Comparative Analysis of Image Segmentation Methods in Power Line Monitoring Systems

This study presents a comparative study of image segmentation methods for power line monitoring using unmanned aerial vehicle (UAV) imagery. The study investigates whether a hybrid segmentation pipeline—combining classical image processing and deep learning—can enhance defect detection under challenging conditions. Three traditional methods (Otsu thresholding, Canny/Sobel edge detection, and k-means clustering) and three deep learning models (UNet, DeepLabv3, Mask R-CNN) were evaluated on a dataset of annotated UAV images, including real and synthetically augmented scenes with lighting, noise, and weather variations. Performance was measured using Intersection over Union (IoU), Pixel Accuracy, and processing time.Traditional methods demonstrated fast inference (0.2-0.4) but limited accuracy (IoU 0.47–0.58; Accuracy 72.5 –82%). Deep learning models significantly outperformed them: UNet achieved 0.85 IoU and 94% accuracy; DeepLabv3 reached 0.88 IoU and 96%; and Mask R-CNN led with 0.90 IoU and 97% accuracy, though at 1.2 seconds per image. The proposed hybrid method combines classical preprocessing for region-of-interest extraction with Mask R-CNN segmentation, achieving 0.89 IoU, 96.5% accuracy, and reduced processing time (0.75 s/image), improving speed by 30% with minimal accuracy loss. Robustness tests showed deep learning and hybrid methods degraded less than 6% under noise, compared to 20% for traditional methods. The results demonstrate that hybrid segmentation provides a practical balance between accuracy and efficiency, suitable for real-time monitoring on resource-limited platforms.