The difference between sheep is small, the similarity is high, it is difficult to distinguish, and the accuracy of long-distance recognition is not high. To solve that, a sheep face recognition model with efficient channel attention mechanism integrating spatial information was proposed to recognize sheep non-contact. The model was based on MobileFaceNet network. The research generated sheep face detector based on YOLO v4 target detection method was used to construct sheep face recognition database. An efficient channel attention integrating spatial information was introduced into the deep convolution layer and residual layer of MobileFaceNet to increase the extraction range of trunk features and improve the recognition rate. Cosine annealing was used to optimize the dynamic learning rate, and finally ECCSA-MFC model was built to realize sheep individual recognition. The experimental results showed that the accuracy of the sheep face detection model based on YOLO v4 can reach 97.91% and can be used as a face detector. In sheep face recognition, the recognition rate of ECCSA-MFC algorithm can reach 88.06% in open set verification and 96.73% in closedset verification. The proposed ECCSA-MFC model had higher recognition rate and lighter weight. The model size was only 4.8MB, which can provide a solution for intelligent breeding in sheep farm.