Artistic style migration method and system based on frequency domain modulation and attention guidance

By using frequency domain modulation and attention guidance, high-frequency structure and low-frequency style components are separated, and parallel cross-attention branches are constructed. This solves the problems of feature dilution and initial feature coupling in art style transfer, and achieves high-quality style transfer results.

CN122453594AActive Publication Date: 2026-07-24SHANDONG UNIV OF SCI & TECH
5 Cites 0 Cited by

Patent Information

Application Number
CN202610920854.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-25
Publication Date
2026-07-24
Estimated Expiration
2046-06-25

AI Technical Summary

Technical Problem

Existing art style transfer methods suffer from feature dilution and initial feature coupling during the generation process when dealing with high-resolution or style-diverse tasks, resulting in insufficient stylization of the generated image or visual artifacts.

Method used

We employ a frequency domain modulation and attention-guided approach, using two-dimensional fast Fourier transform to separate high-frequency structure and low-frequency style components. We construct parallel cross-attention branches to independently process text, content, and style features, and utilize the gradient backpropagation of style-guided energy for latent variable optimization.

Benefits of technology

It effectively decouples content structure and style features, achieving high-quality artistic style transfer. The generated results retain the source image structure while possessing deep style semantics, improving the injection efficiency of style textures and color expressiveness.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention belongs to the technical field of image processing, particularly relates to an artistic style migration method and system based on frequency domain modulation and attention guidance, and aims to solve the problems that an existing method is difficult to effectively separate a content structure from style features, and structural damage or insufficient stylization of a generation result is easily caused. The method comprises the following steps: firstly, realizing effective decoupling of contents and styles from a frequency domain based on a frequency domain modulation initialization strategy of structure perception; a decoupling cross attention mechanism is constructed, and text, content and style features are respectively processed through parallel channels, so that feature interference is avoided, and meanwhile, accurate regulation and control on a generation process are realized; according to the method, an attention-driven guidance mechanism is utilized, style migration is modeled as an iterative optimization problem in a submerged space, and feature distribution is dynamically calibrated through gradient guidance, so that the problem of feature dilution in a deep network is effectively relieved.
Need to check novelty before this filing date? Find Prior Art