Brightness-aware neural adaptation: a hybrid approach for time-dependent low-light image enhancement

Abstract

Low-light image enhancement (LLIE) is essential for improving the visibility and interpretability of images captured under poor or insufficient lighting conditions, which is critical for applications such as autonomous driving, surveillance, and photography. However, most existing supervised approaches depend on large paired datasets that are expensive to collect ainid often fail to generalize across diverse lighting conditions and degradations. In this thesis, we propose a self-supervised learning approach for LLIE using a Time-Aware Zero-Reference Deep Curve Estimation (Zero-DCE) model. Our method eliminates the need for paired training data by learning to enhance images through a self-supervised mechanism that estimates pixel-wise adjustment curves directly from the input. This design makes the model lightweight, efficient, and suitable for real-time enhancement tasks. We gathered a custom dataset in order to train and test our model. Pictures were taken between 3 a.m. and 9 p. m. with five minutes intervals to permit the model to acquire knowledge by constant changes of lighting during the day. By including the awareness of time in the self-supervised process of learning, the model adjusts its improvement strategy to the changing light conditions, which leads towards an improvement in luminance, contrast and hue stability and maintain structural acuity. The proposed framework demonstrates strong potential for real-world deployment in settings where labeled data are limited and lighting is suboptimal, including autonomous vehicles, surveillance systems, and underwater imaging.