牧場乳牛偵測與身分識別;Detection and Identification for Dairy Cows in Pasture

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/84020

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/84020

題名:	牧場乳牛偵測與身分識別;Detection and Identification for Dairy Cows in Pasture
作者:	羅文圻;Lo, Wen-Chi
貢獻者:	資訊工程學系
關鍵詞:	乳牛辨識;乳牛偵測;Cow Identification;Cow Detection
日期:	2020-07-28
上傳時間:	2020-09-02 17:56:20 (UTC+8)
出版者:	國立中央大學
摘要:	乳牛是一種高經濟動物，而其生長發育或發情配種等皆與身分識別有直接關係，因此本研究提出了一牧場乳牛身分識別系統，同時滿足用遠距攝影機拍攝的影像做辨識以及可隨時增刪類別這兩個需求。利用物件偵測找出影像中的乳牛，再用影像對比增強強化特徵並萃取出Auto Encoder 和水平垂直投影三種模態特徵，並為每隻牛分配其專屬的的分類器，新增或減少類別只需增刪分類器即可。每個分類器會分別對三種特徵做量化再由機率神經網路做特徵融合，最後將所有分類器的輸出作整合並推論。實驗結果顯示我們的系統在少量訓練樣本數(10、5)時得到93.5%和88.6%的辨識率，優於ResNet50的86.5%和70.0%，在15 張訓練樣本時則以94.0%略輸給ResNet50 的96.0%。此外在10張訓練樣本時每新增一個類別我們的系統平均降低0.9%的辨識率而ResNet50 則是1.23%。這顯示我們的系統在少量訓練樣本的優勢，以及在新增類別時不僅不須重新訓練，對辨識率的影響也較小，能應對需隨時增刪類別之應用。;Dairy cows are high-economic animals, and their growth and breeding are directly related to identification. Therefore we proposes an identification system for dairy cows in pasture, which can meet the needs of identify by photos taken by remote camera and can add or delete categories easily. Using object detection to find cows in the image, then use image contrast enhancement algorithm to enhance the image and extract the three modal features , including Auto Encoder, horizontal and vertical projection, and assign each cow its own classifier. Adding or reducing categories only needs to add or delete the classifier. Each classifier will extract the three model features separately, and then use the probability neural network to fusion these features, and finally the output of all classifiers are integrate and infer. The experimental results show that our system obtains 93.5% and 88.6% accuracy in fewer training data (10, 5), which is better than ResNet50′s 86.5% and 70.0%. In the case of 15 training data, we got 94% accuracy slightly lost to 96.0% of ResNet50. In addition, adding a new category of 10 training samples, our system reduces accuracy by 0.9% on average and ResNet50 is 1.23%. This shows that our system has the advantage of fewer of training data, and not only doesn’t need to retrain the whole network when adding new categories, but also has less impact on accuracy. Shows that it can cope with applications that need to add or delete categories very often.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	105	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....