Exposure is a fundamental component of a good picture. However, it can be quite challenging to set the camera parameters to get it right. Overexposed shots can be corrected, but this also demands some expertise. In this work, we try to show that this correction can be automated using deep learning. We use conditional adversarial networks in order to correct overexposed images. We mainly build our method on previous work from \cite{isola2017image} that introduced a successful GAN archithecture for image-to-image translation problems. On top of that we make use of more recent techniques to improve the quality of the reconstructions, such as spectral normalization, noise injection and perceptual loss.
Pierre Le Jeune
PhD Student in Deep Learning
My research interests include computer vision, deep learning and applications in low-data regime.