Improving Robustness Using MixUp and CutMix Augmentation for Corn Leaf Diseases Classification based on ConvMixer Architecture

Abstract

Corn leaf diseases such as blight spot, gray leaf spot, and common rust still lurk in corn fields. This problem must be solved to help corn farmers. The ConvMixer model, consisting of a patch embedding layer, is a new model with a simple structure. When training a model with ConvMixer, improvisation is an important part that needs to be further explored to achieve better accuracy. By using advanced data augmentation techniques such as MixUp and CutMix, the robustness of ConvMixer model can be well achieved for corn leaf diseases classification. We describe experimental evidence in this article using precision, recall, accuracy score, and F1 score as performance metrics. As a result, it turned out that the training model with the data set without extension on the ConvMixer model achieved an accuracy of 0.9812, but this could still be improved. In fact, when we used the MixUp and CutMix augmentation, the training model results increased significantly to 0.9925 and 0.9932, respectively.

Keywords

• ConvMixer
• corn leaf diseases
• CutMix
• data augmentation
• MixUp
• robustness