A low-complexity SAR image target detection method for embedded platforms and a readable storage medium

Abstract

The application discloses a low-complexity SAR image target detection method for an embedded platform and a readable storage medium, and at least comprises the following steps: model construction: based on a YOLOv8 architecture, improvements are made to three parts of a backbone network, feature fusion and a detection head, and a low-complexity SAR target detection network model is constructed; model deployment preparation: the low-complexity SAR target detection network model is trained, the low-complexity SAR target detection network model after training is used to detect images in a test set, a detection result is obtained, the detection accuracy and the calculation complexity of the low-complexity SAR target detection network model are calculated; the low-complexity SAR target detection network model is subjected to parameter solidification and precision quantization; model deployment and reasoning: the low-complexity SAR target detection network model after parameter solidification and precision quantization is deployed on an embedded platform for target detection reasoning. Through the three improvements of the efficient backbone network based on the inverted bottleneck, the fast feature pyramid and the decoupled detection head, the calculation complexity is significantly reduced while the feature expression capability is maintained, and the low-complexity SAR target detection network model exhibits a better accuracy-efficiency balance compared with mainstream lightweight models.

Description

Technical Field

[0001] This invention relates to the field of SAR image target detection technology, specifically to a low-complexity SAR image target detection method and a readable storage medium for embedded platforms. Background Technology

[0002] Synthetic Aperture Radar (SAR), with its all-weather, all-time imaging capabilities, has become a core remote sensing tool in fields such as military reconnaissance, marine monitoring, and disaster assessment. With the rapid development of imaging technologies from UAVs, satellites, and missile-borne platforms, the amount of SAR data is growing exponentially, making the efficient extraction of high-value information a key challenge. SAR image target detection, as one of the key technologies, is of great significance for enhancing the practical value of SAR systems.

[0003] In recent years, deep learning-based target detection methods have provided new insights for SAR image interpretation. Existing research mainly falls into two paradigms: CNN-based detection methods and Transformer-based detection methods. SAR image target detection methods based on the classic CNN detection framework have achieved great success. Most studies have significantly improved the generalization ability and accuracy of the models by improving the CNN network structure, including introducing attention mechanisms, multi-scale feature fusion, or data augmentation. Transformer-based SAR target detection methods have made significant breakthroughs in model architecture, denoising strategies, and semantic optimization, especially in the detection of small targets in complex scenes. For example, some studies have improved the performance of SAR target detection models by fusing the advantages of CNN local feature extraction with the global modeling capabilities of Transformer, suppressing speckle noise through multi-level denoising detection Transformer, and using large model techniques to design basic models for SAR image target recognition or constructing large SAR target detection datasets.

[0004] Although these methods have excellent performance, their large number of parameters and computational cost make them difficult to adapt to the needs of embedded platforms, which seriously restricts the real-time application of SAR systems in resource-constrained scenarios. Therefore, it is urgent to research and develop lightweight SAR target detection algorithms. Lightweight target detection algorithms can not only improve processing speed and timeliness, but also better adapt to the needs of embedded platforms and achieve wider application. Existing lightweight detection methods, including the YOLO series, MobileNet and ShuffleNet, have provided an excellent balance between network lightweighting and detection accuracy, but they still have the following limitations in the real-time application of SAR image interpretation: (1) The limited number of channels and layers in existing lightweight methods leads to insufficient feature map details, which weakens the cross-channel information interaction and feature extraction capabilities, especially in complex environments. (2) Existing lightweight methods focus on optimizing inference speed on GPUs, and their adaptability to embedded devices still needs to be further improved, especially when computing power and storage resources are highly limited. Summary of the Invention

[0005] The purpose of this invention is to overcome the contradiction between the computational complexity and inability to deploy high-precision SAR detection models on edge platforms, and the weak feature extraction capability and low accuracy of lightweight models on SAR images. It provides a method that maintains or even improves target detection accuracy on SAR images while significantly reducing model computational complexity and parameter count, and enables efficient, real-time deployment on embedded platforms.

[0006] To achieve the above objectives, this invention provides a low-complexity SAR image target detection method for embedded platforms, comprising at least the following steps: Model Construction: Based on the YOLOv8 architecture, improvements are made to the backbone network, feature fusion, and detection head to construct a low-complexity SAR target detection network model; Model deployment preparation: Train the low-complexity SAR target detection network model, use the trained low-complexity SAR target detection network model to detect images in the test set, obtain the detection results, calculate the detection accuracy and computational complexity of the low-complexity SAR target detection network model; solidify the parameters and quantize the accuracy of the low-complexity SAR target detection network model. Model Deployment and Inference: The low-complexity SAR target detection network model, after parameter solidification and accuracy quantization, is deployed to an embedded platform for target detection inference.

[0007] Optionally, improvements to the backbone network during model construction include: reconstructing the backbone network based on the inverted bottleneck layer, significantly reducing computational load by designing an appropriate channel expansion ratio and employing a depthwise separable convolutional STEM layer.

[0008] Optionally, the improvement of the feature fusion in model construction includes: constructing a fast feature pyramid network, using the P4 feature layer as the main fusion node, and fusing information from the P3 and P5 feature layers to reduce redundant parameters.

[0009] Optionally, improvements to the feature fusion also include employing depthwise separable convolutions to further reduce computational costs.

[0010] Optionally, improvements to the detection head during model building include: adopting a decoupled design for classification and regression tasks, processing them separately using independent pointwise convolutions, and expanding the channels according to different requirements for classification and bounding box extraction to improve the detection efficiency of the detection head.

[0011] Optionally, the model deployment preparation also includes: before training the low-complexity SAR target detection network model, dividing the dataset into a training set, a validation set, and a test set, preprocessing the input images, and then using the training set to train the low-complexity SAR target detection network model.

[0012] Optionally, the training parameters can be adjusted using the validation set.

[0013] Optionally, in the model deployment preparation, the precision quantization includes quantizing the FP32 precision model to FP16 precision.

[0014] Optionally, model deployment and inference also include: optimizing inference performance for different embedded platforms.

[0015] The present invention also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the method described above.

[0016] Compared to the prior art, the beneficial effects of the present invention include at least the following: (1) In the model construction, the backbone network based on the inverted bottleneck avoids feature degradation caused by excessive lightweighting, thus maintaining high accuracy. Traditional lightweighting sacrifices the ability to fuse features at multiple scales. This invention optimizes the feature pyramid (mainly P4 with depth separable convolution), which simplifies the path but ensures that information from P3 (high resolution, rich in small target details) and P5 can be effectively guided to the main feature layer P4. The aim is to quickly fuse features from different scales by reducing the amount of computation, thereby enhancing the detection ability of small targets and complex backgrounds, and making it more advantageous when processing SAR images in complex environments.

[0017] (2) The decoupled detection head of the present invention is specifically designed to provide fast and lightweight computation for classification and regression tasks. While maintaining the expressive power of the model, it significantly reduces the amount of computation, enabling the model to maintain good detection performance in resource-constrained environments. Attached Figure Description

[0018] Figure 1 This is a flowchart of a low-complexity SAR image target detection method for embedded platforms according to the present invention.

[0019] Figure 2 This is a network structure diagram of the improved method of the present invention.

[0020] Figure 3 This is a comparison diagram of the fast feature pyramid network structure of the present invention and the feature fusion structure of the traditional PANet architecture; wherein, (a) represents the feature fusion structure of the traditional PANet architecture; and (b) represents the fast feature pyramid of the present invention.

[0021] Figure 4 This is a visualization of the detection results of the improved method of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Currently, scholars have provided different SAR target detection methods; however, these methods do not address issues such as computational optimization for embedded devices and deployment on embedded platforms.

[0024] This invention provides a low-complexity SAR image target detection method for embedded platforms, comprising at least the following steps: Model Construction: Based on the YOLOv8 architecture, improvements are made to the backbone network, feature fusion, and detection head to construct a low-complexity SAR target detection network model; Model deployment preparation: Train the low-complexity SAR target detection network model, use the trained low-complexity SAR target detection network model to detect images in the test set, obtain the detection results, calculate the detection accuracy and computational complexity of the low-complexity SAR target detection network model; solidify the parameters and quantize the accuracy of the low-complexity SAR target detection network model. Model Deployment and Inference: The low-complexity SAR target detection network model, after parameter solidification and accuracy quantization, is deployed to an embedded platform for target detection inference.

[0025] Example This embodiment provides a low-complexity SAR image target detection method for embedded platforms, which has high detection accuracy for multiple types of targets in complex scenes, and significantly reduces the computational load and parameter count. Figure 1 As shown, it specifically includes the following steps: S1, such as Figure 2 As shown, this invention is based on a lightweight improvement design of the YOLOv8 model, with the main goal of reducing the floating-point operations (FLOPs) of the network while maintaining the detection accuracy of the model. Improvements are made to the backbone network, feature fusion, and detection head, specifically including the following steps: S1-1: The backbone network is designed based on an inverted bottleneck layer, employing a channel expansion strategy within this layer. In the STEM part, a combination of pointwise convolution and standard convolution is used to strictly limit the number of channels, efficiently reducing the input layer size from 640×640 to 160×160, thus reducing computation while stimulating initial channel information. Simultaneously, by limiting the difference in the number of input / output channels (the input and output of each P3-P5 layer), the number of hidden layer channels is ensured to remain within a reasonable range, reducing inter-layer information exchange and enabling the neural network to extract meaningful features more efficiently. This significantly reduces computational complexity while maintaining model accuracy.

[0026] S1-2, as shown Figure 3 As shown in Table 1, the feature fusion part is compared... Figure 3 (a) and Figure 3 (b) Figure 3 (b) compared to Figure 3 (a) simplifies the two connections, demonstrating that the improved feature pyramid structure of this invention simplifies the traditional PANet design by reducing the number of convolutional layers and channels, while retaining the hierarchical feature integration capability and reducing computational cost. Using P4 as the main fusion node, complementary information from P3 (fine-grained details) and P5 (high-level semantics) is efficiently aggregated through an optimized path, optimizing the information flow and minimizing intermediate computational cost, significantly reducing the feature fusion computational cost compared to the traditional PANet architecture.

[0027] Table 1 Comparison of Feature Fusion Methods S1-3, as shown in Table 2, a decoupled detection head is adopted, with the classification and regression networks using independent pointwise operations. By introducing a 1×1 convolution after the traditional convolutional layer, this model enhances the nonlinear interaction between channels and strengthens the nonlinear modeling capability within the local receptive field, thereby improving the performance of classification and detection tasks. Furthermore, the 1×1 convolution reduces the channel dimension to decrease computational overhead, and using two 3×3 convolutions (the 3×3 convolution is followed by the 1×1 convolution) effectively reduces the overall computational cost compared to a single 5×5 convolution. Compared to traditional detection heads, this achieves a nearly 50% reduction in computational cost while maintaining detection accuracy. Proceed to S2.

[0028] Table 2 Comparison of Detection Head Designs S2. Divide the dataset into training, validation, and test sets, and preprocess the input images. Specifically, SARDet-100K is the first large-scale SAR target detection dataset reaching the COCO level, jointly constructed by Nankai University and other institutions. This dataset integrates and standardizes 10 existing SAR detection datasets (such as SAR-AIRcraft, HRSID, SSDD, etc.) from different institutions in China, Europe, and the United States, ultimately forming standardized data containing 116,598 images and 245,653 labeled instances. The dataset covers six key categories: aircraft (accounting for 21% of the total instances), ships (47%), automobiles (5%), bridges (14%), oil tanks (11%), and ports (2%). These target instances are distributed in SAR images with resolutions ranging from 0.1 meters to 25 meters, covering multi-source data from satellites, airborne platforms, to ground radar, exhibiting significant scene diversity and multi-scale characteristics. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio. All images are uniformly scaled to 640×640 pixels, then normalized, and bounding boxes exceeding the image size are checked. Proceed to S3.

[0029] S3. Train the improved low-complexity SAR image target detection network using the training set, and adjust the network training parameters using the validation set to obtain the trained lightweight detection model. Specifically, the training hardware and software environment is: Windows 10 operating system, PyTorch 2.0 deep learning framework, RTX 4060 Ti graphics card with 16GB of VRAM, i5-13490k CPU, and 32GB of RAM. Stochastic gradient descent is used to optimize the model, with a batch size of 16, a learning rate of 0.01, momentum of 0.9, weight decay of 0.001, and 200 iterations. Proceed to S4.

[0030] S4, such as Figure 4As shown, the trained SAR target detection model is used to detect and identify six types of targets in the SAR image: aircraft, ships, vehicles, bridges, oil tanks, and ports, and the detection results are obtained. Proceed to S5.

[0031] S5. As shown in Table 4, the average accuracy is used to evaluate the model's performance in the test results (test set), specifically in two forms: mAP50: the average detection accuracy with a 50% Intersection over Union (IoU) threshold; mAP50-95: the average accuracy across multiple IoU thresholds within the 50%-95% range, providing a more comprehensive evaluation of the model's performance under different positioning accuracy requirements. Simultaneously, floating-point computation and the number of model parameters are used to evaluate the model's complexity. The formula for calculating the average accuracy is as follows: In object detection, mAP is calculated based on IoU (Intersection over Union) between predicted and ground truth boxes. This example uses an IoU threshold of 0.5. Here, TP represents the number of correctly detected targets in a class (IoU ≥ 0.5), FP represents false positives (IoU < 0.5 or duplicate detections), and FN represents false negatives (no real target detected). AP is short for Average Precision, and mAP is short for Mean Average Precision. P represents precision, the proportion of positive predicted samples that are actually positive, and R represents recall, the proportion of positive actually detected samples that are correctly identified. P(R) is the precision-recall function, where n represents the number of classes. In this example, the dataset contains 6 classes, so n is 6. As shown in Table 3, the method of this invention achieves the lowest computational cost and optimal mAP when the input image size is 640, and also achieves performance far exceeding MobileNetv3 when the input image size is 320, demonstrating a better accuracy-efficiency tradeoff compared to mainstream lightweight models. Proceed to S6.

[0032] Table 3 Performance comparison with other advanced lightweight methods S6. Solidify the trained network model parameters and quantize the 32-bit floating-point model to 16-bit precision. Specifically, use the deployment script provided by YOLO to first convert the trained .pt model to .onnx format, fixing the input dimension to 640×640 during the conversion. Then, convert the .onnx format to the format required by different embedded platforms. In this embodiment, deployed on a Huawei Atlas A2 embedded device (Ascend 310), the Huawei ATC tool is needed to convert the .onnx format to the dedicated .om format, and simultaneously quantize the model to FP16 precision. Proceed to S7.

[0033] S7. Deploy the solidified and quantized model to mainstream embedded platforms and optimize inference performance for different platforms. Specifically, this example uses the Huawei Atlas A2 embedded device (Ascend 310) for deployment. Its basic computing capabilities are as follows: INT8 integer operation AI computing power is 8 TOPS; FP16 half-precision floating-point operation AI computing power is 4 TFLOPS; the CPU has four 1.0GHz cores and a power consumption of 8W. The inference environment is built on the Ascend 310 using the CANN toolchain. The converted model is deployed on the Ascend 310 and invoked using ACL commands. AIPP hardware preprocessing and the ais_bench tool are enabled to optimize inference speed. Proceed to S8.

[0034] S8. As shown in Table 4, the real-time processing capability of the model at a specific image scale was tested on an embedded platform, and the inference time was recorded. Specifically, the inference time of the method of this invention and other advanced lightweight methods was tested on an Ascend 310 when the input image size was 640×640. The method of this invention achieved the optimal inference frame rate.

[0035] Table 4. Comparison of inference performance of different lightweighting methods on Ascend 310 embedded devices In summary, this invention provides a low-complexity SAR image target detection method for embedded platforms, comprising at least the following steps: Model construction: Based on the YOLOv8 architecture, improvements are made to the backbone network, feature fusion, and detection head to construct a low-complexity SAR target detection network model; Model deployment preparation: The low-complexity SAR target detection network model is trained, and the trained low-complexity SAR target detection network model is used to detect images in the test set to obtain detection results. The detection accuracy and computational complexity of the low-complexity SAR target detection network model are calculated; Parameter solidification and accuracy quantization are performed on the low-complexity SAR target detection network model; Model deployment and inference: The low-complexity SAR target detection network model after parameter solidification and accuracy quantization is deployed to an embedded platform for target detection inference. Through three major improvements—an efficient backbone network based on an inverted bottleneck, a fast feature pyramid, and a decoupled detection head—the computational complexity is significantly reduced while maintaining feature representation capabilities, exhibiting a better accuracy-efficiency balance compared to mainstream lightweight models. Deployment verification on the Huawei Atlas A2 embedded device (Ascend 310) shows that this method has practical application value.

[0036] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A low-complexity SAR image target detection method for embedded platforms, characterized in that, At least the following steps are included: Model Construction: Based on the YOLOv8 architecture, improvements are made to the backbone network, feature fusion, and detection head to construct a low-complexity SAR target detection network model; Model deployment preparation: Train the low-complexity SAR target detection network model, use the trained low-complexity SAR target detection network model to detect images in the test set, obtain the detection results, calculate the detection accuracy and computational complexity of the low-complexity SAR target detection network model; solidify the parameters and quantize the accuracy of the low-complexity SAR target detection network model. Model Deployment and Inference: The low-complexity SAR target detection network model, after parameter solidification and accuracy quantization, is deployed to an embedded platform for target detection inference.

2. The low-complexity SAR image target detection method as described in claim 1, characterized in that, In model construction, improvements to the backbone network include: reconstructing the backbone network based on the inverted bottleneck layer, significantly reducing computational load by designing an appropriate channel expansion ratio and using a depthwise separable convolutional STEM layer.

3. The low-complexity SAR image target detection method as described in claim 1, characterized in that, In model construction, improvements to the feature fusion include: constructing a fast feature pyramid network, using the P4 feature layer as the main fusion node, and fusing information from the P3 and P5 feature layers to reduce redundant parameters.

4. The low-complexity SAR image target detection method as described in claim 3, characterized in that, Further improvements to the feature fusion include employing depthwise separable convolutions to further reduce computational costs.

5. The low-complexity SAR image target detection method as described in claim 1, characterized in that, In model construction, improvements to the detection head include: adopting a decoupled design for classification and regression tasks, using independent pointwise convolutions for processing, and expanding channels according to different requirements for classification and bounding box extraction to improve the detection efficiency of the detection head.

6. The low-complexity SAR image target detection method as described in claim 1, characterized in that, The model deployment preparation also includes: before training the low-complexity SAR target detection network model, dividing the dataset into a training set, a validation set, and a test set, preprocessing the input images, and then using the training set to train the low-complexity SAR target detection network model.

7. The low-complexity SAR image target detection method as described in claim 6, characterized in that, Use the validation set to adjust the training parameters.

8. The low-complexity SAR image target detection method as described in claim 1, characterized in that, In the model deployment preparation, the accuracy quantization includes quantizing the FP32 accuracy model to FP16 accuracy.

9. The low-complexity SAR image target detection method as described in claim 1, characterized in that, Model deployment and inference also include: optimizing inference performance for different embedded platforms.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, implements the method of any one of claims 1 to 9.

Images