Diabetic retinopathy (DR) is a common complication of diabetes and one of the major diseases leading to blindness in the world. It is difficult to detect DR in the early clinical stage. A computer aided diagnosis method based on convolution neural network is proposed, which can automatically classify the severity of DR according to the fundus images. Various preprocessing methods are used to improve the quality of the input images and various data enhancement methods are used to improve the balance of datasets. Cost-sensitive regularization is used to expand the standard classification loss function based on EfficientNet network architecture. Depending on the degree of difference between the predicted label and the ground truth label, they are applied different penalties. In addition, pre-training on ImageNet dataset is carried on to introduce transfer learning, and the full connection layer of Softmax activation functions are used to achieve better performance of the model. According to the experimental results of two datasets, compared with the recent research results, the model can achieve an improvement of about 5% in quadratic weighted kappa score and 3% in AUC. The introduction of cost-sensitive regularization into the EfficientNet network model can improve the accuracy of diabetic retinopathy classification task and obtain a good model performance.