The implementation of neighbor embedding on grayscale image colorization

Abstract

Image colorization is something fascinating in Image Processing. The issue raised in this research is Grayscale Image Colorization. Grayscale images that only show black and white are reproduced to be color images. The employed methodology was Neighbor Embedding Method, which has been applied before to manage Super-Resolution cases. This method is also implemented by Jun Li on image colorization cases. The method utilized in this grayscale image colorization was an automatic image colorization, requiring two types of images, i.e. grayscale image as the target image and color image as the training or reference image. After undergoing histogram analysis, the colors of the training image were successfully transferred to the target image. The closer the luminance of both images is, the easier the transfer process becomes. Furthermore, the parameters affecting this colorization are patch, K value, and the color range of the target image. In addition to histogram analysis, a research toward Mean Absolute Error (MAE) was also performed for YIQ color model. The results of MAE obtained from the experiment were 0.028311104 for MAE I, which was the smallest value, and 0.025990016 for MAE Q. These results confirm that YIQ color model is relatively applicable in image colorization using this method. Mean Opinion Score (MOS)-based analysis also indicates that colorization results using Neighbor Embedding are relatively in accordance with the image colors in general.