隨著全球寵物飼養數量快速成長，傳統植入式微晶片成本高昂且具健康風險，而取像端影像品質不佳常導致AI身分識別失敗，加上多數模型部署於後端伺服器，資源消耗大且不便使用。本研究提出「應用於犬隻生物特徵註冊之邊緣AI品質感知 取像框架」，將系統部署於行動裝置前端執行。 框架整合「目標影像空間定位」、「互動式引導」與「模糊檢測」三項前端品質感知機制，以輕量化YOLO26n為核心，針對狗臉、狗鼻紋與狗側身建立專用模型。結合Laplacian Variance模糊檢測與YOLO信心度門檻，品質不足時即時以文字與圖示引導使用者調整拍攝角度與距離，僅將通過品質門檻之高品質影像傳送後端進行身分註冊與比對。 實驗結果顯示，三個專用模型於電腦端經5折交叉驗證後，mAP50 分別達 0.987 ± 0.003（狗臉）、0.897 ± 0.023（狗鼻紋）與 0.952 ± 0.010（狗側身），且在實務部署常用信心度門檻（0.55）下仍維持高 Precision 與 Recall，；部署至行動裝置後，記憶體占用從約1620MB降至341MB，平均信心度維持0.83至0.85。四組消融實驗與五種壓力測試證實，完整前端品質感知機制較Baseline減少重拍次數39.5%、縮短完成時間41.9%（p<0.05）。強光與45°角度偏差情境下成功率達85%，但嚴重幾何遮擋與90°極端畸變時表現仍受限制。 本研究提出並實現以品質感知為核心之邊緣AI取像框架，透過在行動裝置上訓練與部署輕量級YOLO模型，有效降低劣質影像所造成的運算負擔，並顯著提升使用者體驗，可作為寵物管理、走失救援及公共服務之實用解決方案。;With rising pet ownership， traditional microchip identification incurs high costs and health risks， while poor capture-stage image quality often causes AI recognition failure. Most models are also deployed on backend servers， leading to high resource consumption. This study proposes an Edge AI Quality-aware Acquisition Framework for Canine Biometric Registration， deployed on mobile devices. The framework integrates object localization， interactive guidance， and blur detection using lightweight YOLO26n models for dog face， nose print， and body. Real-time guidance is provided when image quality is insufficient， ensuring only qualified images are sent for backend registration. After 5-fold cross-validation， the three models achieved mAP50 of 0.987 ± 0.003 (face)， 0.897 ± 0.023 (nose)， and 0.952 ± 0.010 (body). On Android devices， memory usage dropped from ~1620 MB to 341 MB with average confidence maintained at 0.83–0.85. Ablation studies showed the complete front-end quality-aware mechanism reduced retake counts by 39.5% and completion time by 41.9% compared to baseline (p < 0.05). Success rates reached 85% under strong lighting and 45° angle conditions， though performance remained limited under severe occlusion and 90° extreme distortion. This study proposes and implements an Edge AI quality-aware acquisition framework centered on front-end perception. By training and deploying lightweight YOLO models directly on mobile devices， the framework effectively reduces backend computational load caused by low-quality images and significantly improves user experience， offering a practical solution for pet management， lost animal rescue， and public services. Keywords: Quality-aware Front-end， Canine Identification， YOLO26， Mobile Deployment， Image Quality Assessment， Interactive Guidance
