Generative knowledge transfer for ship detection in SAR images

Abstract

This paper proposes a generative transfer learning framework consisting of a knowledge transfer network and a ship detection network to address the data acquisition and labeling problem for ship detection in Synthetic Aperture Radar (SAR) images. The knowledge transfer network generates pseudo-SAR images whose spatial distributions are consistent with labeled optical images and feature distributions are similar to SAR images. These pseudo-SAR images are further used to improve the generalization performance of convolutional neural network-based detection models. Experimental results on SAR Ship Detection Dataset (SSDD) and AIR-SARShip-1.0 Dataset confirm that the pseudo-SAR images generated by our method can benefit the final detection prediction even no labeled SAR image is given at the training stage. The proposed framework can also significantly reduce the probability of missing alarms and false positives in complex backgrounds.