An AI-Assisted Framework for Digital Restoration of Ajanta Murals to Support Inclusive Learning and Cultural Heritage Education

Abstract

Ajanta murals may be considered as one of the most precious cultural heritage artifacts in India. However, due to the factors such as water intrusion, salt efflorescence, and abrasion of surface, many of these paintings have undergone decay in time. Not only does this pose a threat to the preservation of this cultural heritage artifact but also restricts the use of these paintings as rich visual sources in learning about cultural heritage. Traditional digital restoration techniques of these mural paintings using pixel-based inpainting often lack in the representation of the distinctive artistic features of these paintings in terms of texture and coloration.

This paper proposes an AI-driven method of restoring Ajanta mural paintings using Neural Style Transfer (NST). The technique involves transferring the style from a portion of a mural painting to a degraded part of it while retaining the structural information present in the image. In order to stabilize the process of restoration and increase its visual quality, three techniques are employed in addition to NST – Adaptive Gram Matrix Normalization (AGMN), Multi-scale Feature Fusion (MFF), and Dynamic Loss Scheduler (DLS).

AGMN-MFF architecture performs on the benchmark test set which consists of 120 image pairs of content and styles with SSIM value of 0.81, which is 30.6% better than the classical Neural Style Transfer method and minimizes optimization issues by 87%. The framework was tested through restoration of artificially degraded images of Ajanta murals which have been sourced from the IGNCA digital archive database. Canny edge maps are used in order to preserve structural information of the degraded images while mural parts are used for styling purposes. Evaluating stylistic and perceptual aspects, it has been found out that 73% of the images restored with the proposed framework received rating of stylistically convincing against 33% images in baseline method (p=0.003, Wilcoxon signed-rank; κ=0.61).

Thus, the suggested framework shows how artificial intelligence can be used for the digital preservation of heritage and generation of visually coherent restorations of damaged murals. Not only will this framework help researchers in their work in heritage preservation, but it may also be used for educational purposes in order to allow learners, including learners from inclusive education setting, to better interact with India's cultural heritage