Evaluation of a Locally Developed Mobile Application in Malawi to Uniquely Identify Cattle Through Facial Recognition

The current study evaluated a locally developed facial recognition mobile application as a tool for cattle identification in the Malawi’s smallholder systems. The study implemented a Deep Learning Convolutional Neural Network (DCNN) model using a Transfer Learning approach to leverage pre-existing models for feature extraction and improve training efficiency to train the application model. The foundational model selected was EfficientNet-B0. The application was tested on 175 cattle, including Malawi Zebu and dairy crosses, across controlled handling races and open-field environments to simulate real-world conditions. Model Performance was evaluated using standard classification metrics including precision, recall, accuracy, and F1-Score. The application demonstrated outstanding predictive and accuracy performance when applied to restrained Holstein-Friesian dairy cows and their crosses in the handling race with a precision of 0.95 and a recall of 0.91. Nevertheless, its best performance occurred in Malawi Zebu cattle registered and retrieved under open field conditions (precision = 1.00 and recall = 0.95). Specificity testing with unregistered cattle showed an 80% success rate. However, the system matched 20% of unknown animals with existing identities indicating need for further refinement to ascertain ownership to curb stock theft. Despite these limitations, the application shows significant potential as a digital livestock identification tool. Its high precision supports traceability and management, though further development is needed to improve robustness in diverse field conditions.