Combing Single Shot Multibox Detector with transfer learning for ship detection using Chinese Gaofen-3 images

Abstract

SAR images have the advantages of all-whether and all-time, leading to be increasingly used for ship detection to ensure marine transportation and security. Single Shot MultiBox Detector (SSD) has the advantages of fast speed and high accuracy, achieving great success in optical images. In this paper, SSD is applied to ship detection in SAR images. Besides, transfer learning is adopted because it performs well even in small training dataset. More specific, two types of SSD models integrated with transfer learning, namely SSD-300 and SSD-512 with input size of 300 and 512 in both height and width respectively, are applied to ship detection. To evaluate our approaches three SAR images acquired by Chinese Gaofen-3 satellite are used. First, SAR images are cut into patches with size of 256, and then these small images are divided into training, validation, and test with the ratio 70%, 20% and 10% according to machine learning routines. Secondly, both training and validation dataset are resized to feed to SSD with fine tuning VGG 16 to train the model. Finally, test images are used to evaluate our approach. Experimental results reveal that compared to SSD-300, SSD-512 achieves more than 0.02 in probability of ship detection, whereas 0.05 worse in false alarm.