Abstract: View recognition of distribution network towers is a key technology in UAV intelligent inspection. To address the problem of low accuracy of existing deep learning methods in complex background interference, this paper proposes a tower view classification method based on EfficientNet that integrates foreground perception, multi-scale feature fusion, and dual-dimensional attention. First, a Mask-Guided Fusion Module (MGFM) is designed to extract tower foreground masks using the BiRefNet network, enhancing foreground representation and suppressing background interference through a two-stage fusion strategy. Second, a Multi-Scale Attention Aggregation Module (MSAA) is constructed to achieve efficient cross-layer feature fusion through parallel multi-scale convolution, fully integrating shallow details and deep semantic information. Finally, the Convolutional Block Attention Module (CBAM) is introduced to adaptively strengthen view-discriminative features through channel and spatial dual-attention mechanisms, significantly improving the recognition capability for small-sample categories such as top views. Ablation experiments on a self-built multi-view tower dataset show that the proposed method can effectively distinguish different views such as top view, front view, and side view, with significantly improved accuracy compared to other deep learning models, providing technical support for intelligent inspection of transmission lines.

Gao Liu, Changyu Li, Junsheng Lin, Xinzhe Weng, Qianming Wang and Zhenbing Zhao. “A Multi-View Classification Method for Distribution Network Towers Based on Improved EfficientNet”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.12 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0161228

@article{Liu2025,
title = {A Multi-View Classification Method for Distribution Network Towers Based on Improved EfficientNet},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2025.0161228},
url = {http://dx.doi.org/10.14569/IJACSA.2025.0161228},
year = {2025},
publisher = {The Science and Information Organization},
volume = {16},
number = {12},
author = {Gao Liu and Changyu Li and Junsheng Lin and Xinzhe Weng and Qianming Wang and Zhenbing Zhao}
}