Detection of Power Line Insulator Defects Using Aerial Images Analyzed With Convolutional Neural Networks

Proposes a cascaded CNN that first locates insulators in aerial images and then detects defects within them, splitting the task into two detection levels. Earlier hand-crafted or shallow methods only worked under narrow conditions given cluttered aerial backgrounds. Scarcity of defect images was addressed with affine transforms, segment-and-paste onto new backgrounds, Gaussian blur and brightness changes.

Key findings

• Precision 0.91 and recall 0.96 on a standard insulator dataset
• Synthetic augmentation (insulator segmentation and background fusion) used to overcome scarce defect images

Limitations

• Abstract only reviewed
• Evaluation on one dataset; defect type mainly missing insulator caps
• Heavy reliance on synthetic augmentation may not reflect real defect variety