Reversible and Cascaded Lightweight Colour Constancy: Jointly Addressing Illumination Correction and White Balance

Abstract

Colour constancy aims to preserve the perceived colours of objects regardless of variations in illumination intensity and colour temperature. Colour-based algorithms are profoundly influenced by colour constancy. Even with colour augmentation applied during training, high level computer vision tasks such as object detection and image classification are also impacted by colour constancy. Therefore, colour constancy, influenced by both illumination intensity and colour temperature, should serve as a pre-processing step for downstream tasks. These two influence factors aim to two tasks, illumination correction and white balance, which can also mutually affect each other. In this context, we propose a lightweight model that jointly considers white balance and illumination correction to meet the efficiency requirements of pre-processing. Our proposed model enhances and corrects input images using a reversible formula that separately estimates the parameters of linear and nonlinear operations. We employ a cascaded architecture with learnable residual weighted connections to enable modules to focus on their calculation targets and avoid being misled by previous modules during training. Our model achieves state-of-the-art results on exposure correction datasets, competitive results on extremely low-light datasets, and comparable results on white balance datasets with such an over 25 times smaller size, which is only 689KB. Extensive experimental results demonstrate competitive performance of our proposed model in the field of colour constancy, with lightweight architecture showcasing its efficiency benefits.