SAC-YOLO: Efficient Multi-Scale Feature Fusion for Transmission Line Defect Detection

Why It Matters

Accurate, real‑time defect detection reduces outage risk and maintenance costs for power utilities, while the edge‑ready design enables scalable deployment across remote transmission networks.

Key Takeaways

• Replaces SPPF with AIFI for better small defect representation
• Introduces C3K2‑CFBlock to blend local and global features efficiently
• SGMF adds channel‑aware reweighting and bidirectional fusion
• Boosts recall 3.5% and mAP@0.5 to 0.95, a 3.8% gain
• Runs at 277 FPS, enabling real‑time edge deployment

Pulse Analysis

Transmission line operators face mounting pressure to detect minute defects—such as broken conductors or insulator contamination—before they trigger costly failures. Traditional visual inspections are labor‑intensive and prone to human error, especially in complex, cluttered environments. Deep‑learning detectors like YOLO have shown promise, yet standard architectures struggle with the tiny, low‑contrast targets typical of power‑line assets. By focusing on multi‑scale feature fusion and lightweight design, SAC‑YOLO addresses these gaps, offering a practical path to automated, on‑site monitoring.

The core innovations of SAC‑YOLO revolve around three custom modules. The AIFI block replaces the conventional SPPF layer, enriching high‑level semantic maps with intra‑scale context while adding negligible latency. The C3K2‑CFBlock restructures the backbone to merge convolutional detail with global cues, strengthening long‑range dependencies essential for spotting dispersed anomalies. Finally, the Semantic‑Guided Multi‑Scale Fusion (SGMF) module applies channel‑aware reweighting and bidirectional guidance, overcoming the blunt aggregation of classic FPNs and sharpening the model’s focus on small‑scale defects. Collectively, these tweaks elevate feature representation without inflating the parameter count.

Empirical results validate the approach: recall improves by 3.5% and mAP@0.5 climbs to 0.95, a 3.8% uplift over the baseline YOLOv11n, all while sustaining 277 frames per second. This performance level meets the stringent latency demands of edge devices mounted on drones or mobile inspection rigs. For utilities, the technology translates into faster fault identification, reduced downtime, and lower inspection labor costs. Moreover, the model’s modest computational footprint facilitates scaling across vast networks, positioning SAC‑YOLO as a catalyst for smarter, more resilient grid operations.