(英) |
In this paper, we propose a new correcting method of chromatic aberration occurring in color images using Deep Learning. In this proposed method, the existing Deep Learning for denoising (well known as DnCNN) is used for chromatic aberration correction purpose. In a lens optical system for imaging, the wavelength of $G$ is designed to be in focus. Therefore, light $R$ with longer wavelength than $G$ and light $B$ with shorter wavelength are out of focus and may cause chromatic aberration. We propose a method to remove the deterioration of chromatic aberration of $R$ channel and $B$ channel by using the $G$ channel without chromatic aberration. In the case of learning DnCNN for restoration of $R$, it is better to perform correction with CNN learned using only learning data of $R$ and $G$. Moreover, for restoration $B$ channel, it is better using $B$ and $G$ data only than using all $RGB$ data. By separating the two DnCNN networks to be trained for the $R$ or $B$ channels, an accuracy and an efficiency of DnCNN can be improved. Further, the training and correcting process by using the $G$ channel with enhanced contrast, make clear the criteria for $R$ and $B$ channel edge correction. Through experimental results, we show the effectiveness of our proposed method |