HAFusionNet: Visible-Thermal Image Fusion with Cross-Modal Interaction for High-Altitude Substation Defect Detection

Defect detection in high-altitude substations is severely complicated by harsh environmental conditions such as low oxygen, strong ultraviolet radiation, and extreme temperature variations, which markedly limit the reliability of traditional single-modal methods. To address these issues, we propose HAFusionNet, a multimodal defect detection network that fuses visible and thermal imagery for robust equipment inspection in extreme environments. The network is built around three complementary modules: a Gated Cross-Modal Interaction (GCMI) module that performs attention-based, gated feature exchange between the two modalities; an Edge-Guided Difference Gated Fusion (EDGF) module that emphasises modality-complementary differences and boundary cues; and a lightweight Attention-guided Spatial Reweighting (ASR) module for efficient spatial feature refinement. These modules are embedded in a dual-stream backbone with enhanced multi-scale fusion. On a self-constructed high-altitude substation defect dataset, HAFusionNet attains an mAP50 of 0.956, and an mAP50:95 of 0.742, outperforming existing multimodal detectors while remaining efficient. This work offers an effective and robust solution for intelligent defect detection of power equipment in extreme high-altitude environments.