In order to address the issues of the current sheep face pain detection algorithm under complex environments with poor detection accuracy and complex models, this paper proposes a two-stage method of sheep face pain detection based on light yolov5s. First, the weights of the yolov5s model are reduced by combining the Ghostnet structure, feature fusion is performed using the BiFPN structure and the ODConv join to improve detection accuracy. The experimental results show that the number of parameters and complexity of the optimized model are reduced by 38.03% and 50.94%, respectively, compared with the original model, and the accuracy of recognizing sheep faces is improved by 0.2% and the recall rate is increased by 1%. Compared to current mainstream algorithms such as yolov4 tiny and SSD, it not only significantly reduces the number of parameters but also has a better detection performance. Second, pain detection of sheep faces recognised by lightweight yolov5s was carried out by MobileNetV2, and experimental results showed that MobileNetV2 achieved a mean classification accuracy of over 98% for painful sheep, and this illustrates that this two-stage sheep face pain recognition research method has great application value for sheep health breeding.