狗鼻紋路是非接觸式的犬隻身分識別的重要特徵。本論文提出使用深度度量學習（Deep Metric Learning, DML）的自動化狗鼻紋身分識別系統，DML相較典型直接分類方法，其學習的是將輸入映射至具有某種特性的向量空間。透過DML，將灰階狗鼻紋影像輸入至已訓練的DML模型，產生代表該張影像的一維特徵向量，在此向量空間中，同一犬隻不同的鼻紋影像樣本互相群聚，而不同犬隻的鼻紋互相遠離，要衡量兩張鼻紋影像是否屬於同一犬隻衡量此向量空間中代表這兩張樣本的向量的距離即可。深度學習框架成熟，訓練一個神經網路並不難，而DML領域為相對較新的子領域，與典型的深度學習比較上的差異－專為DML設計的損失函數，近年也推陳出新，且深度學習存在許多不同框架，致使了部署上的困難。本論文比較四個DML的損失函數在狗鼻紋上的性能，並將訓練後的神經網路轉至Open Neural Network Exchange（ONNX）的神經網路中介各式，以簡化部署流程。本論文中，輸入影像的狗鼻紋的區域使用語意切割神經網路進行預測，並使用DML搭配KNN的方式對其進行預測。在身分識別模組實驗中，對於訓練時期從未見過的30隻犬隻，每隻4張樣本，分別達到KNN(k=1) 87.3% 與KNN(k=3) 96.7% 的識別率。而完整的識別系統對於訓練時期從未見過的26隻犬隻識別率則為KNN(k=1) 76.9% 與KNN(k=3) 65.4%。;Dog muzzle prints are one of the important feature for non-contact dog identification. This paper proposes an automated dog nasal pattern identification system using Deep Metric Learning (DML), which is a more intuitive classification method than the typical direct classification method. DML learns to map inputs to a vector space with certain characteristics. In this vector space, different muzzle prints of the same dog cluster with each other, while muzzle prints of different dogs are far away from each other. To measure whether two muzzle images belong to the same dog, one calculates the distance between the two vectors representing these two samples. Deep learning frameworks are quite mature and it is not too difficult to train a neural network. However, the DML domain is a relatively new subfield, and its main differences compare to typical deep learning - loss functions designed for DML - have been introduced in recent years, combined with many different frameworks for deep learning, making it difficult to deploy. This paper compares the performance of four DML loss functions on dog muzzle prints images, and transfers the trained neural networks to the Open Neural Network Exchange (ONNX) neural network meddle format, to simplify the deployment process. In this paper, the regions of the dog nasal prints of the input images are predicted using a semantic segmentation neural network, and the prediction is performed using DML with KNN. In the independent identity recognition experiment, the identification rates of 87.3% for KNN(k=1) and 96.7% for KNN(k=3) were achieved for 30 dogs with 4 samples each that had never been seen during the training period. After automating the whole identification process, the identification rate for the 26 dogs never seen during the training period was 76.9% for KNN(k=1) and 65.4% for KNN(k=3).
