Design of a Small-Sample Cow Face Recognition Algorithm for Engineering Applications

Existing cow individual recognition methods typically rely on well-developed cow face databases, with 10 to 50 images per cow, and the training and recognition images are often captured using the same type of sensor. While this approach yields favorable recognition performance, it is impractical in engineering scenarios such as large-scale farms, where obtaining such extensive datasets is challenging or even impossible. Additionally, discrepancies between training dataset images and recognition images are common due to variations in data collection devices. To address these challenges, this paper proposes a practical cow individual recognition algorithm, CFR-Net. The algorithm first detects and extracts cow face images from full-body images captured by cameras or video surveillance. After preprocessing, the extracted faces are matched to corresponding cow IDs for recognition. Experimental results on the same dataset show that CFR-Net outperforms existing methods such as YOLOv8-cls, ResNet50, and InceptionV2 in terms of accuracy. The system, based on CFR-Net, was tested on a dataset comprising 136 cows with 3 to 5 facial images per cow, totaling 532 images, achieving an accuracy of approximately 86%. These results demonstrate the algorithm’s applicability in real-world scenarios with limited datasets.