Soft Thresholding Attention Network for Adaptive Feature Denoising in SAR Ship Detection

Recently, synthetic aperture radar (SAR) ship detection is used in many applications within the marine field, such as fishery management, traffic control, and urgent rescue operations. Meanwhile, deep learning-based methods have bought new capabilities for ship detection in SAR images on account of high accuracy and robustness. However, several challenges remain to be addressed: 1) the shapes of the ships in SAR images have a relatively extreme aspect ratio comparing to the target objects in the optical images, and 2) complex background and clutter noise result in adverse effects for the network to extract prototypical SAR target features, which limit the ship detection performance. To address these issues, this paper proposes two effective approaches to augment the feature extraction ability of the network. Firstly, IOU (Intersection over Union) K-means is carried out to settle the extreme aspect ratio problem. The IOU K-means, as a preprocessing step, clusters a set of aspect ratios from datasets that are suitable for ship detection. Secondly, we embed a soft thresholding attention module (STA) in the network to suppress the impact of noise and complex background. The comparison results with several state-of-the-art object detection algorithms confirm the efficiency and feasibility of proposed approaches.