Augmentation of Imbalanced Training Data for Image Classifier Learning by Image-to-Image Transformation from CG

Abstract

Image classification techniques using machine learning have been applied to various applications, however, a large number of images are generally required for training. For example, in the specific fields such as medical, inspection and agricultural tasks, it is sometimes difficult to obtain enough images, and some classes are imbalanced in training dataset. Data augmentation is a solution to increase the amount of data in such classes. In general, data augmentation methods using computer graphics (CG) does not perform well due to their expressions are not enough to recognize objects in real world. In this paper, we propose a data augmentation method by a data transformation using CG. The method transforms CG images modeled as simplified object into real-like images. Furthermore, a classifier is trained with the dataset balanced by the augmented data. In experiments, the performance of the proposed method was verified in classification of real worlds objects.