Transmission Line Equipment Multilabel Classification Based on Dual-Mask Transformer

Transmission lines (TLs), as the crucial link between power generation and consumption, constitute an indispensable component of modern power systems. Conducting multi-label classification (MLC) on TL images facilitates the extraction of equipment information across diverse scenarios, aiding the implementation of downstream tasks such as target detection and image segmentation. However, occlusions between different equipment, inconsistent appearances, and confounding inter-class relationships significantly impede the accuracy of MLC. In this paper, we propose a novel approach for MLC of TL equipment based on dual-mask Transformer. This method takes both images and labels as joint inputs and utilizes a Transformer to learn the complex dependencies between visual features and labels. Before being fed into the network, image and label masks are designed based on the characteristics of TL images. Specifically, the image mask simulates real-world occlusion issues of equipment, thereby enhancing the model’s robustness, while the label mask strengthens the semantic correlations between equipment and guides the network in updating image features during training. Furthermore, the network adopts a collaborative training strategy by incorporating a branch network onto the main network. This branch network captures different spatial regions occupied by objects from various categories through a class-specific residual attention mechanism, thus improving the quality of basic visual features. Experimental results demonstrate that the proposed method outperforms mainstream MLC methods on TL images. Our source code is available at https://github.com/AHU-psy/-Dual-mask-Transformer.