Damage Sensitive and Original Restoration Driven Thanka Mural Inpainting

Colour-Balanced Edge-Guided Digital Inpainting: Applications on Artworks

“…At the same time, Yu et al [22] include no quantitative result, so we are not able to draw any numeric comparison with their method. While our method does not outperform the SSIM and PSNR reported by [17,23], it is noteworthy to emphasize that the proposed approach works selectively better than [17] for certain areas in the image, as presented in the color difference maps of Figure 12. This selective performance is reinforced by the CNN-based metrics that prove our method handles better the processing of higher level combinations of spatial and chromatic features.…”

“…Letting aside the variations in data, mask type, and precise extent of missing pixels reported for each method, we can compare these results by identifying their common purpose: solving the digital inpainting task for Buddhist mural paintings with deep learning approaches. With these considerations in mind, Wang et al, in [ 23 ], simulate four levels of damage as is our case, even though the ratio of corrupted pixels to non-corruped pixels per level remains unknown. Compared to the IQMs reported by [ 23 ], our work achieves higher PSNR extremes, while lower SSIM extremes.…”

“…With these considerations in mind, Wang et al, in [ 23 ], simulate four levels of damage as is our case, even though the ratio of corrupted pixels to non-corruped pixels per level remains unknown. Compared to the IQMs reported by [ 23 ], our work achieves higher PSNR extremes, while lower SSIM extremes. It seems that the user input in the interactive Deep Image Prior method of [ 25 ] leverages the SSIM to a value of 0.78 for random-walk type of damage, which is higher than our results for the RW category (0.74).…”

“…Wang et al [ 23 ] offer a method for the digital restoration of Tibetan Thanka murals, based on Unet Partial Convolution Network. Their main contributions are related to the mask and training loss design.…”

“…A positive value (yellow) for the contour plots show areas where our proposed method is more color coherent with the original, whereas negative values (blue) represent regions where the other algorithm performs better. It is not trivial to draw a straightforward and fair comparison with other relevant state-of-the-art models, such as [3,23,25], where the exact dataset, train-test split, and model implementation are not available to perform a one-to-one mapping to our approach. So much so, we provide a rough comparison in Table 2, where we outline key aspects of every method, including the minima and maxima of PSNR and SSIM metrics where applicable.…”