SAR图像快速扩散生成方法

doi:10.19665/j.issn1001-2400.20250503

摘要/Abstract

摘要：

现有基于深度学习的SAR目标分类算法性能依赖于大量数据,但受限于高采集成本和目标非合作性等影响,获取的数据量有限,从而导致算法性能显著下降,限制了雷达图像分类的广泛应用。目前基于生成对抗网络的SAR图像生成方法存在生成质量低、易发生模式坍塌和数据影响因子不可控等问题。为此,提出了一种SAR图像快速扩散生成方法,以提高基于深度学习方法的雷达目标识别准确率。通过在扩散模型中嵌入基于互注意力的条件机制,将SAR图像目标的类别和方位角分别进行编码,从而引导扩散模型生成指定类别和方位角的SAR图像。为解决扩散模型生成速度慢的问题,采用快速求解器DPM-Solver加速采样过程,并探究不同超参数对生成质量的影响。在MSTAR数据集上的实验表明,在数据量为60%时,扩散模型生成的增广样本可以将目标分类模型对SAR目标的分类准确率从95.16%提升至97.62%,并且使用DPM-Solver加速生成后的增广样本能够进一步将分类准确率提高到97.98%。

关键词: 扩散模型, 合成孔径雷达, 图像生成, 目标分类

Abstract:

Existing deep learning algorithms for Synthetic Aperture Radar (SAR) target classification rely heavily on large datasets.However,the high cost of radar data acquisition and the non-cooperative targets limit data availability,significantly degrading algorithm performance.This limitation restricts the widespread application of radar image classification algorithms.Current methods for generating SAR images using generative adversarial networks face challenges such as a low-quality output,mode collapse,and uncontrollable factors.To address these issues,we propose an accelerated diffusion-based method for SAR image generation to enhance the radar target recognition accuracy.By embedding a cross-attention conditional mechanism,we encode the class and azimuth of targets to guide the generation of SAR images with specified classes and azimuth angles.To overcome the slow generation speed of the diffusion model,we utilize the DPM-Solver to accelerate the sampling process and explore the impact of different hyperparameters on generation quality.Experiments on the MSTAR dataset demonstrate that with 60% of the data,the augmented samples generated by the diffusion model can enhance the classification accuracy of the SAR target classification model from 95.16% to 97.62%.Moreover,the augmented samples generated by the DPM-Solver can further improve the classification accuracy to 97.98%.

Key words: diffusion model, synthetic aperture radar, image generation, target classification

中图分类号:

TP391

顿皓, 田春娜, 李向阳, 单笑, 郭宇杰. SAR图像快速扩散生成方法[J]. 西安电子科技大学学报, 2025, 52(3): 217-231.

DUN Hao, TIAN Chunna, LI Xiangyang, SHAN Xiao, GUO Yujie. Accelerated diffusion-based method for SAR image generation[J]. Journal of Xidian University, 2025, 52(3): 217-231.

图/表 20

图1

图2

图3

表1

表2

图4

表3

表4

图5

图6

表5

表6

表7

图7

表8

表9

图8

图9

表10

图10

参考文献 38

[1]	谢雯, 滑文强, 焦李成, 等. 采用深度学习的极化SAR地物分类方法综述[J]. 西安电子科技大学学报, 2023, 50(3):151-170.
	XIE Wen, HUA Wenqiang, JIAO Licheng, et al. Review on Polarimetric SAR Terrain Classification Methods Using Deep Learning[J]. Journal of Xidian University, 2023, 50(3):151-170.
[2]	赵鹏菲, 黄丽佳. 一种基于EfficientNet与BiGRU的多角度SAR图像目标识别方法[J]. 雷达学报, 2021, 10(6):895-904.
	ZHAO Pengfei, HUANG Lijia. Target Recognition Method for Multi-Aspect Synthetic Aperture Radar Images Based on EfficientNet and BiGRU[J]. Journal of Radars, 2021, 10(6):895-904.
[3]	MA X, JI K, FENG S, et al. Open Set Recognition with Incremental Learning for SAR Target Classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61(1):1-14.
[4]	GENG Z, XU Y, WANG B N, et al. Target Recognition in SAR Images by Deep Learning with Training Data Augmentation[J]. Sensors, 2023, 23(2):941.
[5]	KANG M, ZHU J Y, ZHANG R, et al. Scaling up GANs for Text-to-Image Synthesis[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE,2023:10124-10134.
[6]	MA X, FANG G, WANG X. DeepCache:Accelerating Diffusion Models for Free[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE,2024:15762-15772.
[7]	WU T, FAN Z, LIU X, et al. AR-Diffusion:Auto-Regressive Diffusion Model for Text Generation[C]// Proceedings of the Advances in Neural Information Processing Systems 36:Annual Conference on Neural Information Processing Systems 2023. San Diego: NIPS,2023:39957-39974.
[8]	DENG J, LU J, ZHANG T. Diff3DETR:Agent-Based Diffusion Model for Semi-Supervised 3D Object Detection[C]// European Conference on Computer Vision. Berlin:Springer,2024:57-73.
[9]	ESSER P, KULAL S, BLATTMANN A, et al. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis[C]// ICML'24:Proceedings of the 41th International Conference on Machine Learning. San Diego: PMLR,2024:12606-12633.
[10]	KINGMA D P, WELLING M. Auto-Encoding Variational Bayes[C]// 2nd International Conference on Learning Representations. Banff:ICLR,2013:1-14.
[11]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative Adversarial Networks[J]. Communications of the ACM, 2020, 63(11):139-144.
[12]	MIRZA M, OSINDERO S. Conditional Generative Adversarial Nets (2014)[J/OL]. [2014-11-06]. https://arxiv.org/abs/1411.1784.
[13]	ODENA A, OLAH C, SHLENS J. Conditional Image Synthesis with Auxiliary Classifier GANs[C]// ICML'17:Proceedings of the 34th International Conference on Machine Learning. San Diego: PMLR,2017:2642-2651.
[14]	CHEN X, DUAN Y, HOUTHOOFT R, et al. InfoGAN:Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets[C]// Proceedings of the Advances in Neural Information Processing Systems 29:Annual Conference on Neural Information Processing Systems 2016. San Diego: NIPS,2016:2180-2188.
[15]	KARRAS T, LAINE S, AILA T. A Style-Based Generator Architecture for Generative Adversarial Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(12):4217-4228.
[16]	SAUER A, KARRAS T, LAINE S, et al. StyleGAN-T:Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis[C]// ICML'23:Proceedings of the 40th International Conference on Machine Learning. San Diego: PMLR,2023:30105-30118.
[17]	GUO J, LEI B, DING C, et al. Synthetic Aperture Radar Image Synthesis by Using Generative Adversarial Nets[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(7):1111-1115.
[18]	ZHAI J, DANG X, CHEN F, et al. SAR Image Generation Using Structural Bayesian Deep Generative Adversarial Network[C]// 2019 Photonics & Electromagnetics Research Symposium -Fall. Xiamen:PIERS,2019:1386-1392.
[19]	李诗怡, 付光远, 崔忠马, 等. 基于多尺度生成对抗网络的SAR图像样本增广[J]. 激光与光电子学进展, 2020, 57(20):175-185.
	LI Shiyi, FU Guangyuan, CUI Zhongma, et al. SAR Aircraft Data Sets Augmentation Based on Multi-Scale Generative Adversarial Network[J]. Laser & Optoelectronics Progress, 2020, 57(20):175-185.
[20]	OH J, KIM M. PeaceGAN:A GAN-Based Multi-Task Learning Method for SAR Target Image Generation with a Pose Estimator and an Auxiliary Classifier[J]. Remote Sensing, 2021, 13(19):3939.
[21]	FENG Z, DAKOVIC M, JI H, et al. Interpretation of Latent Codes in InfoGAN with SAR Images[J]. Remote Sensing, 2023, 15(5):1254.
[22]	HO J, JAIN A, ABBEEL P. Denoising Diffusion Probabilistic Models[C]// Proceedings of the Advances in Neural Information Processing Systems 33:Annual Conference on Neural Information Processing Systems 2020. San Diego: NIPS,2020:6840-6851.
[23]	DHARIWAL P, NICHOL A. Diffusion Models Beat GANs on Image Synthesis[C]// Proceedings of the Advances in Neural Information Processing Systems 34:Annual Conference on Neural Information Processing Systems 2021. San Diego: NIPS,2021:8780-8794.
[24]	SAHARIA C, CHAN W, SAXENA S, et al. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding[C]// Proceedings of the Advances in Neural Information Processing Systems 35:Annual Conference on Neural Information Processing Systems 2022. San Diego: NIPS,2022:36479-36494.
[25]	PEEBLES W, XIE S. Scalable Diffusion Models with Transformers[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Piscataway:IEEE,2023:4195-4205.
[26]	CHEN J, YU J, GE C, et al. PixArt-α:Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis[C]// 12th International Conference on Learning Representations. Vienna:ICLR,2024:1-30.
[27]	ZHENG W, TENG J, YANG Z, et al. CogView3:Finer and Faster Text-to-Image Generation via Relay Diffusion[C]// European Conference on Computer Vision. Berlin:Springer,2024:1-22.
[28]	LU C, ZHOU Y, BAO F, et al. DPM-Solver:A Fast Ode Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps[C]// Proceedings of the Advances in Neural Information Processing Systems 35:Annual Conference on Neural Information Processing Systems 2022. San Diego: NIPS,2022:5775-5787.
[29]	MILDENHALL B, SRINIVASAN P P, TANCIK M, et al. NeRF:Representing Scenes as Neural Radiance Fields for View Synthesis[J]. Communications of the ACM, 2021, 65(1):99-106.
[30]	SONG Y, SOHL-DICKSTEIN J, KINGMA D P, et al. Score-Based Generative Modeling through Stochastic Differential Equations[C]// 9th International Conference on Learning Representations. Virtual:ICLR,2021:1-36.
[31]	HOCHBRUCK M, OSTERMANN A. Explicit Exponential Runge-Kutta Methods for Semilinear Parabolic Problems[J]. SIAM Journal on Numerical Analysis, 2005, 43(3):1069-1090.
[32]	LUAN V T. Efficient Exponential Runge-Kutta Methods of High Order:Construction and Implementation[J]. BIT Numerical Mathematics, 2021, 61(2):535-560.
[33]	KEYDEL E R, LEE S W, MOORE J T. MSTAR Extended Operating Conditions:A Tutorial[J]. Algorithms for Synthetic Aperture Radar Imagery Ⅲ, 1996,2757:228-242.
[34]	LEWIS B, SCARNATI T, SUDKAMP E, et al. A SAR Dataset for ATR Development:the Synthetic and Measured Paired Labeled Experiment (SAMPLE)[C]// SPIE Defense+Commercial Sensing.Algorithms for Synthetic Aperture Radar Imagery ⅩⅩⅥ: 10987. Baltimore:SPIE,2019:39-54.
[35]	张倩. SAR图像质量评估及其目标识别应用[D]. 合肥: 中国科学技术大学, 2011.
[36]	HEUSEL M, RAMSAUER H, UNTERTHINER T, et al. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium[C]// Proceedings of the Advances in Neural Information Processing Systems 30:Annual Conference on Neural Information Processing Systems 2017. Piscataway:IEEE,2017:6629-6640.
[37]	SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the Inception Architecture for Computer Vision[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE,2016:2818-2826.
[38]	CHEN X, LIANG C, HUANG D, et al. Symbolic Discovery of Optimization Algorithms[C]// Proceedings of the Advances in Neural Information Processing Systems 36:Annual Conference on Neural Information Processing Systems 2023. San Diego: NIPS,2023:49205-49233.

类别	训练集		测试集
类别	俯仰角/(°)	数量/张	俯仰角/(°)	数量/张
2S1	17	299	15	274
BMP2	17	233	15	195
BRDM2	17	298	15	274
BTR60	17	256	15	195
BTR70	17	233	15	196
D7	17	299	15	274
T62	17	298	15	273
T72	17	232	15	196
ZIL131	17	299	15	274
ZSU23/4	17	299	15	274
总计	2 746		2 425

类别	训练集		测试集
类别	俯仰角/(°)	数量/张	俯仰角/(°)	数量/张
2S1	15,16	116	17	58
BMP2	16	55	17	52
BTR70	16	43	17	49
M1	14,16	78	17	51
M2	14,16	75	17	53
M35	14,16	76	17	53
M548	16	75	17	53
M60	15,16	116	17	60
T72	16	56	17	52
ZSU23	15,16	116	17	58
总计	806		539

步数	二阶			三阶
步数	均匀	二次型	对数信噪比	均匀	二次型	对数信噪比
10	22.13	8.91	5.00	61.49	15.17	14.15
15	21.51	8.29	5.78	67.75	8.32	7.35
20	21.07	7.45	6.06	48.98	6.44	6.55
25	19.23	6.90	6.03	32.49	6.12	6.18
50	11.31	6.22	6.05	10.72	6.03	6.07
100	7.43	6.09	6.00	5.81	6.03	6.03
250	6.21	6.04	6.00	5.99	6.02	6.02

步数	距离得分
7	4.28
8	3.91
9	4.18
10	5.00
11	5.21

数据集	均值	方差	等效视数	辐射分辨率	IS	距离得分	生成时间/(s·张^-1)
MSTAR数据集	28.17	1 134.27	0.758	3.364	1.187
DDPM	31.89	1 637.14	0.673	3.493	1.119	37.65	6.355
DiT-S/2	33.58	1 823.60	0.682	3.494	1.280	8.83	11.622
文中方法	30.56	1 171.69	0.859	3.219	1.177	3.91	0.041
SAMPLE数据集	49.57	838.55	2.99	2.01	1.305
DDPM	48.29	809.60	2.93	2.02	1.338	6.48	6.381
DiT-S/2	52.29	994.41	2.84	2.08	1.684	37.63	11.602
文中方法	50.39	840.20	3.08	1.99	1.344	5.83	0.041