A Multi-Modal Few-Shot Learning Framework for Foreign Object Segmentation in GIS Inspection

The reliable operation of Gas-Insulated Switchgear (GIS) is crucial for power system safety, yet automatic foreign object inspection within its cavities remains challenging due to low-light conditions and strong reflections. This paper proposes a multi-modal few-shot learning framework for high-precision foreign object segmentation in GIS. To overcome imaging interference, we first establish a dual-light (visible and ultraviolet) image acquisition system and design a lightweight fusion network to adaptively integrate multi-modal features, enhancing scene representation. For the core few-shot segmentation task, we introduce a novel Multi-Similarity Guided Branch Network (MSBNet). This network employs a support-query dual-branch architecture to extract sample prototypes. It features an improved background similarity guidance mechanism to suppress base-class feature interference and a multi-similarity fusion module that synergistically integrates multi-level and multi-metric information, which significantly improves the continuity and boundary accuracy of the segmentation masks. Experiments on our GIS dataset demonstrate that, under extremely limited sample conditions, the proposed method rapidly adapts to unseen foreign object classes and substantially outperforms existing few-shot segmentation baselines.