SCR-YOLO: a small-object detection model for beach litter with spatial and contextual refinements

Detecting small-scale litter in beach environments remains challenging due to weak visual cues, complex backgrounds, and significant scale variations. Existing detectors suffer from information loss during downsampling and insufficient contextual modelling, leading to missed detections of small objects. To address these issues, this paper proposes SCR-YOLO, an improved framework built upon YOLOv8 with targeted spatial and contextual refinements. A Space-to-Depth Convolution (SPDConv) preserves fine-grained spatial information during downsampling, while a Large Selective Kernel Block (LSKblock) dynamically expands the effective receptive field and adaptively fuses multi-scale contextual features through spatially weighted selection, enabling more discriminative representation under complex backgrounds. A dedicated small-object detection head further exploits high-resolution features. Evaluated on a self-constructed 11-category beach litter dataset, SCR-YOLO outperforms YOLOv8 and other state-of-the-art methods, achieving gains of 3.1% in mAP50 and 4.9% in mAP50-95, demonstrating its effectiveness within the evaluated beach scenarios.