SAR Ship Detection via Knowledge Transfer: From Optical Image

Knowledge transfer technique was widely studied in target detection. Yet most of the preceding works were devoted to the transfer between the homologous domain. They were not effective to the knowledge transfer from optical image and SAR data. To solve the problem, a new SAR target detection framework was proposed in this paper. Different from the preceding works, more attentions were paid to the knowledge transfer from optical image to SAR data. It was much more challenging due to the significantly different image formation mechanism between the source domain and the target domain. Specifically, a new generative model based on contrastive language-image pretraining (CLIP) was first presented, by which an intermediate domain between the optical image and the SAR data can be built. The advantages of multi-sensor can be combined. A weighted loss function was then proposed to make the image quality balanced in the intermediate domain. On this foundation, the adversarial learning was introduced to achieve the feature distribution alignment. The problem of cross-domain transfer can be then solved accordingly. The pseudo-label inferred by the previously trained model were finally used to improve the target detection performance. In the end, multiple rounds of experiments were performed. The results clearly showcased the benefits of the proposed method over existing state-of-the-art techniques.