Improving colorization through denoiser with MLP

This paper addresses the problem of generating a plausible colored photograph given a grayscale image. Previous approaches to solve this problem have either relied on significant human input or resulted in desaturated colorizations. Inspired by We explore several architectures/models, approaches and loss functions to understand the best practices to obtain the most aesthetically pleasing colored images. We train a convolutional neural network, and observe that certain loss functions perform better than others due to their inherent properties. In particular, we find that our classificationbased model that uses cross-entropy loss seems to perform the best. We also discuss other state-of-the-art approaches to solve this problem including conditional generative adversarial networks Finally, we provide a qualitative and quantitative analysis of our results, concluding with avenues for future research.