Semantic-guided adaptive dual-stream inference for marine floating litter detection

Marine floating litter detection on autonomous edge platforms, such as Unmanned Surface Vehicles (USVs), faces a fundamental scale-compute conflict: downsampling can obscure the fine-grained features of micro-scale targets, while native high-resolution inference often exceeds the computational budgets of edge platforms. Uniform slicing recovers small-scale target recall but exhaustively processes empty sea surfaces, amplifying false positives from sun glint and sea foam and substantially increasing inference latency. We propose GL-DSI, a Global-Local Dual-Stream Inference framework that replaces blind slicing with semantic-guided adaptive extraction: a low-resolution global stream localizes suspect regions, triggering high-resolution local inference only where small-scale targets likely exist. A Size-Aware Gaussian Attenuation (SAGA) module suppresses boundary artifacts, while the Scale-Semantic Consistency Filter (SSCF) removes environment-induced pseudo-targets. On our marine floating litter dataset (MFL-Dataset), GL-DSI with YOLOv11s reaches an F1-score of 93.4% at 44.6 frames per second (FPS) on the desktop benchmark. Compared to its standard baseline, it improves the detection of highly degraded small-scale targets, boosting recall for targets with areas below 256 px² by 11.0 %age points and lifting nano-scale target recall from 13.5% to 27.0%, while remaining substantially less computationally intensive than exhaustive slicing.