Wu, Lan, Sun, Cheng, Shi, Chen and Zen – AN IMPROVED A.N.N.-BASED LABEL PLACEMENT METHOD CONSIDERING SURROUNDING FEATURES FOR SCHEMATIC METRO MAPS 2024

Description

The paper introduces an improved ANN-based method for automated station label placement on schematic metro maps that explicitly models surrounding (unconnected) features in addition to station points and connected edges. Label–point, label–edge, and label–label overlaps, label ambiguity (distance to other points), passing-line direction, and 16 candidate positions per station are encoded in an N × 69 feature matrix fed to a multilayer feedforward ANN (PyTorch, ReLU). Training used 2,635 stations and testing 413 from 15 cities; maps were vectorized, standardized by character MBR, and candidate labels generated in ArcGIS. Quantitative evaluation shows large improvements over a benchmark: label–point overlap 0.00% vs 4.17%, label–edge 4.12% vs 14.29%, and label–label 20.58% vs 35.11%. The method reduces ambiguous labels (higher Disminimum), keeps labels on the correct passing-line side, and was preferred by ~72–83% of 42 survey participants, though some favored positions were sacrificed. Ablation reveals label–point and label–edge features are critical; remaining adjacent-station overlaps can be further optimized (e.g., genetic algorithms). The approach is effective, extensible, and suggests directions for refinement.