Panoramic Symbol Spotting for Power Design CAD Drawings Based on Primitives and Text Fusion with Locally Sensitive Self-Attention

Panoramic Symbol Spotting (PSS) for power design CAD drawings is crucial for the advancement of intelligent power systems. Existing PSS methods struggle with the large number and uneven sizes of graphical elements in power CAD drawings and fail to leverage text annotations effectively. To address these issues, we propose the PFL-Net model, which utilizes locally sensitive re-ordering and cross-covariance self-attention mechanisms to reduce attention complexity, thereby improving performance on large-scale power design drawings. We also introduce a bilinear pooling method to fuse text and graphical primitive information. Experiments demonstrate that PFL-Net outperforms existing methods in both accuracy and speed, making it highly applicable to power design scenarios.