AMR-Net: Arbitrary-Oriented Ship Detection Using Attention Module, Multi-Scale Feature Fusion and Rotation Pseudo-Label

Abstract

Ship detection is significant and full of challenges in the field of remote sensing. The widely adopted horizontal bounding box representation is not appropriate for ubiquitous oriented ship objects. Complex backgrounds, small objects, and the dilemma of labeling ship datasets are all obstacles that further limit the successful operation of traditional methods in ship detection. In this paper, we propose a multi-task rotation detector using attention module, multi-scale feature fusion and rotation pseudo-label, namely AMR-Net. AMR-Net adds Deformable Convolution Channel Attention Block (DCCAB) to suppress background noise and highlight the foreground. Feature Pyramid Network (FPN) fuses features from different scales, which is beneficial for ship detection. We present Adaptive FPN (AFPN) to determine the optimal number of prediction layers automatically, which can reduce the disturbance of high-level detection of small objects, decrease the size of the model, and adapt to different ship datasets. To enable the detector to achieve excellent performance even when using fewer labeled data, a semi-supervised pseudo-label module is designed, namely Self-Learning Rotation Pseudo-Label (SRP). SRP allows the detector to iteratively self-learn the optimal thresholds, and use the thresholds to refine high-quality rotation pseudo-labels for retraining the model. SRP is optional. If this module is selected, the detector becomes a semi-supervised detector. Extensive both supervised and semi-supervised experiments on remote sensing public datasets HRSC2016 show the state-of-the-art performance of our detector. Experiments on DOTA further illustrate the effectiveness of AMR-Net.