학습 영상 생성 강화를 이용한 악천후 영상의 시인성 개선

초록

Image-to-image translation is a technique that takes an image as input and transforms it into a new image. Since deep learning-based image translation requires a large amount of training data to prevent overfitting, this study proposes a method to efficiently secure training data by generating and selecting fake water-droplet images using Cycle-Consistent Adversarial Network (CycleGAN) and Convolutional Neural Network (CNN) models for the image enhancement under bad weather conditions. Additionally, we present an approach to improve image visibility using the augmented data. By adding the selected images to the existing dataset and training with the augmented dataset, improved water-droplet removal performance can be observed compared to the original training. This paper proposes an additional training step with tone-mapped target images. This approach not only enhances water-droplet removal but also improves detail preservation and contrast ratio.

키워드