Human Perception-based Image Enhancement Using a Deep Generative Model

Amir Nazemi, Shima Kamyab, Zohreh Azimifar, Paul W. Fieguth

Abstract: In this paper we propose a deep model for perceptual image enhancement based on generative modeling. The proposed framework is inspired by the Conditional Variational AutoEncoder (CVAE)which is a well-known deep generative structure. In generative models, there are efficient regularizers for controlling the output distributions using information from input data which lead to accurate and visually plausible results with few parameters. Additionally,we propose to use an image quality assessment network to deter-mine the best result among those obtained by the implemented CVAEs. The proposed CVAE structure models the histogram vectors of different color channels and parameters of image data (i.e.,the networks do not work directly on pixel values). This configuration makes the proposed framework capable of using images of different sizes. Qualitative and numerical evaluations on a related dataset compared to state-of-the-art indicate superiority of the pro-posed framework in improving image quality and content.