A Fine-Grained Semantic Segmentation Method for the Regenerated Built Environment at the Tianjin First Machine Tool Factory: Local–Global Feature Enhancement and Class-Discriminative Constraints

Industrial heritage conservation has shifted toward adaptive reuse, making accurate differentiation of historical components, later interventions, and other architectural elements essential for quantitative spatial analysis and renovation assessment. However, visually similar semantic categories pose challenges in local detail discrimination, local–global contextual integration, and class imbalance. Using the Tianjin First Machine Tool Factory as a case study, we constructed a dataset of 320 directional images with 14 foreground classes from 80 on-site panoramas through four-directional projection and pixel-level annotation. We propose an enhanced Mask2Former for the site’s regenerated built environment (Hg-Mask2Former++), integrating multi-scale local–global feature enhancement (LGFE) and class-discriminative contrastive constraints (CDCCs) through class-balanced sampling and supervised contrastive learning. With parameters increasing from 106.887 M to 108.483 M (+1.596 M; approximately 1.49%), the model achieved 73.92% mIoU and 82.67% mDice in a single ablation run, exceeding the baseline by 2.12 and 1.86 percentage points, respectively. However, the 95% confidence interval for test-set mIoU differences across five paired random-seed experiments was [−3.43, +6.51] percentage points, including zero; a systematic advantage across seeds was not demonstrated. Under the same data partition, training budget, and validation-based checkpoint selection, its mIoU exceeded eight comparison models, including FastVim, TPSegformer, and ViT-P-based models proposed in 2026, with architecture-specific optimization settings. In the single ablation run, Industrial Building and Modern Building IoUs improved by approximately 20 percentage points, but class-level gains varied across seeds. The results reveal class-selective and condition-dependent effects, supporting targeted optimization rather than simply increasing network complexity. Findings are limited to the sampled site area and established data partition and do not demonstrate independent transfer to other industrial heritage sites.