Abstract
A novel Dual-polarization SAR ship target recognition method based on feature and loss fusion deep network is proposed in this paper, to improve the generalization ability, the recognition accuracy of imbalance samples, and the real-time performance of the general deep learning network. The proposed combined network is composed of two parts. The first part is the target region extraction network based on lightweight mini Hourglass network, to eliminate the impact of data imbalance and background noise on the identification accuracy. The second part is a two-channel feature/loss function fusion network based on Efficient B2 backbone network, aiming to solve the problem of dual-polarization image feature fusion and iterative convergence acceleration. The proposed method is tested on SAR image ship slices from the OpenSAR data sets. The experimental results indicates that, the proposed method achieves a recognition rate of 110FPS with recognition accuracy of 87.72%, and exceeds the SOTA by recognition accuracy 3.72% with convergence speed improved by 75.68%. The proposed method can be applied to SAR target recognition with dual polarization, imbalance samples and low-resolution condition for reference.
Highlights
Synthetic Aperture Radar (SAR) plays an irreplaceable role in the military and civilian applications with the stable imaging under all-day and all-weather condition
Traditional SAR automatic target recognition (ATR) methods mainly include three types: the SAR target identification method based on template matching, the SAR target identification method based on sparse representation, and the SAR target identification method based on feature extraction [3]
Aiming at the key technical problems of SAR automatic target recognition system, such as identification accuracy, network reasoning time and network training time, the SAR ship target recognition method based on dual channel twin network, and the target region extraction method based on Mini Hourglass network, is proposed in this paper
Summary
Synthetic Aperture Radar (SAR) plays an irreplaceable role in the military and civilian applications with the stable imaging under all-day and all-weather condition. Most of the existing results are focused on MSTAR data sets, while the results on other target data sets are few, making it difficult to verify the universality of the existing methods on tasks of different target types The reason for this situation is that the public data sets of high-quality SAR images are very limited. Inspired by SE Block, Shao et al [18] proposed CSA block and introduced the channel-spatial attention mechanism to enhance the ability of network feature extraction and obtained 84% SOTA results on the OpenSARship data set. Such performance is far from the demand of practical applications.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
