어두운 환경 감시를 위한 GAN 기반 영상 변환 기법

다국어 초록 (Multilingual Abstract)

The GAN(generative adversarial network) learning-based image transformation can be used for conversion from day images to night images. However, in the case of converting from night images to day images, because the lost information in the night image is predicted and recreated, it exposes more problems than when converting day images into night images. In this paper, we proposed a method of converting day images into night images based on pix2pix. The proposed method reduces the number of hidden units in the pix2pix to solve the problem of gradient vanishing and adjusted the exposure value of the captured day image to create training data for the virtual night image. The experimental results show that the proposed method is superior to pix2pix and Cycle GAN in MOS(mean opinion score) and PSNR(Peak Signal-to-Noise Ratio).

참고문헌 (Reference)

1 이성학, "시각의 공간 주파수 응답을 고려한 고속 바이래터럴필터 분할 영역에서의 HDR 베이스 및 세부 계층 분리" 한국정보기술학회 16 (16): 57-65, 2018

2 Y. Yu, "Unsupervised Representation Learning with Deep Convolutional Neural Network for Remote Sensing Images, Vol. 10667" Springer International Publishing 97-108, 2017

3 S. W. Jung, "Regional Weighted Generative Adversarial Network for LDR to HDR Image Conversion" 697-700, 2020

4 C. Ledig, "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network" 105-114, 2017

5 R. C. Streijl, "Mean opinion score (MOS)revisited: methods and applications, limitations and alternatives" 22 : 213-227, 2016

6 P. Isola, "Image-to-image translation with Conditional Adversarial Networks" 5967-5976, 2017

7 G. Eilertsen, "HDR image reconstruction from a single exposure using deep CNNs" 36 (36): 178-, 2017

8 I. Goodfellow, "Generative adversarial nets" 2672-2680, 2014

9 F. Durand, "Fast bilateral filtering for the display of high-dynamic-range images" 257-266, 2002

10 Y. LeCun, "Backpropagation Applied to Handwritten Zip Code Recognition" 1 (1): 541-551, 1989

11 P. Baldi, "Autoencoders, unsupervised learning, and deep architectures" 27 : 37-50, 2011