Click on “Download PDF” for the PDF version or on the title for the HTML version.


If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Lightweight Cow Face Recognition Algorithm based on Few-Shot Learning for Edge Computing Application

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2021 ASABE Annual International Virtual Meeting  2100557.(doi:10.13031/aim.202100557)
Authors:   Yue-Shao Chen, Cheng-Yu Kuan, Jih-Tay Hsu, Ta-Te Lin
Keywords:   livestock monitoring, face recognition, embedded system, edge computing, few-shot learning, embedding learning

Abstract. Dairy cows often suffer from heat stress problems due to extreme levels of humidity and temperature. Heat stress contributes to the decrease in the feed and drink intake, fertility, breathing rate, and milk production of dairy cows. With the use of modern technology, these behaviors can be automatically monitored by farm owners, thereby ensuring the health of the dairy cows. This research presents a lightweight algorithm for cow face recognition tailored for edge computing application. The proposed algorithm was implemented in an automated dairy cow feeding behavior monitoring system made up of embedded imaging devices. By edge computing, the system can be installed in a dairy farm with improved scalability, efficiency, and data security. A lightweight cow face image recognition convolutional neural network (CNN) model was optimized and trained using few-shot learning (FSL) with a testing accuracy of 0.90. A method in FSL, called embedding learning, was used to enable the cow face recognition model to adapt based on newly acquired training samples. Embeddings were generated that represent lower dimension vectors extracted by the model from the cow face images. In the reduced dimension, the L2 distance of each embedding represented the similarity of each image sample with the support of triplet loss. This research overcomes the non-convergence issue in model training through adaptive training methods that can create a similarity space between embeddings. The techniques in this research may also be applied in other fields that require adaptive face recognition methods.

(Download PDF)    (Export to EndNotes)