A Deep Convolutional Neural Network Based Approach to Extract and Apply Photographic Transformations

Abstract

Sophisticated image editing techniques like colour and tone adjustments are used to enhance the perceived visual quality of images and are used in a broad variety of applications from professional grade image post-processing to sharing in social media platforms. Given a visually appealing reference image that has some photographic filter or effects applied, it is often desired to apply the same effects on a different target image to provide it the same look and feel. Interpreting the effects applied on such images is not a trivial task and requires knowledge and expertise on advanced image editing techniques, which is not easy. Existing deep learning based techniques fail to directly address this problem and offer partial solutions in the form of Neural Style Transfer, which can be used for texture transfer between images. In this paper, a novel method using a convolutional neural network (CNN) is introduced that can transfer the photographic filter and effects from a given reference image to a desired target image via adaptively predicting the parameters of the transformations that were applied on the reference image. These predicted parameters are then applied to the target image to get the same transformations as that of the reference image. In contrast to the existing stylization methods, the predicted parameters are independent of the semantics of the reference image and is well generalized to transfer complex filters from the reference image to any target image.