ZSDECNet: A zero-shot deep learning framework for image exposure correction

Abstract

When shooting street scenes at night, the captured images may be underexposed or overexposed, which seriously affects human visual perception. Therefore, exposure correction is required for these images. Most existing exposure correction methods rely heavily on reference images and the exposure correction results are not thorough. To address these issues, we propose a zero-shot multi-exposure correction method based on S-curves, called ZSDECNet. Our method is divided into two parts: multi-exposure correction and fusion. First, the illumination channel of the image and the corresponding inverted image are subjected to an initial exposure correction. Moreover, by exposure fusion technique, we select the best exposed area for exposure fusion from the exposure correction results of the input image, the illumination channel, and its inverted channel in order to obtain a visually optimal exposure-corrected image. The method is a zero-shot end-to-end training approach that does not require additional data. In addition, the S-type exposure correction curve corrects both underexposed and overexposed areas, which makes it possible to obtain more thorough exposure correction results. Experiments on existing datasets with various exposure conditions (underexposed and overexposed) and on a real nighttime street scene dataset show that our method outperforms state-of-the-art methods.