Context-Preserving Instance-Level Augmentation and Deformable Convolution Networks for SAR Ship Detection
Shape deformation of targets in SAR image due to random orientation and partial information loss caused by occlusion of the radar signal, is an essential challenge in SAR ship detection. In this paper, we propose a data augmentation method to train a deep network that is robust to partial information loss within the targets. Taking advantage of ground-truth annotations for bounding box and instance segmentation mask, we present a simple and effective pipeline to simulate information loss on targets in instance-level, while preserving contextual information. Furthermore, we adopt deformable convolutional network to adaptively extract shape-invariant deep features from geometrically translated targets. By learning sampling offset to the grid of standard convolution, the network can robustly extract the features from targets with shape variations for SAR ship detection. Experiments on the HRSID dataset including comparisons with other deep networks and augmentation methods, as well as ablation study, demonstrate the effectiveness of our proposed method.