Research On Stylization Of QR Code Based On Deep Learning

Quick response code(QR code)is a two-dimensional code,which can be quickly identified by portable mobile devices.It plays an increasingly important role in e-commerce,manufacturing,marketing and daily life.Therefore,the visual beauty of QR code as an important performance worthy of improvement has attracted people's attention.In recent years,the research on QR code beautification algorithms is emerging at home and abroad,and these algorithms have achieved good results.However,most of these methods have some limitations,and tend to enhance the decoding rate of QR code or specialize the visual beauty of QR code.In order to solve this problem,this paper proposes a QR code stylization algorithm based on deep learning method,which not only beautifies the visual effect of QR code,but also ensures the decoding robustness.At the same time,the algorithm has the advantages of fast convergence,parameter reduction,end-to-end training,and can quickly generate stylized QR code.The main work of this paper is as follows:(1)For the task of QR code image stylization,a style modeling method based on the maximum mean deviation of kernel function is proposed,and the style modeling method is applied to QR code image.According to the characteristic distribution of QR code image,a style migration network structure suitable for QR code image is constructed.Firstly,a Gauss model QR beautification algorithm is used to fuse QR code and background image,and then the fused QR code image is input into the trained style migration network to generate the stylized QR code image.(2)Aiming at the problem of QR code decoding rate decreasing due to the stylization process of QR code image,a QR code error correction algorithm based on gradient descent optimization is proposed.A global loss function and a local loss function are designed to balance the stylization degree and decoding robustness of QR code image.By adding these two loss functions to the training of style migration network,we can effectively guarantee the visual effect of style and enhance the decoding robustness of style QR code.In particular,by adjusting the radial range of the Gaussian kernel of the loss function,the consistency of the convergence region is ensured,and the training process is more rapid and stable.An improved image denoising algorithm is proposed to reduce the artifacts and noise caused by QR code image stylization.This algorithm designs a Gaussian kernelfunction for end-to-end training of style migration network.The loss function can reduce the noise of the QR code image on the premise of less impact on the decoding robustness,and further improve the visual effect of the stylized QR code image.