Virtually restoring headless Buddha statues using Deep Learning Techniques

Buddha statues are undoubtfully important for culture and history for worldwide Buddhists, in particular to Sri Lanka since it is the major religion in the country, for representing the rich history of Buddhism. Most of historical Buddha statues which belong to earliest eras of the country, such as in Anuradhapura, had been destroyed or degraded due to natural disasters, vandalism, or aging. It is common that body parts such as the head, has been lost or destroyed in these Buddha statues. While various physical restoration methods exist, they tend to be expensive, and often their effects decrease historical or artistic value of the statues. Instead, the virtual restoration methods represent a non-invasive and affordable solution to the preservation and reconstruction of these heritage assets according to current studies. This study is presenting a specially designed deep learning model using U-Net architecture for segmenting the body of seated headless Buddha statues by removing the background in the input images, which trained on a custom dataset of Buddha statues found in Anuradhapura, Sri Lanka. These segmented images are removing all the unnecessary details from the input image and give a clean statue body for the successive tasks of restoration. The head inpainting phase based on this segmented image will be done with a GAN based architecture, an extension of Pix2Pix architecture, which will virtually inpaint a matching head for the given statues. This two-stage approach; segmentation process followed by inpainting process provides a full pipeline for virtual restoration of a matching head for a headless Buddha statue, while maintaining the cultural integrity and reducing physical intervention to a minimum. This would make the proposed methodology quite efficient, hence a very useful tool in heritage conservation and virtual reconstruction of culturally important artifacts.