A real-time deep learning network for ship detection in SAR images

In the field of ship detection in synthetic aperture radar (SAR) images using deep learning, traditional models often face challenges related to their complex structure and high computational demands. To address these problems, this paper proposes a lightweight end-to-end convolutional neural network called BGD-Net, which is based on anchor-free methods. This network incorporates a novel feature pyramid called Ghost, Cross-Stage-Partial, and Path Aggregation Network (G-CSP-PAN) to enhance the detection performance across targets of varying scales. Additionally, it introduces an efficient decoupled detection head, termed the efficient decoupled head (ED-Head), to enhance the interaction between regression and classification. Furthermore, an optimized loss function named optimized efficient intersection over union (OEIoU) loss is proposed for edge regression. The proposed method is evaluated on public datasets, SSDD and SAR-Ship-Dataset, demonstrating a balance between detection accuracy and efficiency.