Abstract

Abstract: Photorealistic Style transfer models havespecified promising artistic results. This system introduces a deep-learningapproach to photorealistic style transfer that handles an outsized sort ofimage content while transferring the reference style. The approach used in thispaper is build upon the recent work on painterly transfer that separates stylefrom the content of a picture by considering different layers of a neuralnetwork. A neural algorithm of artistic style proposes an iterative algorithmfor neural style transfer. The neural style transfer algorithm involves takinga content image C and a style image P and combine them to produce a new imagethat has the content of C and the style of P. Photorealistic image stylizationinvolves transferring style of a styled photo to a content photo with thelimitation that the stylized photo should remain photorealistic. However, ifgiven a photograph as a reference style, present methods are limited by spatialbiases or improbable artefacts, which should not happen in real photographs.The proposed network architecture enhances photorealism and transfers thestyle. The key factor of our method is wavelet transforms that naturally fitsin deep networks. To preserve the structural information and features we haveproposed a wavelet corrected transfer method in this paper that supportwhitening and colouring transform.

Keywords: Image stylization, Photorealism, Haar wavelet, VGG