Interactive grayscale image colorization with generative adversarial networks

Grayscale image colorization is a classical image editing problem. There are two different methods for colorization. The interaction-based colorization method can generate results based on user interaction. However, this method requires considerable artificial interaction to achieve the desired results. Another method is automatic colorization based on deep learning. However, in this case, the colorization result is unique and cannot be adjusted if the result is incorrect or if the user has additional requirements. In this paper, we combine deep learning with user interaction and propose a grayscale image colorization method based on generative adversarial networks. In this method, a full convolutional neural network is constructed based on the U-net structure as a generator that can process images of any size. The training data is automatically generated by randomly simulating the interactive strokes. The experimental results indicate that this approach can efficiently achieve good colorization results and is capable of generating results based on different user interactions.