摘要 針對犬隻識別於非受控環境下因影像嚴重退化導致系統失效之挑戰，本研究提出一套基於視覺語言模型(VLM，使用模型為CLIP-ViT-B/16)特徵原子化與品質感知動態融合之識別架構。透過鼻紋、臉部及身體三模態解耦設計，結合品質評估模型與Listwise-Transformer，解決單一模態在極端模糊情境下因特徵偏移導致辨識率下滑的問題。 本架構於VLM擷取階段引入特徵攔截策略，刻意繞過模型末端之降維投影層，直接提取未經跨模態壓縮的768維純視覺特徵，此特徵原子化機制實現特徵一次擷取、多處特徵對齊的離線存取機制，確保身份辨識特徵與品質感測立基於相同潛在語義空間，大幅提升比對可靠度。此外，於預處理階段引入平均距離法取代傳統最近鄰法，有效拯救影像退化造成特徵飄移導致誤判之現象，能夠顯著提升單一模態辨識率。 在訓練策略方面，實驗結果顯示雖然極端困難負樣本能逼出線性模型(如羅吉斯迴歸)的效能上限，但卻容易導致非線性模型(如XGBoost)之決策邊界產生碎裂現象，採用隨機負樣本探勘能提供更具普適性的特徵分佈，確保跨演算法架構在面對退化影像時仍能維持優異的決策穩定度。 實驗結果顯示，當多模態對稱退化強度達σ=3.0時，單一模態辨識率跌破10%之際，本系統仍能透過動態權重轉移將整體辨識率穩固於42%。消融實驗與熱力圖視覺化分析證實了隱性品質特徵學習現象，模型不僅能在非對稱情境中精準將權重轉移至清晰模態，亦能在極端對稱退化下憑藉感知全域距離膨脹克服權重均分之僵局；然而純資料驅動在面對過度密集的退化樣本時易引發注意力分散，此時引入顯性品質分數能發揮關鍵的語義定錨與微觀校準作用，有效改善極端雜訊下的決策困境，使系統能在降低特徵工程依賴(縮短訓練時間)與維持防禦強健性之間取得最佳化平衡。 在工程應用部署層面，本研究採特徵萃取與融合決策完全解耦之策略，使系統在不變動底層網路下即可強化防禦。最後，針對物聯網佈署提出雙軌漸進式策略，以低延遲之極端梯度提升(XGBoost)模型搭配非同步深度學習訓練，解決重訓練期間之系統空窗期困境，為強健型生物識別之應用提供具高度實務價值之解決方案。;Abstract To address dog identification failures under severe image degradation， this study proposes a robust architecture using Vision-Language Model (VLM，CLIP-ViT-B/16) feature atomization and quality-aware dynamic fusion. By decoupling three modalities (nose， face， body) and integrating a quality assessment model with a Listwise-Transformer， it mitigates recognition decline caused by extreme blur. During the VLM extraction phase， this architecture introduces a feature interception strategy， deliberately bypassing the model′s terminal projection layer to directly extract the uncompressed 768-dimensional pure visual features. This feature atomization mechanism establishes an offline access framework that achieves single-extraction and multi-alignment. It ensures that identity and quality features share the same latent semantic space， thereby substantially enhancing matching reliability. Additionally， replacing traditional nearest neighbor matching with a mean distance method during preprocessing effectively mitigates misjudgments caused by image degradation-induced feature drift， significantly improving single-modality recognition rates. Regarding training strategies， unlike extreme hard negative sampling which fragments non-linear decision boundaries， random negative sample mining provides generalized feature distributions， ensuring cross-architecture stability.. Experimentally， when multi-modal symmetric degradation reaches σ=3.0 and single-modality accuracy drops below 10%， the proposed system can still stabilize the overall recognition rate at approximately 42%. Ablation studies and attention heatmap visual analysis confirm the phenomenon of "implicit quality feature learning": the model not only accurately transfers weights to clear modalities in asymmetric scenarios， but also overcomes the deadlock of uniform weight distribution under extreme symmetric degradation by perceiving "global distance expansion." However， purely data-driven approaches are prone to causing attention dilution when encountering overly dense degraded samples. Introducing explicit quality scores provides crucial semantic anchoring and micro-calibration， thoroughly breaking through the decision bottleneck under extreme noise. This enables the system to achieve an optimal balance between reducing reliance on explicit feature engineering (shortening training time) and maintaining defense robustness. At the engineering deployment level， this study adopts a completely decoupled strategy between feature extraction and fusion decision-making， enabling the system to enhance its robustness without modifying the underlying network. Finally， for Internet of Things (IoT) deployment， a dual-track progressive strategy is proposed. By pairing a low-latency Extreme Gradient Boosting (XGBoost) model with asynchronous deep learning training， this strategy resolves the dilemma of system downtime during retraining， providing a highly practical solution for robust biometric identification applications.
