Key points are not available for this paper at this time.
This research is the first to adopt a sophisticated automated visual colorization paradigm based on cutting-edge algorithms of conditional generative adversarial networks. Strategically, the world is moving towards the color space of L*a*b and implementing a refined, streamlined methodology to address problems with manual intervention. The proposed model isa combination of UNET and ResNet18 architectures, which undergoes a meticulously coordinated dual-stage training process. The model's ability to produce subtle and generalized colorization is further increased by using SPair 71K in the pretraining phase. The suggested model is compared in this study to the well-established "Palette" baseline, which is renowned for its effectiveness in colorization and image-to-image translation tasks. The model performs exceptionally well in managing color variability, taking care of special components, and reducing color spillovers. With the lower Fréchet Inception Distance (FID) and higher Inception Score (IS) and Colorization Accuracy (CA) than Palette, the results are as follows: FID 3.2, IS 215, and CA 73.0% compared to FID 3.4, IS 212.9, and CA 72.0% for Palette. This shows the efficiency of the proposed model and marks an important development in automated visual colorization approaches.
Kandari et al. (Wed,) studied this question.