Fig. 6: Qualitative comparison with state-of-the-art methods. | npj Heritage Science

Fig. 6: Qualitative comparison with state-of-the-art methods.

From: Supporting historic mural image inpainting by using coordinate attention aggregated transformations with U-Net-based discriminator

Fig. 6

We have highlighted and enlarged specific details next to each image with a red box. a Masked, (b) PConv, (c) GatedConv, (d) PDGAN, (e) Ours, (f) Ground Truth. Note: This figure compares the inpainting performance of different models for mural restoration. a shows the original image with missing sections due to natural degradation. b through (f) display the inpainting results of various models, with each panel presenting the inpainting of the same damaged region. b illustrates the result from PConv, which often produces unrealistic content in the missing areas, showing a lack of precise inpainting in certain regions. GatedConv (c) struggles with larger, irregular holes, leading to suboptimal inpainting results. While PDGAN (d) generates relatively reasonable inpainting, it consistently produces noticeable artifacts that degrade the visual quality of the restoration. In contrast, (e) shows the inpainting results from our method, which provides a significantly better output. The inpainted regions are more seamless, natural, and contextually consistent with the surrounding details. f compares the output from our method with the original (ground truth), clearly demonstrating that our approach outperforms the others in terms of quality and detail preservation.

Back to article page