Combining a single shot multibox detector with transfer learning for ship detection using sentinel-1 SAR images

Abstract

With the capabilities of constant use in any weather condition and a wide coverage area, synthetic aperture radar (SAR) technology is widely used in marine transportation safety and fishing law enforcement. Currently, due to automatic learning of discriminative features, deep learning has achieved enormous success in object detection. Compared to other detection models, the single shot multiBox detector (SSD) has the advantages of high detection accuracy with relatively high speed, which is suitable for inferences in a large volume of SAR images. This paper focuses on the application of the SSD to address ship detection in complex backgrounds. Due to a limited number of datasets, transfer learning is adapted to train the model. Ten scenes of Sentinel-1 SAR images are used to evaluate our approach. Experimental results reveal that 1) transfer learning improves the detection accuracy and overall performance and reduces the false positives, and 2) compared with the faster RCNN and other SSD models, the SSD-512 model with transfer learning achieves the best overall performance, which demonstrates the effectiveness of our approach.