STRE-Diffusion: Few-Shot Spaceborne ISAR Target Recognition via Spatial–Time Relationship Estimation and Improved Diffusion Model
- 9 September 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 61 (6) , 17906-17923
Abstract
In spaceborne inverse synthetic aperture radar (ISAR) space target recognition, the high-speed relative motion restricts the observation time, leading to insufficient and low diversity ISAR datasets and reduced recognition accuracy. In existing methods, the dataset augmentation can be achieved by interpolating the azimuth angle of ISAR images obtained in the observation region. However, they lack the capability to extrapolate images in non-observation region, limiting the effectiveness in enhancing the angular diversity of the dataset. In this article, a few-shot spaceborne ISAR target recognition method via spatial-time relationship estimation and improved diffusion network (STRE-Diffusion) is proposed. The proposed method incorporates the spatial-time relationship (STR) into ISAR images prediction, which not only provides additional information for target motion modeling but also enhances predicted image details. First, visual-domain motion cues are extracted via optical flow estimation. Next, STR is estimated using the optical flow of strong scatterers. Subsequently, a motion cues extrapolation denoising (MCED) module is employed to extrapolate the cross-domain motion cues to the future through inverse diffusion process. In the MCED, a lightweight spatial-time bi-level routing transformer (ST-BiFormer) is proposed for spatial-time feature fusion and alignment. Then the extrapolated motion cues are mapped back to the image domain to obtain the predicted images. Finally, the predicted images are used to construct the extended dataset, which is applied to few-shot recognition. Experimental results show that the predicted ISAR images generated by proposed method have higher authenticity. The few-shot recognition performance based on the extended dataset has been significantly improved.Keywords
Funding Information
- National Natural Science Foundation of China (61971163)
This publication has 46 references indexed in Scilit:
- ExtDM: Distribution Extrapolation Diffusion Model for Video Prediction2024
- Application of deep generative networks for SAR/ISAR: a reviewArtificial Intelligence Review, 2023
- Unified Multivariate Gaussian Mixture for Efficient Neural Image Compression2022
- ISAR Images Generation Via Generative Adversarial Networks2021
- Motion Representations for Articulated Animation2021
- Exploring Simple Siamese Representation Learning2021
- Dynamic Few-Shot Visual Learning Without Forgetting2018
- Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset2017
- Deep Residual Learning for Image Recognition2016
- Very deep convolutional neural network based image classification using small training sample size2015