Food crops are very important to people's production and life, and intelligent recognition of food crop disease images is of great significance. However, in the natural environment, the health and disease samples of food crops often have a long-tailed distribution. Long-tailed data is a difficult point for image classification, and insufficient training of tail data often occurs. In this paper, based on transfer learning, the Bilateral-Branch Network is used as the framework to solve the long-tailed recognition problem. Meanwhile, three re-sampling strategies are designed for the Bilateral-Branch Network: balanced re-sampling, smooth reverse re-sampling and reverse re-sampling. After that, we conduct comparative experiments with different model structures and different re-sampling strategies. Finally, we get the conclusion that the Bilateral-Branch Network can effectively solve the long-tailed recognition problem. Among several re-sampling strategies, reverse re-sampling gives the best results. The best result of our method on the long-tailed dataset of food crop disease images is 94.3%, which is 0.7 percentage points higher than that of the traditional single-branch network model under the same conditions.