Floor Plan Semantic Segmentation Network with Gated Channel-Spatial Fusion and Explicit Boundary Constraint

To address challenges in floor plan segmentation, specifically structural disconnection caused by insufficient long-range dependency capture, and semantic region overflow resulting from inadequate geometric-semantic fusion, an Explicit-boundary Dual-stream Attention Network (EDANet) is proposed. The architecture employs DenseNet121 as the backbone, integrating a Strip Pooling module to establish global context dependencies for anisotropic structures such as corridors. A spatial attention dual-stream decoder is constructed to suppress background noise in skip connections, enhancing edge recovery precision. Furthermore, a Gated Channel-Spatial Fusion Module (GCSFM) incorporates Squeeze-and-Excitation mechanisms for channel re-calibration and utilizes boundary-generated gating to dynamically filter semantic features. Experimental results on Rent3D and Raster-to-Vector datasets indicate that the method achieves mean Intersection over Union (mIoU) scores of 78.0% and 81.0%, respectively, outperforming existing state-of-the-art models. The approach significantly improves segmentation accuracy by directly resolving these two critical issues in heterogeneous floor plan data, effectively restoring structural continuity and preventing semantic overflow.