Dataset Construction and AI Model Development for Line Drawing Restoration of Historic Chinese Murals

Historic Chinese murals possess substantial historical, cultural, and artistic value. However, prolonged exposure to material aging, environmental changes, and human activities can lead to deterioration such as cracking, flaking, and loss [1]. With advances in artificial intelligence, digital methods have increasingly been applied to the analysis and restoration of cultural heritage images [2]. Lines are an important element of artistic expression in Chinese murals, serving to delineate figures and organize pictorial compositions [3]. Extracting line drawings and completing missing lines in damaged murals can help organize surviving pictorial information and provide references for digital research. Previous studies have combined deep learning with image processing to extract line drawings of cultural artifacts and suppress noise [4]. Cracks, flaking boundaries, and boundaries between color regions can interfere with the identification of drawn lines in murals; extraction and denoising methods have been used to reduce such interference [5]. Expanding mural line drawing datasets also remains an active research concern [6]. It is therefore necessary to document image sources, the evidence underlying manual tracing, and the correspondence between source images and line drawings to provide traceable data for model training and evaluation. To address these issues, this study focuses on figures in historic Chinese murals. A paired dataset was established through manual tracing and review, and experiments on line drawing extraction and missing-line completion were conducted using HSPG-Net. The study focuses on data preparation and verification procedures and on the degree of agreement between model-generated outputs and manually traced reference line drawings.

The digital restoration of mural line drawings can be divided into line drawing extraction and missing-line completion. The former identifies surviving drawn lines while suppressing interference from deterioration and the background; the latter completes broken or missing portions based on discernible line structures. In research on general line drawing completion, Sasaki et al. restored incomplete drawings through joint gap detection and inpainting [7]. In mural image restoration, Li et al. used line drawings as structural guidance for the progressive inpainting of damaged regions [8]. These studies provide methodological references for line completion and line drawing-guided mural image restoration, respectively, although their task objectives and input information differ. Different models provide technical baselines for comparative research on mural line drawings. U-Net uses an encoder–decoder architecture with skip connections for pixel-level prediction [9], and modified architectures have been applied to mural crack segmentation [10]. CycleGAN learns mappings between unpaired image domains through a cycle-consistency constraint [11], and its variants have been used for mural image restoration [12]. Pix2PixHD is designed for high-resolution conditional image generation [13], whereas LineDistiller is an open-source line extraction tool for anime, manga, and illustrations [14]. Because these methods were originally developed for different tasks, their training procedures and task adaptations must be specified when they are applied to mural line drawing extraction or missing-line completion.

Images of figures in historic Chinese murals were collected from publicly available digital resources, specialist illustrated catalogs, published literature, and field investigations. Information including site names, mural dates, locations, and image sources was checked. The research team also conducted field investigations and image collection at Yongle Palace in Shanxi and Fengguo Temple in Yi County, Liaoning. In total, 739 images of mural figures were compiled. To address differences in viewing angle, sharpness, resolution, and illumination across sources, working copies of the images underwent cropping, perspective correction, sharpness adjustment, and noise reduction where necessary. Preprocessing prioritized the preservation of discernible pictorial content and spatial relationships among figures, providing an image basis for manual tracing and sample pairing.

The samples represent 12 mural sites, with historical attributions ranging from the Northern Liang period (397–439 CE) to the Ming dynasty (1368–1644 CE). They cover Gansu, Shanxi, Beijing, Hebei, Liaoning, and Tibet and primarily comprise murals in grottoes and Buddhist and Daoist temples. Murals from different historical periods and regions vary in their figure forms and use of lines. Manual tracing should therefore be based on the discernible pictorial information in each mural, avoiding the replacement of its original characteristics with a uniform line style. Corresponding line drawings were manually traced from the verified and processed mural images. The tracing retained artistically meaningful lines, including figure contours, garment folds, and decorative details, while excluding non-pictorial information such as cracks, flaking boundaries, and background textures. Where local details were unclear, specialist illustrated catalogs, published materials, and other clearer images were consulted for cross-checking. No speculative additions were made where sufficient evidence was unavailable. After the initial tracing, the drawings were compared with the source images to identify and correct omissions, positional deviations, and unsupported lines. Figure 1 illustrates the correspondence among the original mural image, the automatically extracted line drawing, and the manually traced line drawing. Some fine lines are faint or missing in the automatic extraction, whereas the manual drawing organizes the discernible drawn lines. The figure illustrates the data preparation procedure and should not be used to generalize about the performance of all automatic extraction methods. The resulting dataset contains 739 pairs of mural images and manually traced line drawings.

The 739 mural image–line drawing pairs were randomly divided into training, validation, and test sets at an approximate ratio of 7:2:1 for parameter learning, monitoring training progress, and final performance evaluation, respectively. All samples were standardized to 1024 × 1024 pixels while preserving the spatial correspondence between each mural image and its manually traced line drawing. Paired data augmentation, including random flipping and rotation, was applied only during training to increase sample variation. The experimental workflow comprises two steps: line drawing extraction and missing-line completion. In the first step, a mural image is used as input to extract surviving figure lines and produce an initial line drawing. In the second step, broken, omitted, and locally missing lines are completed on the basis of the existing line structures. The two steps correspond to the extraction of surviving pictorial information and the prediction of missing line structures, respectively. This study used manually traced line drawings as references to compare model outputs in terms of spatial correspondence, the correct distinction between lines and background, and image structural similarity. Mean intersection over union (mIoU) summarizes the overlap of regions across classes, whereas intersection over union for black line regions (Black IoU) focuses on the overlap of the line regions themselves [15]. Pixel accuracy (Pixel Acc.) is the proportion of pixels correctly classified overall. Mean class accuracy (Mean Acc.) first calculates the accuracy of each class and then averages these values, so that the less numerous line pixels are also represented in the evaluation [16]. Structural similarity (SSIM) compares the generated images and manually traced references in terms of luminance, contrast, and local structure [17]. Under the same calculation settings, higher metric values generally indicate that the generated results are closer to the manually traced references in the corresponding aspects. To examine changes in HSPG-Net outputs during training, line drawings generated at epochs 50, 100, 150, and 200 were selected for visual comparison. Figure 2 shows that at epoch 50, the main lines in some samples were not yet fully formed and considerable spurious responses remained. As training progressed, some figure contours and local lines became clearer. Differences in line strength and detail are still visible between epochs 150 and 200, and the illustrated examples alone are insufficient to establish model convergence. The model trained for 200 epochs was used for subsequent testing and comparison.

Images of mural figures from Yongle Palace, the Mogao Caves at Dunhuang, and Shalu Monastery were selected as test samples for comparing model outputs. Local artificial damage was introduced into the images, while the corresponding manually traced line drawings before masking were retained as references. HSPG-Net, Pix2PixHD, LineDistiller, U-Net, and CycleGAN were compared using the same test images, damaged regions, and reference line drawings. Figure 3 presents the line drawing outputs of the models and enlarged views of local regions. The models differ in contour clarity, fine-line retention, and spurious lines. In the illustrated examples, some figure contours and decorative lines generated by HSPG-Net are clearer. Nevertheless, the outputs require region-by-region comparison with the manually traced references; restoration accuracy cannot be judged solely from greater line clarity or a larger number of lines. As shown in Table 1, HSPG-Net, developed in this study, achieved higher values than the comparison models on all five quantitative metrics, indicating favorable quantitative results. However, examination of individual images shows that fine-line omissions, incomplete local connections, and positional deviations remain in some complex regions or areas with limited information in the source images. The restoration of local line structures and fine details therefore requires further improvement.

The experimental results indicate that HSPG-Net, trained on paired data with manually traced line drawings, can preserve the overall forms of mural figures and the principal relationships among lines with relative stability while restoring broken and missing lines in locally damaged regions. Under the same test conditions, HSPG-Net achieved higher values than the comparison models on all quantitative metrics. This trend is broadly consistent with the visual comparisons and suggests closer agreement with the manually traced references in terms of spatial correspondence and image structure. The main contribution of this study is the construction of 739 pairs of mural images and manually traced line drawings representing different historical periods and regions, together with an AI-based experimental workflow for the digital restoration of mural figure line drawings. These provide data and methodological support for learning the line characteristics of mural figures and completing missing lines, while also supporting the digital preservation of relevant image and line drawing information and subsequent research. Despite relatively stable overall line drawing restoration, fine-line omissions, incomplete local connections, and positional deviations remain in some regions with dense lines, complex structures, or limited information in the original images. Complex local line structures and fine details therefore remain priorities for further model improvement. The dataset also has scope for expansion in sample size, geographical coverage, and painting types. Future research will add mural samples from different periods and regions with varied pictorial characteristics and optimize the model’s ability to identify and restore complex local structures and fine lines, thereby improving its stability and applicability to the digital restoration of line drawings in different types of historic Chinese murals. ·This research was supported by BK21 Four Service Design based Glocal Social Innovation Educational Research Team in Dongseo University.