A Lightweight Radar Ship Detection Framework with Hybrid Attentions

One of the current research areas in synthetic aperture radar (SAR) processing fields is deep learning-based ship detection SAR imagery. Recently, images has achieved continuous breakthroughs precision. However, determining how to strike a better balance between precision and complexity algorithm very meaningful for real-time object real application scenarios, attracted extensive attention from scholars. In this paper, lightweight framework named multiple hybrid attentions detector (MHASD) with mechanisms proposed. It aims reduce without loss First, considering that features are not inconspicuous compared other images, residual module (HARM) developed deep-level layer obtain rapidly effectively via local channel parallel self-attentions. Meanwhile, it also capable ensuring high model. Second, an attention-based feature fusion scheme (AFFS) proposed model neck further heighten object. AFFS constructs develops fresh (HAFFM) upon spatial guarantee applicability The Large-Scale Ship Detection Dataset-v1.0 (LS-SSDD-v1.0) experimental results demonstrate MHASD can speed (improving average by 1.2% achieving 13.7 GFLOPS). More importantly, experiments on Dataset (SSDD) method less affected background such as ports rocks.