MDFNet: a lightweight multi-scale dynamic fusion network for visual measurement of micro-defect in UAV power line inspection

Abstract

The safe and stable operation of transmission lines is essential to the reliable operation of power grids. UAV-based visual inspection has emerged as an efficient paradigm for transmission-line inspection. However, UAV-based visual inspection still faces three major challenges: tiny components/defects, severe background clutter, and strict latency constraints due to limited onboard computing resources. In this paper, we propose MDFNet, a lightweight multi-scale fusion attention network for transmission-line inspection. Specifically, we first design a Multi-Receptive Field Aggregation (MRFA) module to aggregate multi-receptive-field responses, thus preserving fine structural cues for tiny targets. Then, a Global-Context Fusion Module (GCFM) is developed, which performs lightweight cross-scale context modeling via linear attention to enhance interactions among multi-level features. Finally, we design DAIM with parallel global and local attention to emphasize foreground components and suppress background interference. Extensive experiments on three benchmark datasets (Custom, InsPLAD, and CPLID) demonstrate that MDFNet consistently outperforms state-of-the-art methods, achieving significant gains of 13.5%, 8.1%, and 13.2% in small-object Average Precision (APS). With a compact footprint of 18.6 M parameters and an inference speed of 126.1 FPS, the proposed method offers an accurate and computationally efficient measurement solution for real-time power grid infrastructure monitoring