Research On Facial Image Transferring Based On Features Extracted In Frequency Domain

Thanks to the emergence of low-cost image acquisition device and the rapidly developed computer technologies,as well as the increasing demands of games,video and humancomputer interaction applications,facial expression artistic style transferring has become a hot research topic.Many factors lead to the diversity of natural face images.Thus,it is challenging for non-professional users to transfer facial expression and artistic style.Based on the study of the frequency domain characteristics of facial images,we propose two novel methods to transfer facial expression and artistic style.The main achievements are summarized as follows.A novel color facial expression transferring method based on frequency analysis and modified Poisson image editing is proposed to synthesize photo realistic facial expressions with source-object specific expression details.First,we calculate the normalized relative shape movements between the neutral and non-neutral expressions of the source object to generate the non-neutral shape of the target object.We warp the non-neutral expression of the source object and neutral expression of the target object to the non-neutral shape of the target object.The warped target expression does not contain any expression details.Then,we choose YUV color space to separate the luminance(Y channel)and chrominance(U and V channel)components of expression images.In its Y channel,the source expression image is transformed into frequency domain through Fast Fourier Transform.Then,the expression details are extracted from the source object through analyzing their frequency distribution.Finally,the extracted expression details are stitched smoothly to the warped target object with the modified Poisson image editing method.The transferred result is combined with the warped target expressions in U and V channels to synthesize a colorful realistic target expression,which reduces the effect of illumination and skin color variations.A novel facial image artistic style transferring method based on the analysis of multi-band frequency gains is proposed.First,given a natural input facial image and a sample facial image drawn by an artist,we assume that the input and the sample facial images have approximately similar poses and facial expressions.We establish a pixel-wise dense correspondence between the input and the sample faces.Then,the input image and the sample image are decomposed by multi-frequency analysis,respectively.The gain map is calculated in each frequency band,which achieves the artistic style from sample image relative to the input image.Finally,we transfer the background by directly replacing the input background with the sample background.This method makes it easy for people to transfer the facial image with a particular style.