Radar Target Recognition via Keypoint-Based Local Descriptor in the Context of Euclidean Space and Riemannian Manifold

Radar target recognition has been studied widely, yet the problem is still far from a solution. Early studies relied on holistic features or intensity values, which were sensitive to real-world sources of variability. High-level features learned by neural networks have been the topic of much recent research. This family of strategies requires many training samples to estimate model weights, and it suffers from underfitting in the case of limited training samples. In this article, we resort to the local patterns around points of interest. A refined image gradient is first presented. We approximate the gradients of radar images by means of improved ratio operators. The statistical specificities of multiplicative noise can then be exploited. The refined image gradients are used to detect the points of interest, around which local patterns are characterized. Local features are then considered in the context of Euclidean space and a Riemannian manifold. The former views a local feature as a typical Euclidean element and builds sparse signal modeling by means of dictionary learning. The latter regards the batch of local features as an element on a Riemannian manifold and draws the inference in the tangent and abstract kernel spaces. Comparative studies demonstrate that the proposed method can improve recognition performance with good computational efficiency.