SKCSP-YOLOv5: a multi-scale feature enhancement network with attention for bubble detection in gas–liquid two-phase flow

Accurate bubble detection in rod bundle two-phase flow is essential for nuclear reactor safety, underpinning void fraction estimation and thermal–hydraulic model validation, yet existing YOLO-based detectors struggle to jointly resolve small, densely packed, and deformable bubbles. This paper proposes SKCSP-YOLOv5, which integrates Selective Kernel Cross-Stage-Partial (SKCSP) blocks, a Convolutional Block Attention Module (CBAM), and an additional P6/64 detection head that provides a coarse-scale prediction level for the largest individual bubbles, while retaining the same single-bubble detection target across all four detection heads. Across five independent training seeds, SKCSP-YOLOv5 achieves a mean recall of 86.5% and mAP@0.5 of 91.0%, significantly outperforming parameter-matched YOLOv5m, YOLOv8m, and YOLOv12m baselines (19.6–25.9 M parameters; p < 0.001, Cohen’s d > 6) while requiring 25.4 GFLOPs, roughly a third to a half the computational cost of these baselines. Performance is further characterized by bubble size, local density, and occlusion level, together with counting and sizing error, to assess suitability for quantitative bubble characterization beyond aggregate detection accuracy. These results establish SKCSP-YOLOv5 as an accurate, computationally efficient, and statistically validated foundation for automated bubble characterization in nuclear reactor subchannel flows.