A high-fidelity and efficient image restoration method
By designing a dual attention fusion module and a gated convolutional feedforward network, combined with a global mask caching mechanism, the problems of insufficient fusion of shallow and deep features and computational redundancy in image restoration are solved, achieving high-fidelity and efficient image restoration results.
CN122415365APending Publication Date: 2026-07-17JINLING INST OF TECH
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINLING INST OF TECH
- Filing Date
- 2026-05-22
- Publication Date
- 2026-07-17
Smart Images
Figure CN122415365A_ABST
Abstract
本发明公开了一种高保真度且高效的图像复原方法，该方法采用双注意力融合模块并行利用通道与空间注意力对浅层纹理细节与深层语义信息进行自适应加权融合，并在前馈网络中引入深度卷积与门控机制实现动态特征筛选，同时通过以输入配置为键的全局掩码缓存机制消除移位窗口注意力中掩码的重复生成开销；本发明通过构建上述协同架构，解决了现有Swin Transformer类方法在深浅特征融合不足、局部感知缺失与掩码计算冗余方面的问题，大幅提升了复原图像的纹理保真度与计算效率。
Need to check novelty before this filing date? Find Prior Art