Color-UNet++: A resolution for colorization of grayscale images using improved UNet++

Abstract

Colorization is the computer-assisted application of color to a gray scale image, which presents two problems to modern deep learning-based approaches. One is to provide colorization models with both high expressibility and strong learning ability, as current models have difficulty both excelling at coloring and being easy to train. The other is to return a picture without uneven overlap. This paper proposes a deep convolutional network framework called Color-UNet++ for the end-to-end solution of these colorization problems. Color-UNet++ is adjusted to settle gradient dispersion and explosion by capturing more transfer and intermediate results during backpropagation. We adjust the de-convolution structure to solve the problem of uneven overlap. We design the model in YUV instead of RGB color space, with an objective function that is appropriate to the coloring problem and can capture a wide range of colors. A large number of experimental results on LFW and LSUN datasets confirm the method’s superiority.