An Anchor-Free Method Based on Transformers and Adaptive Features for Arbitrarily Oriented Ship Detection in SAR Images
Bingji Chen, Chunrui Yu, Shuang Zhao, Hongjun Song
Abstract
Ship detection is a crucial application of synthetic aperture radar (SAR). Most recent studies have relied on convolutional neural networks (CNNs). CNNs tend to struggle in gathering adequate contextual information through local receptive fields and are also susceptible to noise. Inshore scenes in SAR images are plagued by substantial background noise, so achieving high-accuracy ship detection of arbitrary orientations within complex scenes remains an ongoing challenge when relying solely on CNNs. To address the above challenges, this paper presents an anchor-free method based on transformers and adaptive features, namely, SAD-Det, which can detect rotationally invariant ship targets with high average precision in SAR images. Specifically, a transformer-based backbone network called the ship spatial pooling pyramid vision transformer (SSP-PVT) is proposed to enhance the long-range dependencies and obtain sufficient contextual information for ships in SAR images. In addition, a neck network called the adaptive feature pyramid network (AFPN) is designed to enhance the ability of ship feature adaptation by adding fusion factors to feature layers in SAR images. Finally, a head network called the deformable head (DeHead) is constructed to make the network more adaptable to the characteristics of ships by adaptively detecting the spatial sampling positions of the targets in SAR images. The effectiveness of the proposed method is verified by experiments on two publicly available datasets, i.e., SAR ship detection dataset (SSDD) and rotated ship detection dataset in SAR images (RSDD-SAR). Compared with other arbitrarily oriented object detection methods, the proposed method achieves state-of-the-art detection performance.