In smart livestock, precision livestock systems require efficient and safe non-contact cattle identification methods in daily operation and management. In this paper, we focus on lightweight Convolutional Neural Network (CNN) based cattle face identification in natural background. Particularly, we first construct a fine-grained cattle recognition dataset with natural background. Then, we propose a lightweight CNN model MobiCFNet, containing a two-stage method that can realize one-shot cattle recognition. Finally, a series of experiments are conducted to validate the effectiveness of our proposed network .