Radar Weak Target Detection Based on Deep Convolutional Feature Manifold Network

This paper addresses the problem of radar weak target detection in strong clutter by proposing a deep convolutional feature manifold network detector. Specifically, convolution neural network is exploited to extract deep convolutional features from time-frequency diagrams of radar echo signals. The convolutional features are modeled as symmetric positive definite (SPD) matrices with an eigenvalue-correcting operation. Then, the SPD manifold network is used to reduce the redundant information in high-dimensional SPD manifold by mapping them into a lower-dimensional manifold to maximize discrimination. Since the high-dimensional nonlinearity convolutional features exhibit excellent discriminative power, the proposed algorithm achieves a detection performance improvement. Experimental results performed on a publicly available radar dataset confirm the advantages of the proposed algorithm against the SPD manifold detection method in strong clutter.