A convolutional and variant visual transformer fusion network for small ship detection in SAR images

Synthetic Aperture Radar (SAR) has been widely recognized as an outstanding method for marine monitoring in recent years. With the rapid development of SAR technology and image processing field, significant progress has been made in ship detection techniques based on SAR images. However, detecting small target ships in SAR images remains a major challenge. In this study, we propose a convolutional and variant visual transformer fusion network specifically for small ship detection in SAR images. Primarily, we propose integrating a BiFormer-block at the end of the backbone network to enhance the network's feature extraction capability for small target ships. In addition, we propose a novel NormConv FPN, which for the first time adds new feature fusion layers and detection heads above the shallow layers to maximize retention of small object feature information, and we incorporate normalization-based Attention Modules during feature fusion to ensure focus on critical information. To validate the performance of the method in this paper, experiments were conducted on the SSDD dataset and compared with some state-of-the-art methods, obtaining the highest MAP (Mean Average Precision) score, 97.5%. In addition, ablation experiments were performed to validate the effectiveness of each enhancement element separately. Our code is available at https://github.com/shengyexu/BiFormer-Enabled-YOLOv5 (accessed on 10 August 2023).