FUSEPOINT: a symbol recognition method for defective CAD drawings

The digitalization of aged or poorly scanned CAD drawings is challenging due to damaged elements and inaccurate information, which compromises the performance of existing symbol recognition methods. To address this, we propose FusePoint, a novel recognition framework that treats symbols as point clouds. FusePoint introduces two innovations: 1) the Point Repair and Consistency Enhancement (PRCE) network, a preprocessing module that repairs and normalizes degraded point cloud data, and 2) the Context-Aware and Multi-Scale Fusion (SACF) strategy, which uses attention mechanisms to extract robust features from noisy and incomplete data. Comprehensive experiments on the public FloorPlanCAD benchmark and a self-built AgedCAD dataset with real-world damage show that FusePoint outperforms state-of-the-art methods. It demonstrates significantly improved recognition accuracy and robustness, especially on imperfect drawings, offering an efficient solution for automated processing of non-standard CAD files in engineering applications.