Two-stage ship detection in synthetic aperture radar images based on attention mechanism and extended pooling

Abstract

Object detection in synthetic aperture radar (SAR) images remains a challenging problem due to the particular imaging mechanism of SAR systems. The sizes of targets are relatively small and the scenes are large, indicating that the intersection over union value between the targets and anchors is probably small. In addition, SAR images are severely polluted by speckles under normal conditions. The edges of objects in SAR images are blurred. In this study, we proposed an improved object-detection framework based on a two-stage faster region-based convolutional neural network. A feature-enhancement module based on self-attention mechanism is designed to learn a target-oriented feature map, where the spatial attention and channel attention work simultaneously, such that targets can be highlighted and the speckles can be suppressed to a certain extent. Two separate feature maps serve as the input of the region proposal network to isolate the classification and regression tasks. Because only one category exists in the ship-detection task, correctly distinguishing targets from the background results in the correct detection of ships. In the final classification stage, an extended region-of-interest pooling operation is performed on the potential proposals and contexture information. The usage of extra information can improve target fine-tuning in the final network. To avoid ignoring small targets, we carefully set the anchors’ parameters based on the analysis of ground truth and select the appropriate shapes of feature maps. With the help of these modifications, the proposed method can detect small, weak, and dense targets in SAR images. Ablation experiments over networks with different configurations prove that the proposed modules are working. Experiments on real SAR Gaofen-3 and Sentinel-1 images demonstrate the efficiency of the proposed object-detection framework.