Dual-branch CNN–transformer with curve estimation and shadow-aware colour correction for low-light image enhancement

Low-light images are perceived to be of lower quality and less efficient in terms of visual performance due to poor visibility, excessive noise levels and colour inaccuracies. Conventional enhancement techniques tend to lose detail while maintaining a balance of natural hues, and many modern deep learning models struggle with varying illumination conditions and dark shadows. In this paper, we present a dual-branch CNN-Transformer (DBCT) architecture that combines the local feature extraction capabilities of a convolutional neural network with the global contextual awareness of a Transformer. The network divides the input images into two parallel paths: a Reflectance branch to restore the fine structure and an illumination branch to increase the brightness. We evaluated this method on well-known paired and unpaired datasets that have noise and colour loss in real-world low-lighting conditions. The results consistently show that our method makes the images brighter, more clearly illuminated and have more natural colours. The new framework is consistently better than the current methods for all quantitative measures and also provides better visual output, particularly in dark areas where the traditional methods often lack detail and tone fidelity.