Semantic segmentation in power grid scenarios using scale-transforming transformer

Semantic segmentation of power grids is challenging due to size variations and intricate deformations caused by different shooting distances and angles. Traditional hierarchical architectures and pyramidal methods can learn multi-scale features to address size variations but struggle with deformations due to fixed aspect ratios of features. To address this issue, we propose a scale-transforming transformer (STT) approach. Our approach’s novelty lies in a scale-transforming module (STM), which implements cost-effective aspect ratio adjustments, patch splitting, and patch combining. This process generates local patches comprising various versions of the original patch, characterized by distinct aspect ratios and scales. In particular, this approach ensures that the output and input feature maintain uniform dimensions. We also control computational loads through a channel grouping strategy, which deploys different STMs in distinct feature groups. Consequently, our STM seamlessly integrates into existing transformer models to build STT models. Experiments show that our STT models achieve state-of-the-art performance.