Vision Hypergraph Feature Enhancement and Task Decoupling Network for SAR Ship Detection

Synthetic aperture radar (SAR) provides significant advantages for ship detection and has been extensively applied. In recent years, numerous one-stage networks for SAR ship detection have been developed as the result of the rapid advancement of deep learning. However, despite the remarkable success of these detectors, they still encounter several critical challenges. First, the complex inshore backgrounds and the multiscale of ships persistent challenges for SAR ship detection. Second, the independence of regression and classification tasks in most one-stage networks constrains further enhancements in detection accuracy. To address these issues, a vision hypergraph feature enhancement and task decoupling network (VHTDNet) is proposed. Specifically, the vision hypergraph computation is applied in neck to enhance feature representation. Furthermore, a channel-spatial joint task decoupling head is proposed for regression and classification tasks. This detection head includes a dynamic feature interaction and task decoupling module to facilitate feature interaction and task-specific decoupling in the channel dimension, as well as spatial offset and spatial focus modules to enhance decoupled features in the spatial dimension by using interactive features. Comprehensive experimental results on two public SAR ship datasets indicate the performance and generalization ability of VHTDNet.