Reference-free low-light image enhancement by associating hierarchical wavelet representations

For computer vision and image content understanding, the low-light image becomes an obvious challenge as it suffers from poor contrast and illumination. Hence, low-light image enhancement(LLIE) technology has attracted considerable attention. However, most of the existing LLIE methods based on deep learning technology rely on referenced images to guide supervised training. It is very difficult to capture the referenced images in a real-world scene. To remedy this, we propose a reference-free low-light enhancement framework by estimating pixel-wise curves with wavelet decomposition and association. Typical LLIE methods apply illumination adjustment on the RGB image roughly. In contrast to that, we re-examine current LLIE pipelines and propose a fine-grained image enhancement framework by abstracting and associating frequency prior. Firstly, we decompose the image to the frequency domain by wavelet transform. Then, we realize frequency-away message passing with hierarchical wavelet decomposition. Finally, reference-free loss functions are applied with the consistent association between wavelet frequencies. These losses are used to guide a fine-grained reference-free LLIE paradigm. We verify our method through comprehensive experiments, and prove that our model is superior to other relevant methods under the quantitative index and visual perception.