實驗結果顯示，SSAF 在 M58 工業零件資料集上取得 65.74% 的 Top-1 準確率，相較於現有最佳方法顯著提升了 18.93 個百分點。本研究證明了以語義先驗橋接極端虛實鴻溝的有效性，為隱私受限環境下的工業零件辨識提供了一套穩健的解決方案。;Source-Free Domain Adaptation (SFDA) is essential for privacy-sensitive industrial scenarios. However， in extreme CAD-to-Real settings—where models must transfer knowledge from clean synthetic CAD images to real factory photographs riddled with metallic reflections， oil stains， and cluttered backgrounds—the vast visual discrepancy causes feature space collapse without source data guidance， rendering models unable to discriminate between part categories.
To address this， we propose the Structural Semantic Alignment Framework (SSAF). The key insight is that while visual appearance is inherently unstable across the CAD-to-Real gap， the textual semantics describing a part remain invariant. SSAF leverages Vision-Language Model (VLM) semantics as stable anchors: per-class Semantic Anchors are constructed in the VLM′s semantic space， and a cosine-similarity classifier continuously pulls unstable visual features toward their corresponding semantic centers， replacing the conventional trainable linear head. A Global Centering technique corrects the systematic directional bias in CLIP features caused by training data imbalance， ensuring anchors and target-domain features interact on a consistent geometric basis. Robustness is jointly safeguarded by a Mean Teacher architecture—where a temporally averaged， stable teacher model supervises a more volatile student model—and a calibrated sampling strategy that retains only high-confidence pseudo-labels.
Experiments on the M58 industrial part dataset show that SSAF achieves a Top-1 accuracy of 65.74%， outperforming the current state-of-the-art by 18.93 percentage points. This demonstrates that grounding adaptation in semantic priors effectively bridges the extreme CAD-to-Real gap， providing a robust solution for industrial part recognition under strict data-privacy constraints.
