Arbitrary Style Transfer of Facial Image Based on Feed-Forward Network and Its Application in Aesthetic QR Code

Abstract

QR code has become essential in daily-life because of the popularity of mobile devices. The visual effect of a conventional QR code is not ideal. Consequently, many good aesthetic algorithms have been proposed. However, both the decoding rate and visual effect of a QR code cannot be guaranteed simultaneously when facial image serves as the background. We propose an arbitrary style transfer of facial image based on feed-forward network as a preprocessing algorithm for an aesthetic QR code. The deep characteristics of content image and style image are unified in the same layer of convolutional neural networks in our style transfer network. Styles are changed. The result of style transfer is restricted with semantic segmentation result, color uniform regularization of facial image and repeating restriction similarity constraints. Experimental results show that both the decoding rate and visual effect of a QR code are guaranteed when our method is used in background preprocessing.