Colorization of Grayscale Images Using Generative Adversarial Network

Abstract

The automatic colorization of the grayscale images is an appealing area in the field of image processing. It has a major application in the restoring of aged or degraded images. In this work, we perform the colorization of grayscale images using a Generative Adversarial Network. This method requires less human interference. Especially, a conditional generative network is used to get the desired output. A piece of extra information or condition is fed to both generator and discriminator model. In this thesis, two color spaces have been used: CIE-LAB color space and YCbCr color space. Model is trained over both the color spaces and their results are compared. Experimental results from both the color spaces show that the CIE-LAB color space is more accurate than the YCbCr color space. Images generated by model trained over CIE-LAB color space are sharper and brighter.