Abstract: At present, the detection of external defects in potatoes mainly depends on manual feature extraction and is consequently inaccurate. A classification method for fast and accurate online grading of potatoes based on the lightweight convolutional neural network is proposed in this paper. First, the Xception network model was trained using ImageNet dataset to establish a pre-trained network model. Then, based on the trained Xception network, this method replaced the Softmax classifier in the original Xception network with category 5 tags, and the potato defect dataset was trained in the Xception framework with transfer learning．Finally, based on the trained external defect recognition model, the classification performance of 5 tags defects was tested. The experimental results showed that when the learning rate was 0.000 01, the overall performance of the network model was optimal, the training accuracy rate was 98.88%, and the loss value was 0.034 9. Compared with nine other neural networks with different depths, the proposed lightweight network model had better recognition performance with average recognition accuracy of 96.04% under the same sample conditions. The model’s processing time was shorter than that of the ResNet152 network, with better recognition effect, and the recognition rate of the network model was 6.4 frames/s. This study could provide theoretical support for online classification of potatoes.

Key words: potato; external defect; transfer learning; Xception network; grading