Aesthetics-Driven Active Reinforcement Learning for Color Enhancement

Yuanhang Gao, Qi Zhu, Yutian Fu, Dong Liang

Abstract: Existing learning-based color enhancement approaches focus on improving quality by learning a mapping from input images to enhanced images. However, they ignore the progressive nature of the enhancement process and fail to consider the aesthetics of the images, resulting in visually unappealing outcomes. In this paper, we propose AD-ARL, an aesthetics-driven active reinforcement learning model for color enhancement, to explicitly model the progressive nature of the expert enhancement process. We treat color enhancement as a Markov Decision Process, where each pixel, acting as an agent, actively explores based on the input image and environmental feedback and determines the optimal action that maximizes the reward. After multiple iterations, the optimal sequence of enhancement actions has been obtained. Furthermore, we utilize image aesthetic scores to compute reward and provide aesthetic feedback, ensuring that the enhanced image aligns with human aesthetic perception. Extensive experiments show that our method effectively enhances images in terms of visual perception and image quality. Our results on various benchmarks demonstrate the superiority of our approach over state-of-the-art methods.