A Robust Method for Extracting Anti-Loosening Lines for Turbine Unit Bolts Under Multi-Factor Influences

Dynamic inspection of hydropower-generator rotors is hindered by long imaging distances, metallic reflections, and motion-induced degradation of anti-loosening marking lines. We present a task-decoupled vision framework that first localizes bolts and then segments their marking lines within scale-normalized regions of interest. The YOLO-SE detector combines spatial-to-depth convolution with efficient multi-scale attention to preserve small-object features and suppress illumination interference. The segmentation stage integrates a MobileViT2 backbone, attention-enhanced atrous spatial pyramid pooling, and Dice cross-entropy loss to recover global context and sparse foreground boundaries. In laboratory experiments, YOLO-SE achieved 99.4% mean average precision at 0.5 intersection-over-union (mAP50) and 99.8% recall, exceeding YOLO26n by 3.2 and 3.4 percentage points, respectively. The complete segmenter achieved 98.67% mean intersection-over-union (mIoU) and 95.38% target-line IoU; across five seed-matched runs, YOLO-SE obtained 99.32 ± 0.10% mAP50 and the segmenter obtained 98.61 ± 0.09% mIoU. Site-specific validation on M12 rotor-clamp bolts supported field applicability under changes in scale, viewpoint, and illumination. The complete cascade required 53.1 million parameters, 207.7 GFLOPs, 2.38 GB peak GPU memory, and 31.8 ms per frame, meeting a 30 FPS GPU budget under the evaluated conditions.