Precise segmentation method for slender power targets based on multi-scale perception and location-sensitive learning

In power systems, precise wire segmentation holds significant technical value for ensuring the safe operation and maintenance of electrical facilities, yet achieving high-precision and efficient intelligent segmentation under complex scenarios involving background interference and significant target scale variations remains challenging. This paper proposes MSHNet, a novel segmentation architecture integrating multi-scale head networks with scale- and location-sensitive loss functions, based on multi-scale feature analysis and location-aware learning theories. It overcomes the limitations of conventional segmentation networks in perceiving geometric features and modeling spatial contextual relationships. The core innovation lies in synergizing physical-space multi-scale feature fusion mechanisms with deep neural network adaptability. By constructing scale-aware head modules and position-sensitive loss functions, it unifies multi-scale spatial information decoding with pixel-level feature optimization, achieving refined wire segmentation and semantic understanding in complex scenarios. Key contributions include: (1) Enhanced U-Net-based multi-scale head network design; (2) Geometry-aware scale- and location-sensitive loss function development; (3) Co-optimization of encoder-decoder networks and feature analysis modules. An intelligent experimental platform for power line segmentation is established to realize precise wire extraction and semantic segmentation detection in power transmission/distribution scenarios. This research expands the theoretical framework of electrical vision inspection and provides innovative technical solutions for intelligent facility maintenance in complex environments.