RECDFS: Rotation Equivariant Convolution and Delayed Feature Statistics for arbitrary neural style transfer

Sun, Ya-Qi, Xiang, Jin, Rosin, Paul L. ORCID: https://orcid.org/0000-0002-4965-3884, Lai, YuKun ORCID: https://orcid.org/0000-0002-2094-5680, Zhao, Hui-Huang, Li, Zhi and Xie, Xiao-Lan 2026. RECDFS: Rotation Equivariant Convolution and Delayed Feature Statistics for arbitrary neural style transfer. Journal of King Saud University - Computer and Information Sciences 38 , 366. 10.1007/s44443-026-00795-3

PDF - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB)

Abstract

The aim of arbitrary style transfer is to render an image with the style of a reference image while preserving the original content. A key challenge is the extracting and maintaining global image information due to the local nature of convolutional neural networks (CNNs). Therefore, arbitrary style transfer methods incorporate a limited representation of the content. Furthermore, incoherence in local style transfer also contributes to the generation of low-quality images. In order to significantly reduce content bias and generate high-quality images, we introduce a novel arbitrary style transfer approach named RECDFS (Rotation Equivariant Convolution and Delayed Feature Statistics). In detail, we propose a Consistent Neural Content Mapping (CNCM) module based on Rotation Equivariant Convolution and attention mechanism. CNCM enhances feature extraction and attends to key-style representations. Based on CNCM, we design a delayed global feature statistics (DGFS), which is used to align the second-order statistics of content and style features. Additionally, we design two loss functions to optimize detail enhancement. Importantly, our method achieved state-of-the-art results compared with twelve classic methods.

Item Type:	Article
Date Type:	Published Online
Status:	Published
Schools:	Schools > Computer Science & Informatics
Publisher:	Elsevier
ISSN:	1319-1578
Date of First Compliant Deposit:	7 May 2026
Date of Acceptance:	18 April 2026
Last Modified:	04 Aug 2026 13:09
URI:	https://orca.cardiff.ac.uk/id/eprint/186841

Actions (repository staff only)

Edit Item