Photograph Enhancement Via Imitation-to-innovation Training Scheme

Abstract

Photographs are acknowledged as a major carrier of visual information, especially in online interactions. The visual quality of photographs significantly influences the efficiency of the daily interaction. Photograph enhancement aims to improve the visual quality of photographs by modifying the pixel values while retaining original semantic information. As photograph retouching software requires operators to take professional training, an automatic photograph-enhancing system can benefit non-expert photographers and save experts from tedious retouching tasks. Modern automatic photograph-enhancing systems utilize convolutional neural networks (CNNs) to approximate the mapping relationship between raw images and manually edited versions.

In this thesis, we present a novel deep learning framework based on an imitation-to-innovation training scheme. Our method integrates a bilateral grid data structure and an adversarial generative network (GAN) to achieve high time-efficiency and appealing retouched output. We also present a bilateral loss function to maintain the piecewise smoothness. Our experimental results demonstrate that our method is capable of recovering vibrant colourization and sharpness from underexposed photographs in microseconds.