Implementation of Adjustable Light Enhancement Based on Deep-Retinex-Net

Abstract

The Deep-Retinex-Net (DRN) method enhances lowlight images by decomposing them into reflectance and illumination components, which are independently processed using convolutional neural networks (CNNs) and then recomposed. However, a significant limitation of this approach is its inability to adaptively adjust output brightness according to user-specific requirements, this is because the mapping relationship it established is fixed on the light and dark relationship of the image pairs in the dataset. To address this issue, this paper proposes the Control-Deep-Retinex-Net (C-DRN) framework. The proposed method introduces modifications in dataset construction, optimizes the training strategy, and refines the network architecture. By incorporating a maximum likelihood estimation (MLE) model, C-DRN enables continuous and customizable brightness adjustment—from extremely dark to normal lighting conditions—during the low-light enhancement process. Experimental results demonstrate that the framework achieves 96.3% accuracy in brightness control, significantly improving the model’s adaptability and learning capability. The approach not only produces visually pleasing results but also supports precise and user-oriented illumination customization.