SAR Ship Target Detection for SSDv2 under Complex Backgrounds
Yuan Chen, Jie Yu, Xu Yang · 2020
With the launch of space-borne satellites, more synthetic aperture radar (SAR) images are available than ever before, thus making dynamic ship monitoring possible. Object detectors in deep learning achieve top performance, for example, Faster R-CNN, ResNet, FPN, YoLo and SSD Net. Compared with YoLo, SSD algorithm performs better in accuracy and speed, but it is not good at small targets detection like SAR images. To solve this problem, we proposed an improved method. We added a deconvolution module and prediction module on the basis of SSD. The deconvolution module is mainly used to integrate the high-level semantic information into the feature information of the low-level network so as to improve the detection accuracy. The prediction module, which is composed of residual network, can extract depth features and input them into regression task and classification task., we call it SSDv2. We test our model in the SAR-Ship-Dataset. SAR-Ship-Dataset was created using 102 Chinese Gaofen-3 images and 108 Sentinel-1 images. It consists of 43,819 ship chips of 512 pixels in both range and azimuth. These ships mainly have distinct scales and backgrounds. The result shows that our SSDv2-512*512 achieves 91.07% mAP, outperforming a state-of-the-art method SSD.