SegS-YOLO: A YOLO-Based Instance Segmentation Approach for Small-Scale Ship Targets in SAR Images

Existing target recognition and segmentation methods based on Synthetic Aperture Radar (SAR) images often overlook the unique characteristics of small-scale vessels. Especially when sea clutter and speckle noise are present in SAR images, the segmentation performance of existing networks is not ideal when dealing with images containing a large number of small-scale ship targets. This paper presents an improved model-Segment Small-You Only Look Once (abbreviated as SegS-Yolo). While maintaining a strong segmentation ability for multi-scale ship targets, it enhances the segmentation capacity for small-scale ships in SAR images. This network optimizes the core architecture of YOLOv8-Seg. Principally, the newly designed SubImg-FADNet enhances the backbone's ability to retain redundant ship detail features, thus reducing the risk of losing these details. Additionally, it adopts a redesigned Multi-dimensional Feature Fusion Pyramid (MF-FPN) structure as the network's neck to enhance the network's feature fusion capabilities. To improve the network's segmentation accuracy for small-scale ship targets, we incorporate an enhanced SIF-C2MA feature fusion module in the backbone network and the Dynamic Snake Convolution PKI Block (DSCPKIBlock). Moreover, we also utilize a lightweight upsampling operator (CARAFE), a Sub-image Deep Feature Fusion module (SIF), a Cross Stage Partial Stage (CSPStage), and a CGA-based single-level and multi-channel fusion scheme (CGA-SMFusion) to construct a brand-new Multi-dimensional Feature Fusion Pyramid (MF-FPN) neck structure. The effectiveness of this network has been verified on the SSDD dataset, indicating that it exhibits good segmentation accuracy in the experimental environment. Meanwhile, we tested its generalization ability on the LS - SSDD dataset, which is dedicated to small - target detection (lacking instance annotations), and the high - resolution HRSID dataset. The experimental results show that SegS - Yolo outperforms other competing models. Its Average Precision (AP50) is 5% higher, and the Average Precision for Small targets (APS) is 5.6% higher, demonstrating its superiority over other models.