Data-Driven Sea Clutter Suppression via an Image-to-Image Neural Network and Consistency Loss
Qizhe Qu, Hao Chen, Weijian Liu, Binbin Li, Yongliang Wang
Abstract
Marine target detection under diverse sea conditions has been a research hotspot. An effective sea clutter suppression method is supposed to be beneficial to marine target detection. However, traditional model-driven suppression methods are likely to suffer from the model mismatch problem. To overcome these problems, we propose a data-driven sea clutter suppression method based on an image-to-image neural network. The proposed method first takes complex-valued range-Doppler (RD) images as inputs by concatenating real and imaginary parts. The designed network is then capable of capturing significant phase information and learning deep features of both targets and clutter. Finally, clutter-free RD images are predicted directly by the proposed network. Meanwhile, a consistency loss is elaborately designed to suppress clutter residuals and noise points with high energy. Measured results with the Sea-Detecting Radar Data-sharing Program (SDRDSP) database verify the advantages of the consistency loss and show that the proposed network outperforms its real-valued counterpart. Compared with four classical clutter suppression methods, the proposed method achieves the best suppression performance with relatively lower computational complexity.