A light-weight deep learning framework for Low Light Image Enhancement

In computer vision, deep learning-based methods for improving low-light images have gained popularity. The proposed lightweight end-to-end deep neural network architecture is designed by minimizing the number of trainable parameters while optimizing design choices for efficiency and ensuring fast inference time. The proposed architecture consists of denoising, enhancing, and fusion modules designed to enhance image visibility, and contrast and reduce noise while preserving content and color information. We used a modified convolutional neural network (CNN)-based framework for exposure fusion that is designed to identify and rectify hidden degradation within dimly light images and highly adaptive to diverse lighting conditions. However, after conducting quantitative experiments, we have found that the proposed method outperforms the state-of-the-art TTST by about 0.48 dB and EDiffSR by 1.48 dB. Our lightweight method accounts for 8.28% and 6.77% of the computational cost (FLOPs) of TTST and EDiffSR respectively, and requires just 1.91% and 1.35% of their trainable parameters additionally.