Fine-grained Potato Disease Identification Based on Contrastive Convolutional Neural Networks

Early identification of potato diseases is of great significance for reducing yield losses. The identification of different types of diseases has achieved great success. However, for different periods of different disease, it is difficult to distinguish due to similar symptoms and fine-grained, so there are few related studies. In this study, we proposed a convolutional neural network based on contrastive learning to identify fine-grained potato diseases. Different from the previous unsupervised contrastive learning used in pre-training, the proposed model adds a projection head to the backbone network of Vgg16 to extract the contrastive representation features, and then integrates the contrastive loss with the classification loss to form a joint loss. Finally, an end-to-end supervised contrastive convolutional neural network is constructed, which is easier to train while reducing the transmission error. Experimental results show that the proposed model achieves an average recognition accuracy of 97.24%, which is higher than 90.28% of Resnet50, 90.62% of Resnet101, 93.06% of AlexNet, 94.44% of Inception V3, and 94.79% of Vgg16. It shows that the model has an obvious effect on classification task with similar features, and has practical significance for fine-grained potato disease identification.