by weighted fusion of spatial features; After performing point convolution projection on the weighted features, 1×1 convolution is used to compress the number of channels to the target dimension; after regularizing the compressed features using Dropout, a shortcut path for residual connections is constructed. Add the output of the main path and the shortcut path and activate them.

[0008] Preferably, S1 further includes: converting a single-channel SAR amplitude image into a three-channel pseudo-color input tensor via channel copying and then inputting it into a lightweight encoder-decoder network.

[0009] Preferably, S2 includes: S21. First layer encoding: After the SAR amplitude image passes through the initial convolution module, it enters the first layer of the encoder for downsampling. S22, Second layer encoding: The features output from the first layer of the encoder are input into the second layer of the encoder for downsampling and feature transformation. S23, Third layer encoding: The features output from the second layer of the encoder are input into the third layer of the encoder. After processing by two SCAF-DW residual blocks, a deep semantic feature map is output.

[0010] Preferably, S3 includes: S31. The features output from the third layer of the encoder are directly input into the third layer of the decoder; the decoder consists of a transposed convolution and a SCAF-DW residual block. S32. Add the features output by the third layer of the decoder (naturally aligned) to the features output by the second layer of the encoder element by element and then feed them into the second layer of the decoder. S33. The features output from the second layer of the naturally aligned decoder are added element-wise to the features output from the first layer of the encoder and then input into the first layer of the decoder to generate the final feature map.

[0011] Preferably, S4 includes: The spatial dimension is compressed using a global average pooling layer. The channel feature vectors are obtained, and then passed through a fully connected layer and The function yields the category probability distribution. , The output is the category label with the highest probability.

[0012] Secondly, a synthetic aperture radar image target recognition system based on lightweight neural networks includes: A lightweight encoder-decoder network is constructed, wherein each layer of the encoder contains two SCAF-DW residual blocks, and each layer of the decoder contains one SCAF-DW residual block. SAR amplitude images are acquired and input into the lightweight encoder-decoder network. Deep semantic feature maps are extracted using an encoder with asymmetric topology. Multi-scale features from the encoder are fused using the decoder through parameterless skip connections to generate the final feature map. The classification and output module is used to input the final feature map into the classification head and output the class label with the highest probability. The synthetic aperture radar image target recognition system based on a lightweight neural network is used to implement the synthetic aperture radar image target recognition method and its steps based on a lightweight neural network as described in the first aspect.

[0013] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. The lightweight architecture proposed in this invention maintains high recognition performance while compressing the number of model parameters to an extremely low level. On the FAIR-21 dataset (publicly available on GitHub, containing only images and annotations) built by the inventors, the model achieved an overall classification accuracy of 73.2% using only 0.90M parameters, outperforming many mainstream lightweight models. This achievement effectively demonstrates that through architectural innovation, it is possible to maintain excellent discriminative ability while compressing the model size to an extremely low level, providing a practical technical path for building "small but precise" SAR recognition models.

[0014] 2. Provides a reproducible lightweight technology paradigm. The complex lightweight objective is decomposed into three modular design dimensions: asymmetric topology, parameter-free skip connections, and adaptive feature calibration, enabling component-based configuration of lightweight strategies. This technical approach provides a clear and transferable design reference for lightweight transformation of encoder-decoder architectures in resource-constrained tasks, lowering the development threshold for high-performance lightweight models.

[0015] 3. It possesses the potential for direct deployment in edge computing scenarios, lowering the technical application threshold. The model of this invention not only has extremely low parameter count (<1M), but its computational complexity (Floating Point Operations, or FLOPs) has also been carefully optimized, falling within the typical computing power range of embedded platforms. This characteristic allows it to be directly deployed in spaceborne processors, UAV onboard computing units, or portable reconnaissance equipment with strict limitations on power consumption, memory, and computing resources, without the need for complex secondary compression or adaptation, enabling real-time SAR image target recognition and greatly promoting the practical application of the technology. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall architecture of the lightweight neural network (LW-ResUNet) of this invention; Figure 2 This is a flowchart illustrating the structure of the depth-separable residual block (SCAF-DW) with embedded SCAF module in this invention. Figure 3 This is a schematic diagram of the internal structure of the lightweight adaptive feature calibration module (SCAF) in this invention; Figure 4 This is a schematic diagram of a portion of the tank image samples from the FAIR-21 dataset used in this embodiment of the invention; Figure 5 This is a schematic diagram of a portion of the tower crane images from the FAIR-21 dataset used in this embodiment of the invention; Figure 6 This is a schematic diagram of a portion of the Boeing 747 samples from the FAIR-21 dataset used in this embodiment of the invention. Detailed Implementation

[0017] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0018] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0019] The core improvements of this invention's lightweight neural network (LW-ResUNet) are reflected in three aspects: (1) An asymmetric computational topology is adopted to concentrate computational resources on the encoder to enhance feature extraction.

[0020] (2) A parameterless cross-branch fusion strategy is adopted, which eliminates the parameters in the skip connection.

[0021] (3) Depth-separable residual blocks are embedded in the SCAF attention module to form a lightweight adaptive feature extraction unit. These designs together achieve a lightweight reconstruction of the encoder-decoder architecture.

[0022] Experimental data description: The SAR target recognition dataset (named FAIR-21) used in Embodiment 1 of this invention was created by the inventors. This dataset is publicly available on GitHub (URL: https: / / github.com / angelandeagle / LW-ResUnet-dataset), and contains only SAR amplitude images and their corresponding target category labels; it does not include the structure, parameters, or training code of the lightweight neural network described in this invention. The public can reproduce the experimental process of this invention using this publicly available dataset. Example 1: like Figure 1 The synthetic aperture radar image target recognition method based on lightweight neural networks shown includes the following steps: S1. Construct a lightweight encoder-decoder network; each layer of the encoder in the lightweight encoder-decoder network contains two SCAF-DW residual blocks, and each layer of the decoder contains one SCAF-DW residual block; obtain the SAR amplitude image and input it into the lightweight encoder-decoder network, such as... Figure 1 As shown; S11. Network Architecture Construction. A lightweight encoder-decoder network (LW-ResUNet) is constructed, with its backbone symmetrically composed of three encoder layers and three decoder layers. The encoder is responsible for feature extraction and downsampling, while the decoder is responsible for feature fusion and upsampling. Each encoder layer contains two SCAF-DW residual blocks, and each decoder layer contains one SCAF-DW residual block, as detailed below. Figure 2 As shown, the SCAF-DW residual block is implemented as follows: The SCAF-DW residual block is the basic unit that constitutes the encoder and decoder. Now let Let be the set of real numbers. and The module has 10 input channels, which are the height and width of the input feature map. The number of output channels is The number of channels in the intermediate hidden layer is In this module = That is, the first one Convolution turns the channels from Transform to Depth-wise convolution keeps the number of channels constant, the second one... Convolutional Preservation The internal processing flow is as follows: Let the module input feature tensor be represented as: Input this feature tensor into point convolution to expand the channels and generate expanded features. : .

[0023] The expanded features are processed using depthwise convolutional spatial filtering. Depthwise separable convolutions are used to efficiently extract spatial features and obtain feature values. : in Group size Step length .

[0024] Lightweight Adaptive Feature Calibration (SCAF) is performed on the spatial features. This step adaptively enhances the features, such as... Figure 3 As shown. Specifically: 1. Compression: Channel statistical descriptors are generated through global average pooling. : 2. Excitation: Channel weights are generated through a fully connected layer with a channel compression ratio r and a sigmoid function. : , It is the channel compression ratio.

[0025] 3. Adaptive Fusion: Using Learnable Scalars Weighted features are obtained by weighted fusion. in, This indicates channel-by-channel multiplication. Initialize it to 0.5 and update the module during training as a residual branch embedded in the main convolution path.

[0026] Pointwise convolution projection is applied to the weighted features, and 1×1 convolution is used to compress the number of channels to the target dimension. ,generate in For the second The weight matrix is ​​used to adjust the number of channels from... Project to .

[0027] right Using Dropout regularization, a Dropout layer is introduced to prevent overfitting and generate... : Construct a shortcut path for residual joins and generate : .

[0028] Residual fusion and output: The outputs of the main path and the shortcut path are added together and activated to obtain the features. : , , S12. Input data standardization. The single-channel SAR amplitude image is converted into a three-channel pseudo-color input tensor by channel duplication. In this embodiment, it is set This tensor will serve as the input to the network.

[0029] S2. Feature extraction of encoder path (asymmetric topology) S21, First-level encoding. Input After the initial convolutional module, the code enters the first layer of the encoder. This layer consists of two cascaded SCAF-DW residual blocks, where the first block performs a 2x downsampling and outputs a feature map. .

[0030] S22, Second layer encoding. The second layer of the feature input encoder. This layer also consists of two SCAF-DW residual blocks, performing downsampling and feature transformation to output a feature map. .

[0031] S23, Third-level coding. Features The input is the third layer of the encoder. After processing by two SCAF-DW residual blocks, the output is a deep semantic feature map. .

[0032] S3, Feature fusion of the decoder path (parameterless cross-branch connection, PF-CBC): The decoder fuses multi-scale features from the encoder through parameterless skip connections.

[0033] S31, Top-level decoding and direct connection. Encoder top-level features. It is directly used as the input to the third layer of the decoder. D3 consists of a transposed convolution (achieving 2x upsampling) and a SCAF-DW residual block, outputting features. .

[0034] S32, Mid-layer decoding and first cross-layer fusion. Perform parameterless cross-branch fusion: combine decoder features. With encoder features Perform element-wise addition. This operation is feasible only if the overall network design ensures sufficient channel numbers for both methods. With space size Natural alignment. The blending formula is: The FC2 feature is input into the second layer of the decoder to generate decoder features. S33: Shallow decoding and second cross-layer fusion. Perform parameterless cross-branch fusion again: combine decoder features. With encoder features Perform element-wise addition. The fusion formula is: Features after fusion The input is processed by the first layer of the decoder, and the final high-resolution feature map is output. .

[0035] S4. Classification Decision Output. This involves processing the feature map output from the decoder. The data is fed into the classification head. First, the spatial dimension is compressed using a global average pooling layer. The channel feature vectors are obtained, and then passed through a fully connected layer and The function yields the category probability distribution. , The final output is the category label with the highest probability.

[0036] Table 1: Experimental comparison of LW-ResUNet with several representative lightweight architectures and classic networks.

[0037] As shown in Table 1, the proposed LW-ResUNet model achieves the best overall accuracy of 73.21%, surpassing all comparison models. Its accuracy is 2.70 percentage points higher than MobileNetV3. Notably, despite ResNet18 having 12.5 times more parameters than LW-ResUNet, LW-ResUNet still maintains a 1.56 percentage point advantage. In terms of the macroscopic average F1 score, reflecting cross-class balance performance, LW-ResUNet also performs best (0.5320). Table 2 shows the specific recognition accuracy metrics of the LW-ResUNet network for each target class.

[0038] Table 2: LW-RESUNET Category Reports

[0039] Figure 4 This is a schematic diagram of a partial sample of tank images from the FAIR-21 dataset used. Figure 5 This is a schematic diagram of a portion of the tower crane images used in the FAIR-21 dataset; Figure 6 This is a schematic diagram of a portion of the Boeing 747 samples used in the FAIR-21 dataset.

[0040] Table 3: Comparison of parameter-free cross-branch connections and standard skip connections

[0041] Table 3 compares the proposed PF-CBC with traditional skip connections using 1×1 convolutions. Experiments were conducted based on the LW-ResUNet architecture (including asymmetric topology and SCAF modules), testing two different skip connection implementations: the traditional "standard skip + 1×1 convolution" and the proposed PF-CBC. The PF-CBC strategy improves performance on all classification metrics, increasing overall accuracy from 0.7128 to 0.7321. This demonstrates that element-level addition achieved through cross-layer channel matching is more effective at preserving feature size than convolutionally aligned connections. Eliminating 1×1 convolutions not only reduces the number of parameters but also avoids information loss that may result from channel projection, thereby enhancing discriminative power.

[0042] Table 4: Comparison of SCAF module ablation charts

[0043] Table 4 compares the performance of the no-attention module (wo-SE), the standard compressed stimulus module (SE), and the proposed SCAF module on the LW-ResUNet infrastructure (including asymmetric topology and PF-CBC strategy). The baseline (wo-SE) OA is 72.03%. After introducing the standard SE module, the OA is still 72.02%, showing no improvement and even a slight decrease in other metrics. This indicates that directly applying the classic channel attention mechanism in a highly lightweight network may have limited benefits due to adaptability issues. The SCAF module, however, introduces learnable adaptive fusion coefficients. It achieves smooth dynamic fusion of original features and channel calibration features, avoiding simple feature rescaling, and improves OA to 73.21%, verifying its greater flexibility in lightweight networks and its adaptability to complex features of SAR images.

[0044] Table 5: Ablation Study of Encoder / Decoder Residual Configuration

[0045] The core structural design of this paper lies in the asymmetric distribution of the number of residual blocks between the encoder and decoder stages. Specifically, the network consists of three encoder stages and three decoder stages. Each encoder stage stacks two residual blocks sequentially (the first one performs downsampling), while each corresponding decoder stage uses only one residual block (preceded by an upsampling layer). This results in the total number of residual blocks in the encoder path for feature extraction (6 blocks) being twice that in the decoder path for feature fusion (3 blocks), a configuration we refer to as (2:1). To verify the effectiveness of this design, we compare it with a symmetric configuration (3 encoder residual blocks: 3 decoder residual blocks) and an inverse asymmetric configuration (1:2) through ablation experiments. Quantitative comparison results are detailed in Table 5.

[0046] This embodiment, through the above steps, completes an efficient and lightweight identification process from SAR image input to target category output. The embodiments described above merely illustrate the implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A synthetic aperture radar image target recognition method based on lightweight neural networks, characterized in that, Includes the following steps: S1. Construct a lightweight encoder-decoder network, where each layer of the encoder contains two SCAF-DW residual blocks and each layer of the decoder contains one SCAF-DW residual block; acquire SAR amplitude images and input them into the lightweight encoder-decoder network. S2. Use an encoder with asymmetric topology to extract deep semantic feature maps; S3. Use the decoder to fuse multi-scale features from the encoder through parameterless skip connections to generate the final feature map; S4. Input the final feature map into the classification head and output the class label with the highest probability.

2. The synthetic aperture radar image target recognition method based on a lightweight neural network according to claim 1, characterized in that, In S1, the SCAF-DW residual block includes: convolving the input points of the SAR amplitude image X to expand the channels, generating expanded features; and using depthwise convolutional spatial filtering to process the expanded features. Spatial features are obtained through depthwise separable convolution; lightweight adaptive feature calibration is then performed on the spatial features.

3. The synthetic aperture radar image target recognition method based on a lightweight neural network according to claim 2, characterized in that, Lightweight adaptive feature calibration includes: Global average pooling is used to compress the input features to generate channel statistical descriptors; Channel statistical descriptors are used to generate channel weights via a fully connected layer with a channel compression ratio of r and a sigmoid function. Based on channel weights, use learnable scalars Weighted features are obtained by weighted fusion of spatial features; After performing point convolution projection on the weighted features, 1×1 convolution is used to compress the number of channels to the target dimension; after regularizing the compressed features using Dropout, a shortcut path for residual connections is constructed. Add the output of the main path and the shortcut path and activate them.

4. The synthetic aperture radar image target recognition method based on a lightweight neural network according to claim 1, characterized in that, S1 also includes: converting a single-channel SAR amplitude image into a three-channel pseudo-color input tensor via channel copying and then inputting it into a lightweight encoder-decoder network.

5. The synthetic aperture radar image target recognition method based on lightweight neural network according to claim 1, characterized in that, S2 include: S21. First layer encoding: After the SAR amplitude image passes through the initial convolution module, it enters the first layer of the encoder for downsampling. S22, Second layer encoding: The features output from the first layer of the encoder are input into the second layer of the encoder for downsampling and feature transformation. S23, Third layer encoding: The features output from the second layer of the encoder are input into the third layer of the encoder. After processing by two SCAF-DW residual blocks, a deep semantic feature map is output.

6. The synthetic aperture radar image target recognition method based on lightweight neural network according to claim 1, characterized in that, S3 include: S31. The features output from the third layer of the encoder are directly input into the third layer of the decoder; the decoder consists of a transposed convolution and a SCAF-DW residual block. S32. Add the features output by the third layer of the decoder (naturally aligned) to the features output by the second layer of the encoder element by element and then feed them into the second layer of the decoder. S33. The features output from the second layer of the naturally aligned decoder are added element-wise to the features output from the first layer of the encoder and then input into the first layer of the decoder to generate the final feature map.

7. The synthetic aperture radar image target recognition method based on lightweight neural network according to claim 1, characterized in that, S4 include: The spatial dimension is compressed using a global average pooling layer. The channel feature vectors are obtained, and then passed through a fully connected layer and The function yields the category probability distribution. , The output is the category label with the highest probability.

8. A synthetic aperture radar image target recognition system based on a lightweight neural network, characterized in that, include: A lightweight encoder-decoder network is constructed, wherein each layer of the encoder contains two SCAF-DW residual blocks, and each layer of the decoder contains one SCAF-DW residual block. SAR amplitude images are acquired and input into the lightweight encoder-decoder network. Deep semantic feature maps are extracted using an encoder with asymmetric topology. Multi-scale features from the encoder are fused using the decoder through parameterless skip connections to generate the final feature map. The classification and output module is used to input the final feature map into the classification head and output the class label with the highest probability. The synthetic aperture radar image target recognition system based on a lightweight neural network is used to implement the synthetic aperture radar image target recognition method and its steps based on a lightweight neural network as described in claim 1.

Images