This paper proposes a novel data augmentation method for enhancing the performance of deep-learning-based object detection in underwater environments. With deep-learning-based methods, system performance is highly dependent on the learning dataset. In extreme conditions such as underwater environments, however, it is difficult to obtain sufficient image data. To make this easier, we propose an image generation method that augments learning data by applying the special conditions of underwater environments. The proposed method generates virtual underwater images by using the optical properties of water from images taken above ground. The generated data can be used in place of underwater data, thus reducing the training effort. Experimental results show that the proposed method is effective.
Paper
Full text
Data Augmentation Method for Object Detection in Underwater Environments
Semantic Scholar · Environmental Science · 2019
Abstract
This paper proposes a novel data augmentation method for enhancing the performance of deep-learning-based object detection in underwater environments. With deep-learning-based methods, system performance is highly dependent on the learning dataset. In extreme conditions such as underwater environments, however, it is difficult to obtain sufficient image data. To make this easier, we propose an image generation method that augments learning data by applying the special conditions of underwater environments. The proposed method generates virtual underwater images by using the optical properties of water from images taken above ground. The generated data can be used in place of underwater data, thus reducing the training effort. Experimental results show that the proposed method is effective.