基于流匹配的图像快速修复方法
By enhancing the encoder's collaborative architecture through rectified flow and multi-expert knowledge, and combining linearized ordinary differential equations and a stochasticity elimination module, the bottlenecks of traditional flow matching methods in terms of restoration efficiency and quality are solved, achieving fast and high-fidelity image restoration. This is applicable to medical image reconstruction, digital preservation of cultural heritage, and image restoration for autonomous driving perception.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DALIAN NATIONALITIES UNIVERSITY
- Filing Date
- 2025-10-27
- Publication Date
- 2026-07-17
AI Technical Summary
Existing image inpainting techniques face significant bottlenecks in balancing inpainting efficiency and quality. Traditional flow matching methods lack adaptive modeling of damage patterns, resulting in a disconnect between the inpainting results and the semantics of the original image. Furthermore, they are computationally expensive and difficult to meet the needs of real-time applications.
We employ rectified flow as the core of flow matching, and combine it with a multi-expert knowledge-enhanced encoder and a randomness elimination module to construct a collaborative architecture. We optimize the flow matching process through multi-scale feature pyramids, dynamic expert routing, and linearized ordinary differential equations. We train the system using multi-attention fusion and composite loss functions to achieve fast and high-fidelity image restoration.
It significantly reduces sampling steps and computational costs, improves the semantic consistency and detail integrity of the restoration results, meets the needs of efficient and high-fidelity restoration, and is suitable for medical image reconstruction, digital protection of cultural heritage, and restoration of perception images for autonomous driving.
Smart Images
Figure CN121563840B_ABST