犬隻鼻紋具有非侵入、個體獨特且終身穩定等特性，具備應用於寵物管理、走失協尋與保險理賠之潛力。然而實務影像常受模糊、低光照與低解析度等劣化影響，使深度度量學習模型於自動化辨識流程下之效能明顯下降；固定式前處理亦難以同時兼顧清晰影像與重度劣化影像之需求。 為解決上述問題，本研究提出一個基於流形約束特徵處理控制之品質感知自適應生物特徵識別架構。方法上，以DA-CLIP 建構品質語義流形，將輸入影像映射為劣化語義嵌入，並由流形約束MLP 控制器輸出前處理控制參數；其參數矩陣再透過Sinkhorn–Knopp 疊代投影至Birkhoff多胞形，以提升控制決策之穩定性與可解釋性。系統整合ROI 擷取、自適應前處理與Swin-Tiny 搭配AdaFace 之辨識模組，形成可隨影像品質動態調整之推論流程。 本研究使用MIAT 實驗室自建犬隻鼻紋資料集，共含130 個身分、1，078 張影像；模型訓練採identity-disjoint 切分，最終於220 張獨立測試影像上以leave-one-out 協定評估。實驗結果顯示，本方法於整體測試集之Rank-1 與Rank-5 分別達92.27% 與96.36%，相較Swin-T+AdaFace+前處理增強（S1）基線之91.36% 與95.91% 進一步提升；於訓練階段未出現之MIAT-3 子集上，Rank-1 由81.94% 提升至84.72%。需注意的是，mAP 由S1 的88.70% 微幅下降至88.62%。因此，本研究不將整體Rank-1 的增益視為唯一貢獻，而是將方法價值定位於訓練階段未出現子集與重度劣化情境下之穩健性提升，以及流形約束對控制器決策穩定性的改善。;Dog noseprint recognition is a non-invasive biometric modality with persistent individuality over an animal’s lifetime. However， real-world captures often suffer from blur， low-light conditions， and low resolution， while fixed preprocessing cannot adapt to varying image quality. This thesis proposes a quality-aware adaptive biometric recognition framework based on manifold-constrained control of feature processing. A DA-CLIP encoder maps each input into a quality semantic manifold， and a manifold-constrained MLP controller predicts preprocessing parameters from the degradation embedding. The parameter matrix is projected onto the Birkhoff polytope through Sinkhorn–Knopp iterations to improve the stability and interpretability of adaptive decisions.The framework integrates ROI extraction， adaptive preprocessing， and a Swin-Tiny + AdaFace recognition module. This study uses the MIAT laboratory self-built dog noseprint dataset， which contains 130 identities and 1，078 images; model training follows an identitydisjoint split， and final evaluation is conducted on 220 independent test images under a leave-one-out protocol. The proposed method achieves Rank-1 = 92.27% and Rank-5 = 96.36%， improving over the Swin-T + AdaFace + preprocessing augmentation (S1) baseline of 91.36% and 95.91%， respectively， and improves Rank-1 on the MIAT-3 subset that is absent from training from 81.94% to 84.72%.It should be noted that mAP slightly decreases from 88.70% for S1 to 88.62%.Therefore， the contribution of this study is not positioned solely as an overall Rank-1 gain， but rather as improved robustness under training-subset shift and severe-degradation conditions， together with more stable controller decisions enabled by the manifold constraint.
