Graph-Enhanced Structural Reasoning for Feature-Fusion-Based Character Line-Art Segmentation

Conventional pixel-wise segmentation does not fully exploit long-range relations among semantic components in character line drawings, where sparse textures, broken strokes, and overlapping contours make local evidence unreliable. We propose a graph-enhanced structural reasoning framework built on adaptive multi-level feature fusion. Hierarchical visual features are aligned and spatially reweighted to preserve fine strokes, part-scale structure, and global semantics. Soft component assignments induce a structural graph whose nodes represent semantic prototypes and whose edges combine learned feature affinity, a trainable symmetric gate, and a soft topology prior. Normalized graph propagation transfers evidence across distant but related regions, after which the refined node representations are reprojected to the dense feature map. The objective augments conventional region losses with probabilistic label-transition regularization, prototype supervision, and explicit boundary prediction. We further derive a conditional perturbation bound for graph propagation, prove the non-expansiveness of the assignment-induced dense smoothing operator, and establish a region-coherence bound that characterizes how within-region variation depends on graph connectivity. Equal-budget experiments use five paired seeds on the controlled CLA-Structure benchmark and three paired seeds on 800 public freehand face sketches with native part labels. The full character model achieves 63.03% mean IoU; its paired difference from the boundary-plus-consistency control is −0.08 percentage points (95% interval −0.52 to +0.36). Corrected coherence scoring, topology and global-mixing controls, and additional corruptions expose the limits of graph-specific gains. The evidence supports a reproducible study of semantic context, while providing no basis for universal superiority or a contraction guarantee for the trained network.