A Study for Semi-supervised Learning with Random Erasing

Due to the recent popularization of various data classified by computer, machine learning is attracting great attention. A common method of machine learning is supervised learning, which classifies data using a large number of class labeled training data called labeled data. To improve the processing performance of supervised learning, it is effective to use Random Erasing in data augmentation. However, since supervised learning requires much labeled data, the cost of manually adding label information to an unclassified training case (unlabeled data) is very high. In this paper, we propose a method for achieving high classification accuracy using Random Erasing for semi-supervised learning using few labeled data and unlabeled data. In our evaluation, we confirm the availability of the proposed method compared with conventional methods.