Objective:To solve the problem of low accuracy of tobacco grading. Methods:An improved tobacco leaf grading model based on convolutional neural network was proposed. According to the VGG16 network structure, the network model was built in a custom way. The traditional convolution was replaced by the hole convolution, which increased the image receptive field while avoiding. The loss of image features was changed, and the activation function was changed to Leaky_relu. The data distribution was corrected, and the hard saturation problem of the ReLU function was solved. 41 levels of tobacco leaf pictures were used for testing. Results:The grading accuracy rate of the test algorithm was 95.89%, which was 10.46% higher than the traditional SVM algorithm, and 7.87% higher than the classic VGG16 algorithm. The loss rate finally converged to 0.13. Conclusion:Compared with the original model and traditional feature extraction Methods, this algorithm has improved the accuracy of tobacco leaf classification.