Assessment of deep learning algorithms for damage segmentation in Indian murals
Anshul Kumar Yadav, Ronit Kunkolienker, N.A. Dhiraj · International Journal of Arts and Technology · 2025
Murals on walls of the havelis in Rajasthan have become damaged, and image inpainting has emerged as a potent solution to the problem. Masking damage is a crucial step for inpainting algorithms to avoid learning from damaged regions. Therefore, this study focuses on damage identification, utilising prominent architectures, including U-Net and its derivatives, as well as adversarial networks. The study also explores the effect of dense conditional random field (dCRF) and major voting ensemble algorithm. The results show that using dCRF modelling and an ensemble approach improves the average structural similarity index measure (SSIM) score from 0.9672 to 0.9703 for the test dataset. Alone, dCRF improves the output of the worst-performing model, Pix2Pix (modified), by 5.02%. The suggested method also outperforms the generic adversarial image translation networks for mural damage segmentation on the test dataset by up to 5.45% in the mIoU and 3.61% in the mean DSC score.