Detection of Cassava Leaf Diseases Using Self-supervised Learning

Abstract

Identifying infected plants such as cassava leaves in advance and removing them early can effectively increase production. The traditional approach is to use many image sets to improve the classification accuracy of neural networks. However, this must be established on the premise that the data sets have enough labels that botanists can classify. We applied self-supervised learning to image classification and trained a reliable cassava disease detection model to reduce the difficulty and cost of collecting these tags. We propose a new hybrid loss that combines the classification and contrastive losses for the whole classification process. Experiments show that our method performs well in the lacking of labeled data. Specifically, the model uses 2/3 of the total data set and reaches an accuracy (90%) close to supervised learning (91.59%) by adding the contrastive term. In addition, we also prove that the change of the percentage of data without labels is linearly independent of the model detection accuracy.