Supervisors: Dr. Radu Timofte
As the quality of cameras integrated into portable devices improves every year, so does the the technology behind dedicated, system cameras that have far less constraints regarding size and battery consumption. Hence, the task of improving the quality of images taken by compact cameras remains one of the principal tasks in computer vision. The existing approaches to end-to-end image enhancement use standard convolutional neural network architectures and leave it to these complex structures to learn the particular mapping from low to high quality images. However, it is possible to model the mapping between input and target domains by adding specialized blocks that incorporate our a priori knowledge about the domains. In this work, we introduce explicit linear and gamma corrections that take advantage of our observation that the pixel intensities in images captured by DSLR cameras are spread more uniformly across the intensity levels that it it is the case for images captured by integrated cameras. We evaluate multiple network structures trained using different losses and achieve improved perceptual quality in comparison to the existing approaches.