Automatic Image Colorization based on Multi-Discriminators Generative Adversarial Networks

This paper presents a deep automatic colorization approach which avoids any manual intervention. Recently Generative Adversarial Network (GANs) approaches have proven their effectiveness for image colorization tasks. Inspired by GANs methods, we propose a novel colorization model that produces more realistic quality results. The model employs an additional discriminator which works in the feature domain. Using a feature discriminator, our generator produces structural high-frequency features instead of noisy artifacts. To achieve the required level of details in the colorization process, we incorporate non-adversarial losses from recent image style transfer techniques. Besides, the generator architecture follows the general shape of U-Net, to transfer information more effectively between distant layers. The performance of the proposed model was evaluated quantitatively as well as qualitatively with places365 dataset. Results show that the proposed model achieves more realistic colors with less artifacts compared to the state-of-the-art approaches.