A Semi-Supervised Method Using Cycle Consistency and Multi-Perspective Dilated for Sar-to-Optical Translation

作     者：Wang, Jinyu Yang, Haitao Liu, Zhengjun Chen, Hang 

作者机构：Department of Space Equipment The Academy of Equipment Beijing101416 China Department of Automation Measurement and Control Harbin Institute of Technology Harbin150001 China Space Engineering University China 

出 版 物：《SSRN》 

年 卷 期：2024年

核心收录：

主　　题：Radar imaging 

摘      要：Synthetic Aperture Radar (SAR) is a powerful imaging sensor capable of operating in any weather and at any time, despite its image interpretation poses challenges. Translating SAR images to optical (S2O) can serve as a beneficial supplement for target detection, image alignment, and image fusion. The existing S2O translation employs solely unsupervised or supervised techniques. However, the supervised method requires a substantial quantity of aligned data support, while the unsupervised method fails to guarantee the quality of image generation. This research presents a novel semi-supervised S2O generating approach that can obtain domain-level and pixel-level mapping relationships. A dual-branch cycle consistency loss has been developed to independently constrain supervised and unsupervised modules. To improve global and sparsity feature extraction, the generator combines sliding dilated convolution with multi-respective parallel branching. Using just 1100 pairs of training sets, extensive experiments have demonstrated that the suggested strategy outperforms baseline methods in both subjective and objective standards. Where FID is optimized by 23.41%, 49.25%, 48.57%, 16.73%, and 42.32% on the five land classes. © 2024, The Authors. All rights reserved.