Low-Light Image Enhancement with Normalizing Flow

To enhance low-light images to normally-exposed ones is highly ill-posed, namely that the mapping relationship between them one-to-many. Previous works based on pixel-wise reconstruction losses and deterministic processes fail capture complex conditional distribution of normally exposed images, which results in improper brightness, residual noise, artifacts. In this paper, we investigate model one-to-many via a proposed normalizing flow model. An invertible network takes images/features as condition learns map into Gaussian distribution. way, can be well modeled, enhancement process, i.e., other inference direction network, equivalent being constrained by loss function better describes manifold structure natural during training. The experimental existing benchmark datasets show our method achieves quantitative qualitative results, obtaining better-exposed illumination, less noise artifact, richer colors.