In order to solve the problems with manual assessment of individual sheep’s pain, which includes the requirement for a high level of human experience on the subject matter, a lack of pain recognition accuracy, and extended delay for the treatment for sheep, spatial transformer visual geometry group network (STVGGNet) was proposed as an improved model to the current mainstream deep learning model visual geometry group network (VGGNet). The STVGGNet algorithm introduced the spatial transformer networks and increased the area of analysis and in return improved the level of recognition of a sheep’s facial expression with regards of pain. Additional 887 images were added to the pre-existing dataset of sheep’s facial expression images. However, because the new image dataset remained low in quantity, the model also utilized ImageNet for transfer learning and fine-tuning classification between painful and non-painful sheep’s facial expressions. The experimental results showed that the best performance accuracy of STVGGNet in training stood at 99.95% with the best validation results upwards of 99.06% vs the VGGNet model which yielded 99.80% and 95.07% respectively. Therefore, with STVGGNet’s improved accuracy and strong robustness to classify pain within a sheep’s facial expression, it provided technical support for the intelligent development of sheep disease detection in animal husbandry.