Normalizing Flow Learning Based Restoration for Low-Light Images with Massive Zero-Element Pixels

The low-light image enhancement task learns the mapping of low-light images to normal images by training paired or unpaired images, but it does not restore the missing areas composed of massive zero-element pixels(ZEP). In this paper, we propose a normalizing flow learning based restoration for low-light images with massive zero-element pixels(LLI-MZEP), named ZMAR-SNFlow, which is composed of a zero-element mask attention based Restormer (ZMAR) encoder and a strengthened normalizing flow (SNFlow). A new attention mechanism, zero-element mask attention (ZMA) is proposed in mask-based Transformer blocks to repair missing regions. Existing image restoration methods are trained with multiple loss functions, but we use a single negative log-likelihood loss for training, so that SNFlow can learn the conditional distribution of high-light images. Experimental results on three benchmark datasets show that the proposed ZMAR-SNFlow method can achieve state-of-the-art (SOTA) performance in low-light image restoration tasks.