目的 唐卡作为人类非物质文化遗产热贡艺术的重要表现形式之一，承载着重要的历史文化价值。在实地采集过程中发现，由于保存条件有限，许多唐卡作品出现了裂痕、破损、水渍、霉点等问题，传统手工的修复方式效率低，且存在导致唐卡二次受损的风险。此外，使用传统图像修复方法和基于深度学习的图像自动修复方法，在修复唐卡时，往往会产生不符合唐卡纹理结构的不合理结果。因此提出了一种线稿引导的交互式唐卡图像修复网络LSFNet。方法 该方法由三部分组成，首先是唐卡艺术家指导的交互式线稿修复，使得修复的线结构更加接近真实唐卡图像；其次是风格纹理修复阶段，通过构建空间风格纹理模块，来学习唐卡图像整体风格和特征，并结合通道注意力和全连接层，捕获全局信息并进行融合，获得初步修复特征；最后是精修复阶段，引入线性注意力模块，实现全局信息传递，增强模型对唐卡图像内容的理解能力。结果 以在青海省黄南州同仁市实地采集到的唐卡图像为基础，创建了唐卡修复数据集，并通过模拟破损区域，制作了掩码数据集，在创建的数据集上进行训练测试。与DeepFillv2，EdgeConnect，DFNet，HiFill，T-Former等几种图像修复方法进行定量、定性和主观实验对比分析。结果表明，该方法有良好的修复效果，在唐卡数据集上的PSNR(peak signal to noise ratio)、SSIM(structural similarity)和LPIPS(learned perceptual image patch similarity)三个评价指标结果均优于其他对比方法。与性能第2的模型相比，PSNR和SSIM分别提高了10.55%和1.8%，LPIPS降低了57.98%。此外还进行了消融实验进一步验证了交互式线稿修复，风格纹理修复和精修复三个模块的有效性。结论 本文通过采用交互式线稿修补的方法，能够有效地对破损唐卡图像进行修复，获得符合唐卡内容风格的修复结果。

Objective The Regong art originates from the Longwu River valley in the Tibetan region of Huangnan, Qinghai Province, and has flourished there, forming a unique regional artistic style. In 2009, it was inscribed on the UNESCO Representative List of the Intangible Cultural Heritage of Humanity. As one of the significant manifestations of Regong art, Thangka carries the rich historical and cultural heritage of the Tibetan region, which possesses significant historical, cultural, and artistic values. During the field collection process, we discovered that due to limited preservation conditions, many Thangka works exhibited cracks, tears, water stains, mold spots, and other issues. However, traditional restoration methods are not only inefficient but also prone to causing further damage to the Thangka, which is not conducive to the inheritance and development of Regong art. Therefore, there is an urgent need to conduct research on image restoration for the collected damaged Thangka images. However, when attempting to restore Thangka images using currently popular enhancement and restoration algorithms, we encountered issues such as blurred texture lines and misaligned repairs. This is primarily because the complexity and diversity of Thangka make it difficult for models to learn its unique structural and textural characteristics. Method Therefore, an interactive thangka image restoration network LSFNet guided on line draft repair is proposed.This method consists of three parts. Firstly, the interactive line restoration guided by Thangka artists makes the restored line structure closer to the real Thangka image; Secondly, there is the style and texture restoration phase, where an overall style and texture module is constructed to learn the overall style and characteristics of Thangka images. By integrating channel attention and fully connected layers, this phase captures global information and fuses it to obtain preliminary restoration features. Lastly, the refinement restoration phase introduces a linear attention module during downsampling to capture local and global dependencies, acquire features of different scales, further refine the restoration, eliminate restoration traces, and enhance the image restoration effect. PatchGAN is adopted as the discriminator, dividing the input image into multiple receptive fields and conducting independent binary classification judgments for each receptive field to determine whether it possesses the texture characteristics of the target image. This approach effectively enables pixel-level supervision of the generated image, improving the quality of image restoration. Result This paper created a Thangka restoration dataset with a total of 25,000 images using field research and collected data. The Canny algorithm was employed to extract edge line arts, resulting in a line art dataset. Additionally, Photoshop tools were used to simulate damaged areas of the Thangka, generating 5,000 mask maps. Another 1,000 mask maps were selected from public mask datasets to enhance restoration capabilities under different damage conditions, combining them into a mask dataset containing 6,000 mask maps. All datasets have a resolution of 256×256 pixels. The proposed method in this paper was trained, tested, and compared with other restoration methods such as DeepFillv2, EdgeConnect, DFNet, HiFill, and T-Former on the dataset created in this study. The results indicate that this method has good repair performance, and its PSNR (peak signal to noise ratio), SSIM (structural similarity), and LPIPS (learned perceptual image patch similarity) evaluation metrics on the Tangka dataset are all superior to other comparative methods. Compared with the performance second model, PSNR and SSIM increased by 10.55% and 1.8% respectively, while LPIPS decreased by 57.98%. Experimental results demonstrate that the proposed interactive Thangka image restoration method based on line art repair can effectively restore damaged Thangka images, and the restoration results are closer to authentic Thangka images. Conclusion This article uses an interactive line drawing repair method, guided by Thangka artists, to repair damaged areas of the line drawing. Then, through style texture repair, the Thangka image style features are learned. Finally, through fine repair, the repair results are further optimized. The experimental results show that this method can effectively repair damaged thangka images and obtain repair results that conform to the style of thangka content.