GANmera: Reproducing Aesthetically Pleasing Photographs Using Deep Adversarial Networks

摘要

Generative adversarial networks (GANs) have become increasingly popular in recent years owing to its ability to synthesize and transfer. The image enhancement task can also be modeled as an image-to-image translation problem. In this paper, we propose GANmera, a deep adversarial network which is capable of performing aesthetically-driven enhancement of photographs. The network adopts a 2-way GAN architecture and is semi-supervised with aesthetic-based binary labels (good and bad). The network is trained with unpaired image sets, hence eliminating the need for strongly supervised before-after pairs. Using CycleGAN as the base architecture, several fine-grained modifications are made to the loss functions, activation functions and resizing schemes, to achieve improved stability in the generator. Two training strategies are devised to produce results with varying aesthetic output. Quantitative evaluation on the recent benchmark MIT-Adobe-5K dataset demonstrate the capability of our method in achieving state-of-the-art PSNR results. We also show qualitatively that the proposed approach produces aesthetically-pleasing images. This work is a shortlisted submission to the CVPR 2019 NTIRE Image Enhancement Challenge.