Text Effects Transfer via Statistical Methods

Abstract

Style transfer is the process of transferring appearance features from one given source image to another target picture&rsquos content where the outcome is a new image, which has the style of first image and content of the second one. Text effects transfer or typography is a subset of style transfer in which styles of different text images are transferred. Style transfer between images has also color transfer and texture transfer. Different methods have been represented for style transfer in previous years. These methods try to do color and texture transfer between images by the least amount of error. Color transfer methods transfer color style between two images. Texture transfer methods are classified the parametric and nonparametric methods. In this thesis, nonparametric transfer challenges of text effects are evaluated and three distance-based statistical approaches for different cases are proposed. The first method which is named NPD-TET, transfers various text effects easily and with good accuracy between source image and target image. NPD-TET takes a set that consists of raw source image, the image with the effect of source and target raw image, and produces the image with target effect. In NPD-TET, first stylized patches are identified with normalized distances and optimal scale and then by applying them on the objective function, the new texture is combined with multiple scales. The experimental results show that our approach is better than existing style transfer methods for various text effects and its effectiveness on Persian characters has been examined. Then, Extended NPD-TET algorithm has been proposed, which is a Extended version of the NPD-TET algorithm. In this algorithm, there is no need for raw image of the source to transfer effect and the effect transfer is done by only using two images. One image is the image with source effect and the other one is the raw target image. In this algorithm first the raw image of source is reconstructed completely from the image that has the source effect by using a proposed destylization network. Then images are put into NPD-TET algorithm and text effect transfer is done. We also proposed a special and more comprehensive method named TES-CI where exchenges the text effects between two stylized images where there is no raw image. Because of complete reconstruction of raw images by destylization network and also use of NPD-TET algorithm this method optimizes text effets switching and preserves local and global information.

Key Words : Style Transfer, Text Effect Transfer, Typography, Statistical Method, Non Parametric Method, Effect Switching, Persian Character.