Optical Image-Assisted Zero-Shot Learning for Unknown Target Recognition in SAR Images

Zero-shot learning (ZSL) for unknown target recognition in synthetic aperture radar (SAR) images is of great significance as the training samples of certain targets may not be available under real-world scenarios. Existing ZSL algorithms usually integrate SAR images with semantics for unknown target recognition in SAR images. However, for these ZSL algorithms, semantics can only provide very limited information and some semantics may not be helpful compared with optical images. To this end, we develop an optical image-assisted ZSL (OZSL) algorithm, which includes three loss terms. Specifically, the first loss term explores intermodal correlation between SAR images and optical images. Subsequently, the second loss term explores the interview correlation among SAR images. However, combining these two loss terms is still unable to recognize unknown SAR images. Considering this, the third loss term is designed by exploiting additional optical images to assist the above two loss terms. Afterward, these three loss terms are jointly optimized to acquire the optimal solutions of the related variables involved in them by the corresponding optimization algorithms. Finally, according to the Kullback–Leibler divergence, OZSL exploits the optimal solutions to achieve unknown SAR image recognition. The results verify the superiority of OZSL.