Enhanced safety risk detection in power operations via scene augmentation and patch-wise knowledge distillation

Real-time detection of operational behaviors and personal protective equipment (PPE) in on-site power operation scenarios is essential for reducing safety hazards. Compared with generic image benchmarks, on-site surveillance faces cluttered backgrounds, crisscrossing transmission lines, high-contrast vegetation, and extreme scale variations where PPE objects occupy only a few pixels in wide-area views. To enhance safety risk detection in power operations, we propose a method combining scene augmentation (SA) and patch-wise knowledge distillation (PWKD). SA simulates diverse environmental shifts and enhances structural and textural cues for robust scene-level and content-level augmentation. PWKD further transfers knowledge in a patch-wise manner to mine fine-grained features from complex backgrounds and improve small-object detection. Compared to the baseline, our method improves mAP by 4.9 and 4.8 points on the S-PPE and PCSM datasets, respectively, and achieves gains of 4.9 and 5.5 points in the small-object metric.