Abstract: The quality grading of Chinese chestnut had an important influence on the standardization and homogeneity of Chinese chestnut products. Accurate classification was helpful to the standardization of Chinese chestnut products and gave full play to the maximum value of each grade of Chinese chestnut. In view of the low efficiency and low accuracy of Chinese chestnut classification, which mostly depended on machines and manpower, this paper proposed to use the deep learning method to realize the automation and intelligence of Chinese chestnut classification. The classical convolutional neural network LeNet-5 model was improved by adding 5 layers of convolution layer and 2 layers of pooling layer to deepen the network, so as to extract chestnut features more accurately. At the same time, the input layer was modified to the image size of 256×256 color images. The activation function was improved to Leaky ReLu, and the Dropout algorithm was added to alleviate the over fitting phenomenon. Adam was used as the optimizer to optimize the network parameters. Comparing the improved LeNet-5 model with the original LeNet-5 model, AlexNet and VGG16 model, it was found that the improved LeNet-5 model had an average recognition accuracy of 99.68%, an accuracy of 99.34%, and a recall of 99.35% on the test set, which was superior to the other three models. It took only 0.19 seconds to recognize a sample. The improved LeNet-5 model could achieve a good classification of Chinese chestnuts and meet the needs of factories for automatic classification of Chinese chestnuts.

Key words: Chinese chestnut grading, deep learning, convolutional neural network, improved LeNet-5 model