Abstract: Aiming at the problem of low accuracy of classifying chestnuts by artificial or mechanical vibrating screen, a chestnut grading method based on shallow convolutional neural network was proposed. The 5 481 images of five levels of chestnut were captured by millet mobile phone and applied to the training, verification and testing of the convolutional network model.By learning the network structure of EfficientNet, the designed shallow convolutional neural network（Efnet-1） was composed of 1 ordinary convolutional module and 3 MB convolution modules to form a chestnut image feature extractor. The feature extractor connected a classifier composed of the global average pooling layer, hidden layer and output layer. The relevant hyperparameters were optimized during the training of the Efnet-1 model. The chestnut grading performance of Efnet-1 and deep learning model AlexNet was analyzed. The grading accuracy of Efnet-1 was 98.68%, and the proportion of bad chestnuts being classified into good chestnuts was not more than 0.9%. The chestnut image classification time of Efnet-1 was 62 ms. The improved convolutional neural network model Efnet-1 grades chestnuts quickly and accurately, which provides a technical basis for the automatic classification of chestnuts.