Implicit Neural Representation With Imaging Geometry for SAR Target Recognition

Although research on synthetic aperture radar (SAR) automatic target recognition is well established, many studies assume that sufficient data are available, i.e., a large number of training samples. However, acquiring sufficiently labeled SAR data is challenging, making it difficult to obtain adequate samples from various viewpoints for each target in practical scenarios. Previous methods rely on synthesizing from neighboring viewpoints or using data-driven generative networks to extend the data, which are relatively straightforward but lack interpretability. To address this, we propose an implicit neural representation model based on the imaging geometry of SAR, named SAR-INR, to represent the 3-D electromagnetic backscatter intensity of a target. This model is capable of synthesizing SAR images of unseen viewpoints, using only a limited number of viewpoint images for training. Specifically, our method employs a neural network to represent an SAR scene, using the spatial coordinates and beam orientation of query points as inputs to predict their density and intensity. We introduce a rigorous SAR epipolar geometry modeling to inverse the pixel position of the 2-D SAR image into 3-D spatial grids. The parameters of the network are optimized using the actual pixel values from the SAR images as ground truth to supervise the rendered outputs for these grids. Once trained, this model can synthesize SAR images from unseen viewpoints, effectively enhancing data utilization. Experimental results on real data show that our method not only expands the ability to synthesize unseen perspectives but also significantly enhances performance with limited data samples.