MLEU

Abstract

This paper presents the method for tackling the challenge of fully automatically image colorization. We improve U-net by fusion multi-level feature from the pre-trained ImageNet to enhance the model under the small datasets. Furthermore, we reduce the unbalance colors by the enhancement distribution over quantized colors based on the smoothness of the prior distribution. The experiments in the DIV2K dataset show that our results are very encouraging. Our method improves PSNR as well as colorizes the images under complex textures.