Complex-valued Conditional Diffusion Model with Physical Scattering Information for SAR Image Target Recognition

Deep learning-based methods require samples for training, while synthetic aperture radar (SAR) images are limited due to the imaging environment, which restricts the recognition performance of SAR images. To access this issue, we develop a novel Complex-valued Conditional Diffusion Model incorporating scattering information (CDMSI), which solves the SAR small-sample problem by generating simulated complex SAR image for SAR image target recognition. We first design a complex diffusion model to generate complex SAR images with the amplitude and phase information, and then the labels are incorporated into complex diffusion model to guide the generating direction of the images from specific class. Comparing to traditional real-valued domain generative models, the complex-valued domain conditional diffusion model inherently exploits the algebraic structure of complex SAR images, offering a principled way to model intricate dependencies in data while maintaining interpretability. Our method is constructed under a complex network framework, which can explore more detailed and abundant features of SAR images for recognition. In addition, the physical mechanism of SAR imaging is incorporated into the complex networks to design a target recognition model, which learns the scattering center features of SAR images. The generated complex SAR images are utilized for model training, which can effectively improve the feature representativeness of SAR images under limited samples, and thus gain high recognition performance. Results on the measured MSTAR dataset validate that our method can generate complex SAR images with high quality, and also achieve superior SAR image recognition accuracy with limited available samples.