Reliable Recognition and Hard-Boundary Evaluation for Archiving Images of Huizhou Architectural Heritage

Huizhou architectural heritage images are important visual materials for regional heritage documentation, digital archiving, and conservation support. However, Huizhou architecture shares many visual cues with Jiangnan white-walled and grey-tiled buildings and other visually similar Classical Chinese architectural forms. In practical image archiving, the critical risk is therefore not only whether a model can achieve high four-class accuracy, but also whether visually similar Classical Chinese buildings are incorrectly archived as Huizhou architecture. To address this problem, this study constructs the HuizhouStyle-4C image recognition task and proposes a reliability-oriented evaluation workflow for Huizhou architectural heritage image archiving. The workflow combines group-aware leakage control, hard-boundary diagnosis, and an out-of-fold (OOF)-safe fixed decision rule. A structure-aware baseline classifier, SAB-ResNet, is further used to incorporate architectural boundary and facade-structure cues into the recognition process. The evaluation includes group-aware five-fold cross-validation, an independent holdout subset from the curated dataset, and a hard-boundary test focused on the distinction between Huizhou architecture and visually similar Classical Chinese samples. Under this within-dataset held-out evaluation setting, the OOF-safe balanced point achieved an Independent Holdout Macro-F1 of 0.9501 and a Hard Binary F1 of 0.8182. The results indicate that Huizhou architectural heritage image archiving benefits from not only conventional four-class recognition, but also explicit diagnosis of near-neighbour misclassification risks and fixed decision rules selected without test-set labels. Additional stronger backbone comparisons further show that competitive held-out four-class performance does not eliminate the need for hard-boundary diagnosis.