Experimental study demonstrates enhanced surface defect detection on transmission towers using YOLOv11-TA-Convformer, indicating improved inspection speed and reduced maintenance costs.
Key Points
Develop an enhanced single-stage object detection model to accurately detect structural surface damages on transmission towers amid complex backgrounds and small-target constraints.
Constructed the YOLOv11-TA-Convformer network by replacing the backbone C3K2 module with Convformer and incorporating Triplet Attention to strengthen critical feature extraction.
Developed an augmented image dataset encompassing four damage categories: pier cracks, steel corrosion, missing bolts, and structural steel cracks.
Evaluated model optimization using SGD, Adam, and RMSProp optimizers across varied learning rate strategies and compared performance with SENet, CBAM, and state-of-the-art baselines.
Stochastic gradient descent (SGD) delivered the most rapid convergence and the lowest final loss during training.
The YOLOv11-TA-Convformer architecture achieved a mean average precision (mAP) of 0.873, outperforming all baseline models and alternative attention mechanisms.
Field engineering deployment increased overall inspection efficiency by 2.8 times and lowered annual maintenance costs by 17.3%.