Improvement Research of Zero-DCE-Based Algorithm for Low-Light Style Transfer in Generative Cartoon GANs

Among the rapidly developed generative artificial intelligence, CartoonGAN is attracting a lot of attention as it can overcome the limitations of Cartoon production occurring in fields such as Cartoon and animation. CartoonGAN has a problem of not properly preserving the subject shape and edge of a low-light image. To this end, research should be conducted to recover the shape of the subject and improve the structural features of the image through the adjustment of brightness and noise. Therefore, in this study, CartoonGAN, a Cartoon style image generation model using deep learning, proposes a method of applying Zero-DCE to the pre-processing process to overcome the limitations arising from low-light image processing. Through this process, the brightness and color of a low-light image are improved without a reference image. The performance of the proposed algorithm was evaluated through image quality evaluation indicators such as a peak signal-to-noise ratio (PSNR) and a structural similarity index (SSIM). The experimental results showed that the proposed algorithm is effective in improving the performance of CartoonGAN. In addition, it was confirmed that the proposed algorithm effectively improved the brightness of the image through the measurement of the brightness improvement, and a performance evaluation for the recovery of subject information was also performed by introducing a new evaluation method. These results show that CartoonGAN can be improved and effectively applied to low-light image processing. Therefore, this study contributes to the development of image generation technology based on deep learning and presents an important direction to provide high performance in real-world application fields.