Retinex-Inspired Cooperative Game Through Multi-Level Feature Fusion for Robust, Universal Image Enhancement

Existing approaches to enhancing distorted images frequently grapple not only with the dual challenges of optimizing visual fidelity and computational efficiency but also tend to be ineffectual in uncharted and intricate scenarios. Herein, we present a Retinex-inspired cooperative game based image restoration technique termed RICG to address the difficulty of navigating model performance and efficiency in different kinds of environments within a unified model. Specifically, we propose a two-step pipeline, comprising self-supervised illumination disentanglement and adjustment. The zero-shot illumination disentanglement is trained through a novel camera response Transformer (CRT), followed by illumination adjustment using a dual-discriminator feature pyramid network (DDFPN) incorporating an self-attention regularization. It is worth mentioning that we devise a specialized training process to reconstruct the optimal restored image through cooperative game. We substantiate the diverse advantages of RICG over existing methods through a meticulous and comprehensive evaluation process, illustrating its versatility in unexplored and convoluted circumstances. (Implementation code can be accessed at https://github.com/Ruiqi-Mao/RICG).