Image-Enhancement-Based Data Augmentation for Improving Deep Learning in Image Classification Problem
Weihang Zhang, Yuma Kinoshita, Hitoshi Kiya
Abstract
In this paper, we propose a novel data augmentation method based on image enhancement. When training CNN models for image classification, it is required to prepare sufficient training data taken in various conditions. However, traditional data augmentation methods are limited to physical transformation, and shooting conditions of images, so exposure conditions lack for data augmentation. Therefore, we utilize an image enhancement method to generate images with different exposures, although enhanced methods are generally used for generating high quality images. Experimental results show that the proposed method improves the classification accuracy of a CNN model. The results also demonstrate that combining the proposed method with other existing data augmentation methods provides further improvement of the classification accuracy.