QR code arbitrary style transfer algorithm based on style matching layer

doi:10.1007/s11042-023-17231-7

Return

QR code arbitrary style transfer algorithm based on style matching layer

2023-10-05

PRE

Hai-Sheng Li *

Jingyin Chen

Huafeng Huang

DOI：10.1007/s11042-023-17231-7

Original

Original request for help

Save

Abstract

En 中文

A single style transfer model can realize the Quick Response (QR) code beautification task. However, the number of styles in the single-style transfer model is limited, and the training must be carried out in advance to expand to new styles, which requires a lot of training time and memory. Therefore, the flexibility of the model needs to be further improved. The current mainstream arbitrary style transfer algorithm solves the flexibility problem and can complete the transfer task of any style after one training. Unfortunately, the existing arbitrary style transfer algorithms have issues such as heavy computational burden, distorted content, and blurred details, which are unsuitable for QR code beautification tasks. This paper proposes an arbitrary style conversion algorithm based on a style-matching layer to solve the above problems. Our algorithm obtains the style information using a small convolutional neural network and then transmits the second-order statistics of style images through a linear matrix multiplication to achieve arbitrary style transfer. Simulation results show that the proposed algorithm can efficiently complete the general style transfer task and has a higher visual quality and testing speed than the existing arbitrary style transfer models. Meanwhile, the algorithm can flexibly combine multi-level style features to complete the QR code style beautification task. Furthermore, there is a great affinity between the image generated by the proposed algorithm and the original image without distorted content and fuzzy details. Therefore, the generated style beautified QR code has high visual quality and recognition.

Keywords:

QR code

Image style transfer

Deep learning

Convolutional neural network