Automated Image Colorization using Generative Adversarial Networks

Abstract

Automated colorization of grayscale images is one of the fundamental challenges in the Computer Vision(CV) domain. Self-supervised learning methods are an effective approach to learn general visual features automatically, without having to manually annotate image datasets. Generative Adversarial Networks (GAN) are applied for the colorization method as they are capable of learning visual characteristics from any image without the need for annotated data. They possess a loss function to train the mapping and have ability to learn the mapping from input to output image. The effectiveness of different color spaces (LAB, YUV, and HSI) for image colorization using GAN is investigated. The evaluation of colorization quality is carried out by calculating the pixel accuracy using Peak Signal-to-Noise Ratio (PSNR), and assessing structural integrity using Structural Similarity Indexing Method (SSIM). Using such methods, a comparative approach is demonstrated to study colorization effectiveness of the old black and white images. The SSIM and PSNR values obtained for different color space models are LAB (0.947 and 32.69 dB), YUV (0.842 and 27.65 dB) and HSI (0.851 and 28.72 dB) respectively. Among the three models, LAB color space performed the best with the SSIM and PSNR value of 0.947 and 32.69 dB.