融合可微分渲染的SAR多视角样本增广

贾赫成; 蒲欣洋; 王燕妮; 符士磊; 徐丰

doi:10.12000/JR24011

融合可微分渲染的SAR多视角样本增广

doi: 10.12000/JR24011

电磁波信息科学教育部重点实验室复旦大学上海 200433

基金项目: 国家自然科学基金(61991422)

详细信息

作者简介:
贾赫成，博士生，主要研究方向为SAR图像解译

蒲欣洋，博士生，主要研究方向为遥感图像解译、视觉基础模型

王燕妮，博士生，主要研究方向为深度强化学习在电磁领域中的应用

符士磊，博士生，主要研究方向为SAR图像解译、SAR三维重建

徐　丰，博士，教授，主要研究方向为SAR图像解译、电磁散射建模、智能信息技术

通讯作者:
徐丰 fengxu@fudan.edu.cn

责任主编：仇晓兰 Corresponding Editor: QIU Xiaolan
中图分类号: TN95
计量
- 文章访问数: 207
- HTML全文浏览量: 42
- PDF下载量: 29
- 被引次数: 0
出版历程
- 收稿日期: 2024-01-16
- 修回日期: 2024-03-21
- 网络出版日期: 2024-03-28
- 刊出日期: 2024-04-28

Multi-view Sample Augumentation for SAR Based onDifferentiable SAR Renderer

Key Lab for Information Science of Electromagnetic Waves (MoE), Fudan University, Shanghai 200433, China

Funds: The National Natural Science Foundation of China (61991422)

More Information

Corresponding author: XU Feng, fengxu@fudan.edu.cn

摘要

摘要: 合成孔径雷达(SAR)因其全天候、全天时的监测能力在民用和军事领域得到广泛应用。近年来，深度学习已被广泛应用于SAR图像自动解译。然而，由于卫星轨道和观测角度的限制，SAR目标样本面临视角覆盖率不全的问题，这为学习型SAR目标检测识别算法带来了挑战。该文提出一种融合可微分渲染的SAR多视角样本生成方法，结合逆向三维重建和正向渲染技术，通过卷积神经网络(CNN)从少量SAR视角图像中反演目标三维表征，然后利用可微分SAR渲染器(DSR)渲染出更多视角样本，实现样本在角度维的插值。另外，方法的训练过程使用DSR构建目标函数，无需三维真值监督。根据仿真数据的实验结果，该方法能够有效地增加多视角SAR目标图像，并提高小样本条件下典型SAR目标识别率。
- 合成孔径雷达(SAR) /
- 可微分SAR渲染器(DSR) /
- 卷积神经网络(CNN) /
- 三维重建 /
- 多视角样本生成
Abstract: Synthetic Aperture Radar (SAR) is extensively utilized in civilian and military domains due to its all-weather, all-time monitoring capabilities. In recent years, deep learning has been widely employed to automatically interpret SAR images. However, due to the constraints of satellite orbit and incident angle, SAR target samples face the issue of incomplete view coverage, which poses challenges for learning-based SAR target detection and recognition algorithms. This paper proposes a method for generating multi-view samples of SAR targets by integrating differentiable rendering, combining inverse Three-Dimensional (3D) reconstruction, and forward rendering techniques. By designing a Convolutional Neural Network (CNN), the proposed method inversely infers the 3D representation of targets from limited views of SAR target images and then utilizes a Differentiable SAR Renderer (DSR) to render new samples from more views, achieving sample interpolation in the view dimension. Moreover, the training process of the proposed method constructs the objective function using DSR, eliminating the need for 3D ground-truth supervision. According to experimental results on simulated data, this method can effectively increase the number of multi-view SAR target images and improve the recognition rate of typical SAR targets under few-shot conditions.
- Synthetic Aperture Radar (SAR) /
- Differentiable SAR Renderer (DSR) /
- Convolutional Neural Network (CNN) /
- 3D reconstruction /
- Multi-view sample generation

HTML全文

图 1 SAR目标视角覆盖不足的原因分析

Figure 1. Analysis of insufficient view coverage for SAR targets

下载: 全尺寸图片幻灯片

图 2 SAR多视角样本增广算法框架

Figure 2. Overall framework of multi-view sample augumentation for SAR

下载: 全尺寸图片幻灯片

图 3 目标和阴影区域预处理

Figure 3. Preprocessing of target and shadow regions

下载: 全尺寸图片幻灯片

图 4 目标重建网络结构

Figure 4. Structure of the target reconstruction network

下载: 全尺寸图片幻灯片

图 5 目标照射图和阴影图生成示意

Figure 5. Illustration of illumination map and shadow map generation

下载: 全尺寸图片幻灯片

图 6 目标重建结果可视化

Figure 6. Visualization of target reconstruction results

下载: 全尺寸图片幻灯片

图 7 SAR车辆目标生成结果可视化

Figure 7. Visualization of SAR vehicle target generation results

下载: 全尺寸图片幻灯片

图 8 目标区域和背景区域像素的Gamma分布拟合

Figure 8. Gamma distribution fitting of pixel values in target and background regions

下载: 全尺寸图片幻灯片

图 9 不同视角数量下的目标重建结果评估

Figure 9. Evaluation of target reconstruction results with different numbers of views

下载: 全尺寸图片幻灯片

图 10 T72实测样本的目标和阴影区域分割示意

Figure 10. Illustration of target and shadow area segmentation for T72 measured samples

下载: 全尺寸图片幻灯片

图 11 基于实测数据重建的目标三维模型可视化

Figure 11. Visualization of target 3D models reconstructed based on measured data

下载: 全尺寸图片幻灯片

图 12 生成图像中的目标和阴影区域

Figure 12. Target and shadow areas in generated images

下载: 全尺寸图片幻灯片

图 13 阴影区域形态学处理示意

Figure 13. Illustration of morphological processing in shadow areas

下载: 全尺寸图片幻灯片

图 14 不同阴影分割准确性下的实验结果

Figure 14. Experimental results under different shadow segmentation accuracies

下载: 全尺寸图片幻灯片

图 15 各类别中间结果的三维模型可视化

Figure 15. Visualization of intermediate 3D models for each category

下载: 全尺寸图片幻灯片

图 16 各类别生成样本与真值可视化对比

Figure 16. Visualization comparison of generated samples and ground truth for each category

下载: 全尺寸图片幻灯片

图 17 扩充前后识别结果混淆矩阵

Figure 17. Confusion matrices of recognition results before and after augmentation

下载: 全尺寸图片幻灯片

表 1 仿真数据集信息

Table 1. Information of the simulated dataset

实验数据集	入射角$ \alpha $	方位角$ \beta $
训练集	{30°, 40°, 50°, 60°}	{0°, 45°, 90°, ···, 315°}
测试集1	{30°, 40°, 50°, 60°}	{30°, 75°, 120°, ···, 345°}
测试集2	{35°, 45°, 55°, 65°}	{30°, 75°, 120°, ···, 345°}

下载: 导出CSV

表 2 各目标重建结果的mIoU

Table 2. mIoU of reconstruction results for each target

目标类型	训练集	测试集1	测试集2
车辆	0.8410	0.8375	0.8401
飞机	0.8629	0.8521	0.8554
风车	0.7860	0.7858	0.7880

下载: 导出CSV

表 3 小样本识别实验数据配置

Table 3. Configuration of experimental data for few-shot recognition

训练数据	入射角$ \alpha $	方位角$ \beta $	样本数量
训练集	{30°, 40°, 50°, 60°}	{0°, 45°, 90°, ···, 315°}	320
测试集		{5°, 15°, 25°, ···, 355°}	1440
扩充1		{0°, 30°, 60°, ···, 330°}	480
扩充2		{0°, 20°, 40°, ···, 340°}	720
扩充3		{0°, 10°, 20°, ···, 350°}	1440

下载: 导出CSV

表 4 各类别重建三维模型结果评估

Table 4. Quantitative evaluation of intermediate 3D models for each category

目标	mIoU	$ {{L}}_{\mathrm{f}\mathrm{l}\mathrm{a}\mathrm{t}} $	目标	mIoU	$ {{L}}_{\mathrm{f}\mathrm{l}\mathrm{a}\mathrm{t}} $
BRDM2	0.7919	0.0036	ZIL131	0.7995	0.0050
BTR60	0.7812	0.0029	ZSU234	0.6000	0.0029
D7	0.6653	0.0039	T72	0.6406	0.0032
2S1	0.8052	0.0030	BMP2	0.7982	0.0022
T62	0.7321	0.0040	BTR70	0.7781	0.0030

下载: 导出CSV

表 5 各类别散射纹理Gamma拟合参数估计

Table 5. Estimation of Gamma fitting parameters for scattering texture of each category

目标	背景区域		目标区域
目标	参数a	参数b	参数a	参数b
BRDM2	2.583	0.014	1.104	0.147
BTR60	2.125	0.022	1.094	0.185
D7	3.898	0.009	0.909	0.245
2S1	3.597	0.012	1.145	0.169
T62	3.331	0.010	1.070	0.191
ZIL131	4.076	0.008	1.080	0.157
ZSU234	3.144	0.009	0.955	0.216
T72	2.550	0.018	1.066	0.202
BMP2	2.561	0.018	1.027	0.191
BTR70	2.610	0.018	1.179	0.141

下载: 导出CSV

表 6 样本扩充与小样本识别结果

Table 6. Results of few-shot recognition with sample augmentation

训练数据	样本总数	旋转增强	准确率(%)
训练数据	样本总数	旋转增强	VGG16	ResNet50	HRNet	Swin-tiny
原训练集	320	–	89.65	92.50	85.35	89.79
原训练集	320	√	95.69	96.87	94.86	95.83
扩充1	800	–	94.24	96.94	95.76	94.24
扩充2	1040	–	96.32	97.50	95.83	94.65
扩充3	1760	–	95.90	97.29	95.69	96.39
扩充3	1760	√	97.78	99.03	97.85	98.96

下载: 导出CSV

表 7 ResNet50的各类别识别结果准确率(%)

Table 7. Per-category recognition accuracy using ResNet50 (%)

训练数据	BRDM2	BTR60	D7	2S1	T62	ZIL	ZSU	T72	BMP2	BTR70
原始数据集	85.16	98.62	97.95	97.93	96.50	100	99.31	88.82	86.96	75.60
扩充1	99.26	100	100	99.31	100	100	100	100	90.62	82.25
扩充2	98.57	100	100	100	100	100	100	100	92.31	85.37
扩充3	99.29	100	100	100	100	100	98.63	99.31	86.90	89.04

下载: 导出CSV

参考文献(26)

[1]	ZHANG Liangpei, ZHANG Lefei, and DU Bo. Deep learning for remote sensing data: A technical tutorial on the state of the art[J]. IEEE Geoscience and Remote Sensing Magazine, 2016, 4(2): 22–40. doi: 10.1109/MGRS.2016.2540798.
[2]	MA Lei, LIU Yu, ZHANG Xueliang, et al. Deep learning in remote sensing applications: A meta-analysis and review[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 152: 166–177. doi: 10.1016/j.isprsjprs.2019.04.015.
[3]	ZHU Xiaoxiang, TUIA D, MOU Lichao, et al. Deep learning in remote sensing: A comprehensive review and list of resources[J]. IEEE Geoscience and Remote Sensing Magazine, 2017, 5(4): 8–36. doi: 10.1109/MGRS.2017.2762307.
[4]	SUN Xian, WANG Bing, WANG Zhirui, et al. Research progress on few-shot learning for remote sensing image interpretation[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 2387–2402. doi: 10.1109/JSTARS.2021.3052869.
[5]	HUANG Zhongling, PAN Zongxu, and LEI Bin. What, where, and how to transfer in SAR target recognition based on deep CNNs[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(4): 2324–2336. doi: 10.1109/TGRS.2019.2947634.
[6]	SHORTEN C and KHOSHGOFTAAR T M. A survey on image data augmentation for deep learning[J]. Journal of Big Data, 2019, 6(1): 60. doi: 10.1186/s40537-019-0197-0.
[7]	HENDRYCKS D, MU N, CUBUK E D, et al. AugMix: A simple data processing method to improve robustness and uncertainty[C]. 8th International Conference on Learning Representations, Addis Ababa, Ethiopia, 2010.
[8]	ALFASSY A, KARLINSKY L, AIDES A, et al. LaSO: Label-set operations networks for multi-label few-shot learning[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 6541–6550. doi: 10.1109/CVPR.2019.00671.
[9]	SCHWARTZ E, KARLINSKY L, SHTOK J, et al. Δ-encoder: An effective sample synthesis method for few-shot object recognition[C]. 32nd International Conference on Neural Information Processing Systems, Montréal, Canada, 2018: 2850–2860.
[10]	ANTONIOU A, STORKEY A, and EDWARDS H. Data augmentation generative adversarial networks[EB/OL]. https://arxiv.org/abs/1711.04340v3, 2018.
[11]	MALMGREN-HANSEN D, KUSK A, DALL J, et al. Improving SAR automatic target recognition models with transfer learning from simulated data[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(9): 1484–1488. doi: 10.1109/LGRS.2017.2717486.
[12]	GUO Jiayi, LEI Bin, DING Chibiao, et al. Synthetic aperture radar image synthesis by using generative adversarial nets[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(7): 1111–1115. doi: 10.1109/LGRS.2017.2699196.
[13]	SONG Qian, XU Feng, ZHU Xiaoxiang, et al. Learning to generate SAR images with adversarial autoencoder[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5210015. doi: 10.1109/TGRS.2021.3086817.
[14]	GUO Qian and XU Feng. Learning low-dimensional SAR target representations from few samples[C]. 2022 International Applied Computational Electromagnetics Society Symposium, Xuzhou, China, 2022: 1–2. doi: 10.1109/ACES-China56081.2022.10065101.
[15]	LIU Shichen, CHEN Weikai, LI Tianye, et al. Soft rasterizer: A differentiable renderer for image-based 3D reasoning[C]. 2019 IEEE/CVF International Conference on Computer Vision, Seoul, Korea (South), 2019: 7708–7717. doi: 10.1109/ICCV.2019.00780.
[16]	WANG Nanyang, ZHANG Yinda, LI Zhuwen, et al. Pixel2Mesh: Generating 3D mesh models from single RGB images[C]. 15th European Conference on Computer Vision, Munich, Germany, 2018: 52–67. doi: 10.1007/978-3-030-01252-6_4.
[17]	WEN Chao, ZHANG Yinda, LI Zhuwen, et al. Pixel2Mesh++: Multi-view 3D mesh generation via deformation[C]. 2019 IEEE/CVF International Conference on Computer Vision, Seoul, Korea (South), 2019: 1042–1051. doi: 10.1109/ICCV.2019.00113.
[18]	FU Shilei and XU Feng. Differentiable SAR renderer and image-based target reconstruction[J]. IEEE Transactions on Image Processing, 2022, 31: 6679–6693. doi: 10.1109/TIP.2022.3215069.
[19]	KATO H, USHIKU Y, and HARADA T. Neural 3D mesh renderer[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 3907–3916. doi: 10.1109/CVPR.2018.00411.
[20]	XU Feng and JIN Yaqiu. Imaging simulation of polarimetric SAR for a comprehensive terrain scene using the mapping and projection algorithm[J]. IEEE Transactions on Geoscience and Remote Sensing, 2006, 44(11): 3219–3234. doi: 10.1109/TGRS.2006.879544.
[21]	FU Shilei, JIA Hecheng, PU Xinyang, et al. Extension of differentiable SAR renderer for ground target reconstruction from multiview images and shadows[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5217013. doi: 10.1109/TGRS.2023.3320515.
[22]	Moving and stationary target acquisition and recognition (MSTAR) public release data[EB/OL]. https://www.sdms.afrl.af.mil/index.php?collection=mstar.
[23]	SIMONYAN K and ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]. 3rd International Conference on Learning Representations, San Diego, USA, 2015.
[24]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 770–778. doi: 10.1109/CVPR.2016.90.
[25]	SUN Ke, XIAO Bin, LIU Dong, et al. Deep high-resolution representation learning for human pose estimation[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 5693–5703. doi: 10.1109/CVPR.2019.00584.
[26]	LIU Ze, LIN Yutong, CAO Yue, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]. 2021 IEEE/CVF International Conference on Computer Vision, Montreal, Canada, 2021: 10012–10022. DOi: 10.1109/ICCV48922.2021.00986.