Semantic Segmentation with the Mixup Data Augmentation Method

Abstract

The mixup data augmentation method is a method that creates new images via a linear function from multiple images. In this paper, it is examined whether the mixup data augmentation method improves the U-Net model’s segmentation capability. In this study, artifact segmentation was performed with histopathological images. The dataset used was examined into three different groups: (1) images that are produced through traditional data augmentation methods like flipping and rotation; (2) images that are produced through only the mixup method; and (3) images that are produced through both the traditional and mixup methods. According to the findings, the use of the mixup method in combination with the traditional data augmentation methods improved the model’s average Dice coefficient value for artifact segmentation of histopathological images.