Deep Learning-Based Inpainting for Reconstructing Severely Damaged Handwritten Javanese Characters

Long-term storage at museums can damage ancient Javanese manuscripts; for instance, temperature changes and other factors may cause parts of the script to disappear. The Javanese script exhibits similarities among its letters, making the reconstruction process challenging, particularly when dealing with severe damage to the script's characteristic areas. To address this issue, we conducted a character painting technique that utilizes deep learning architecture, specifically the convolutional autoencoder, partial convolutional neural network, UNet, and ResUNet. The dataset contains 12,000 handwritten Javanese characters. We evaluated the restoration of missing characters using SSIM and PSNR metrics. The ResUNet achieves the best performance compared to other methods, with an SSIM value of 0.9319 and a PSNR value of 18.9507 dB. According to this study, the ResUNet models can reconstruct Javanese manuscripts with strong performance, offering an alternative solution to ensure the preservation and accessibility of these valuable historical documents for future generations.