基于生成对抗网络的纺织品文物虚拟修复方法研究OA

Currently,the virtual restoration of textile artifacts primarily relies on traditional algorithms such as the Criminisi algorithm.Since these traditional approaches fill missing regions by exploiting local pixel similarity,they exhibit significant limitations when dealing with textile artifacts that feature complex and irregular patterns,especially in restoration tasks involving large-scale missing areas.To address these challenges,this study proposes a restoration model that integrates the multi-scale feature fusion capabilities of U-Net with the multi-head attention mechanism of Transformer networks,facilitating high-fidelity restoration for textile artifacts characterized by intricate patterns and extensive damage. A U-Net encoder-decoder architecture is employed for the restoration model.The encoder performs progressive downsampling with a stepwise increase in the number of attention heads to achieve feature extraction of textile relics,spanning from local details to global semantics.Conversely,the decoder performs progressive upsampling with a stepwise decrease in attention heads,leveraging skip connections to facilitate multi-scale content reconstruction of missing areas from a global-to-local perspective.To address the highly complex patterns and chromatic characteristics of textiles,a dual-branch parallel spatial-channel attention layer is introduced.Specifically,the channel branch performs weighted feature fusion via self-attention weights,which enhances critical features while suppresses redundant information.Simultaneously,the spatial branch expands the receptive field via dilated convolutions to capture extensive multi-scale contextual information.The two branches work synergistically to guide the model to focus on the most task-relevant features amid the complex information of textiles.To mitigate information loss during conventional downsampling and to adapt to the intricate damage patterns of textile relics,a synchronized feature-mask downsampling module is implemented.This module performs Pixel Unshuffle operations concurrently on both the feature maps and the damage masks,interlacing them along the channel dimension.This ensures the model maintains a continuous perception of the artifacts'damaged areas even in deeper network layers. Experimental results demonstrate that the proposed restoration model consistently achieves stable performance on unseen textile relics with damage ratios ranging from 0.01%to 60%.Even in high-damage scenarios(40%-60%)where reference information is severely limited,the model maintains a robust restoration capability.Comparative experiments verify that our model outperforms state-of-the-art deep learning models across various evaluation metrics,demonstrating its superiority in the specialized task of textile relic restoration.Furthermore,ablation studies confirm the efficacy of the model's core components.The removal of either the spatial-channel attention layer or the damage-aware downsampling module leads to a decline in performance.This validates that these modules contribute critically to handling complex textile features and irregular damage patterns. This study systematically establishes a deep learning restoration paradigm tailored for textile relics,providing an effective new technical approach for their virtual restoration.Its neural network architecture design offers valuable reference for the future research and development of textile relic restoration models.