Work

ST-GVMNet: a structure-texture dual-stream network fusing gating mechanism and V-Mamba for Dunhuang mural image inpainting

Year: 2026

Type: article

Abstract: Dunhuang murals are afflicted by complex degradation phenomena, including extensive content loss, blurred textures, and pervasive mould-spot corrosion. Existing image inpainting methods often fail to handle these intricate degradation patterns effectively, primarily due to their limited ability to model long-range dependencies, susceptibility to structural flow noise, and inability to preserve high-frequency details. To address these challenges, we propose ST-GVMNet, a Structure-Texture Dual-Stream Network that synergistically integrates Gating Mechanisms with the Visual Mamba (V-Mamba) architecture for Dunhuang mural inpainting. Adopting a coarse-to-fine, decoupled strategy, the network consists of two core components: a Structure Reconstruction Network (SRN) and a Texture Refinement Network (TRN). In the SRN stage, we develop a Structure Gate Fusion Module (SGFM) to adaptively filter structural features and suppress edge noise through spatially weighted gating operations. For texture inpainting, we introduce a Multi-scale Dilated VMamba (MD-VMamba) module, which leverages 2D selective scanning and dilated feedforward blocks to capture both global contextual information and fine-grained local details. Furthermore, a Variance-Gated Aggregation Module (VGAM) is proposed to dynamically recover high-frequency nuances by utilizing local variance as a saliency indicator. Extensive experiments on the Dunhuang mural and Thangka datasets, as well as standard benchmarks, demonstrate that ST-GVMNet achieves state-of-the-art (SOTA) performance in both quantitative metrics and qualitative visual fidelity. Notably, our method exhibits exceptional structural coherence and robustness when inpainting images under extreme mask ratios.

Source: Journal of King Saud University - Computer and Information Sciences

DOI: 10.1007/s44443-026-00858-5

Authors: Rui Tian, Tongchen Wu, Dandan Feng, Lulu Wang, Zihao Xin

Institution: Lanzhou Jiaotong University

Language: English

Topic: Generative Adversarial Networks and Image Synthesis

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $1,990

Funder: National Natural Science Foundation of China