PTIDNet: a lightweight convolutional network with channel-split self-attention for insulator defect detection
Abstract
Overhead-line insulators are constantly exposed to harsh weather, and undetected defects, such as self-explosion-induced disc loss, can trigger wide-area power outages. However, insulator defects captured by unmanned aerial vehicles (UAVs) typically suffer from strong background interference, large scale variation, and extremely low target-pixel ratios, so existing lightweight detectors struggle to control theoretical computational complexity while maintaining accuracy. Building on YOLOv8n, this paper proposes PTIDNet, a lightweight detector. A Multi-Path Hybrid Block (MPHB) is designed at the top of the backbone, splitting channels between a convolutional path and an attention path processed in parallel. An Adaptive Multi-Path Pooling module (AMPM) employs dynamic kernels to reduce information loss for small targets during downsampling. A Context-Aware Fusion Network (CAFN) encodes features along separate “defect-region” and “surrounding-context” paths before fusing them. On the public CPLID-D and CPLID-N insulator defect datasets, PTIDNet improves mAP50-95 over the YOLOv8n baseline by 6.4 and 10.3 percentage points, respectively, with only 2.48M parameters; comparisons with MobileNet-SSD, EfficientDet-D0, and a channel-attention-enhanced YOLOv8n, together with a systematic report of parameter count, GFLOPs, and model size (MB) as theoretical complexity indicators, further indicate that PTIDNet achieves a good balance between accuracy and lightweight design.