| Abstract: | 國內現階段酪農業所面對的最大困境為勞動力不足，首當其衝為牛隻管理上的問題。目前推動之智慧化農場管理旨在以偵測技術進行精準管理，目前的生物識別多以虹膜、口套、耳標等作為辨識目標，惟上述目標皆需極高的影像品質，在實務上難以達到，加上維護成本，故本研究擬以牛隻原有的斑痕與臉部特徵作為辨識目標。
本研究先以25 隻牛做為實驗樣本，共計121 張影像進行訓練與測試。先以YOLOv4
偵測，擷取臉部與斑痕特徵影像並分割，通過Triplet 三元組神經網路取得影像特徵向量，計算樣本間的歐氏距離，進行相似度比對，最後得到牛隻身份識別結果。實驗結果可發現使用單一特徵作為識別條件，牛臉辨識率為92%，斑痕辨識率為88%。最後通過混合神經網路的方式，設計了一個雙模態神經網路，經由PNN 機率神經網路算出牛隻樣本的機率進行身份識別，辨識率可達96%。因此證明在少樣本實驗情境下，透過雙模態神經網路可以有效地提升生物識別效能。;The biggest dilemma faced by dairy farming in Taiwan currently is the lack of labor, especially the problem of cattle management. The current promotion of intelligent farm management aims at accurate management with detection technology. At present, biometric identification mostly uses iris, muzzle, ear tag, etc. as identification targets. However, the above targets all require extremely high image quality, which is difficult in practice. In addition to the maintenance cost, this study intends to use the original scars and facial features of cattle as the identification target.
In this study, 25 cows were used as experimental samples, and a total of 121 images wereused for training and testing. First, YOLOv4 is used to detect, capture and segment feature images of faces and scars, obtain image feature vectors through Triplet network, calculate the Euclidean distance between samples, compare the similarity, and finally get the cattle identification result. The experimental results show that using a single feature as the recognition condition, the recognition rate of cow face is 92%, and the recognition rate of spots is 88%.
Finally, through the method of hybrid neural network, a dual-modal hybrid neural network is designed, and the probability of cattle samples is calculated through the PNN probability neural network for identification, and the recognition rate can reach 96%. Therefore, it is proved that the dual-modal hybrid neural network can effectively improve the biometric identification performance under the few-sample experimental situation. |
