小樣本學習臉部辨識演算法應用於乳牛採食與溫度監測

Dairy Cow Face Recognition Based on Few-Shot Learning for Feeding Behavior and Eye Temperature Monitoring

陳約劭(Yue-Shao Chen)

指導教授 ： 林達德

國立臺灣大學/生物資源暨農學院/生物機電工程學系/碩士(2021年)

https://doi.org/10.6342/NTU202101767

摘要

在亞熱帶乳牛產業中，熱緊迫為乳牛飼養與生產管理的重要問題。熱緊迫會影響乳牛的採食量、飲水量、生育能力、呼吸速率和產乳量。在這些行為中，乳牛採食量的變化為乳牛熱緊迫的重要指標，亦直接影響乳牛之產乳量。本研究利用小樣本學習之臉部辨識演算法監測乳牛採食行為，以解決傳統模型遇到新類別就需花費大量時間重新訓練的問題，同時提升辨識效能。影像系統使用Raspberry Pi 3B+作為邊緣運算系統，搭配ArduCam攝影鏡頭來擷取乳牛採食時之臉部影像進行臉部辨識。乳牛採食行為辨識有兩步驟：偵測及辨識。YOLOv4-Tiny用於偵測乳牛臉部位置，模型F1-score為0.98。臉部辨識模型以小樣本學習演算法訓練MobileNetV2，利用online triplet loss損失函數來實現。模型輸出特徵向量之L2距離即為相似度，且可用距離閾值判斷是否為新目標牛隻，其F1-score為0.91。本研究驗證實驗在辨識基準模型訓練19類新增5類的情況下，各類100張訓練影像即可達到平均準確率0.97。最終驗證系統預測個別牛採食時間與人工計算其R^2=0.98。本研究亦利用所開發之牛臉辨識模型加入牛眼偵測模型，同樣以YOLOv4-Tiny為模型架構，其表現F1-score為0.92，再利用熱影像儀擷取個別牛眼溫度資訊，其溫度範圍落在±0.3°C內。本研究進一步將自動辨識系統所得個別泌乳牛採食時間監測資料，以分娩後天數區分為三個類別分析個別乳牛採食時間與溫濕度指標(Temperature and humidity index, THI)關聯性。分析結果顯示離分娩日期越短的牛隻其所受THI影響較小，而泌乳中至後期之牛隻，2日前平均THI對於採食時間有負相關性。牛眼溫度監測中得到所收溫度資訊受環境溫度影響，其R^2=0.89，而在相同環境溫度時偵測牛眼溫度範圍穩定，未來可應用於乳牛異常溫度之偵測。

關鍵字

熱緊迫 ； 嵌入式系統 ； 小樣本學習 ； 臉部辨識 ； 牛眼溫度

並列摘要

In sub-tropical dairy farms, heat stress is an important issue in dairy cattle feeding and production management. Heat stress affects feeding intake, water intake, fertility, respiration rate, and milk production of dairy cows. Of these behaviors, fluctuating feeding intake can be used as an indicator to alert farmers that dairy cows are suffering heat stress. This paper proposes an adaptive face recognition algorithm for monitoring the feeding behavior of dairy cattle using an embedded imaging system. The imaging system uses a Raspberry Pi 3B+ as an edge computing platform, and an ArduCam camera module for capturing images of cows’ faces during feeding. There are two steps in feeding behavior recognition in the proposed model: face detection and face recognition. YOLOv4-Tiny was used to detect and crop cows’ faces from the acquired images, and the final F1-score is 0.98. The few-shot learning algorithm was applied to train MobileNetV2 for dairy cow face recognition, and online triplet loss was used in the training process. The L2 distance between the model output features represent the similarity. The analysis revealed that an average accuracy of 97% could be achieved for 5 newly added categories to the based model of 19 categories after retraining of the adaptive model. This research verifies that the image system estimated the feeding time of individual cow with an R^2=0.98 when compared with manual observation. This study also integrated the cow eye detection model with the developed cow face recognition model for automated eye temperature measurement. The F1-score of the cow eye detection model is 0.92. The cow eye detection results were then mapped to the thermal imaging camera to collect individual cow eye temperature information, and the eye temperature values range are within ±0.3°C. The feeding time of individual dairy cows acquired from the automated monitoring system was further analyzed to investigate the effect of temperature and humidity index (THI) on feeding behavior. The results showed that dairy cows group in early lactation were less affected by THI. While for cows in mid- to late lactation, the average THI before 2 days was negatively correlated with feeding time. The temperature information received from the cow eye temperature monitoring was influenced by the environment temperature with R^2=0.89. While the detected cow eye temperature value is stable at the same environment temperature, and this yields the potential of detecting abnormal cow temperature in the future.

並列關鍵字

heat stress ； embedded system ； few-shot learning ； face recognition ； cow eye temperature

參考文獻

行政院農業委員會。2019。農業統計年報。台北：行政院農委會。網址：https://agrstat.coa.gov.tw/sdweb/public/book/Book.aspx。出版日期：2019-07-10。

蔡雨錡。2018。乳牛熱緊迫影像監控系統之建置與資料分析。碩士論文。台北: 臺灣大學生物機電工程學研究所。

官承譽。2019。基於深度學習之泌乳牛採食行為影像監測系統。碩士論文。台北: 臺灣大學生物機電工程學研究所。

Achour, B., Belkadi, M., Filali, I., Laghrouche, M., and Lahdir, M. 2020. Image analysis for individual identification and feeding behaviour monitoring of dairy cows based on Convolutional Neural Networks (CNN). Biosystems Engineering, 198, 31-49.

Argüeso, D., Picon, A., Irusta, U., Medela, A., San-Emeterio, M. G., Bereciartua, A., and Alvarez-Gila, A. 2020. Few-Shot Learning approach for plant disease classification using images taken in the field. Computers and Electronics in Agriculture, 175, 105542.