Handwritten Javanese Character Reconstruction with Extensive Damage Areas Using Deep Learning-Based Inpainting Techniques

Abstract

Ancient Javanese manuscripts stored in several museums have been damaged, so some parts of the script are missing. The requirement for repairing broken Javanese script is maintaining the correct stroke shape and consistent font style. The Javanese script has similarities between the letters, so if there is damage to the area that is characteristic of the script, it will not be easy to reconstruct. In addition, the reconstruction of the Javanese script will also experience difficulties if the damage area is large. Character improvement can be done using character inpainting techniques. Due to the limitations of textual data, more research on character inpainting needs to be done. In this study, the implementation of the character painting technique is proposed to reconstruct the handwritten Javanese script with the condition of a wide area of damage or damage to the area that is characteristic of the script. The proposed method uses Deep Learning architecture, namely Convolutional Autoencoder, Partial Convolutional Neural Network, UNet, and ResUNet. These methods were evaluated qualitatively and quantitatively on a dataset consisting of 12000 handwritten Javanese characters. Restoration of missing character sections is evaluated using SSIM and PSNR metrics. Overall, the ResUNet method achieves the best performance compared to other methods. Because the ResUNet method has the concept of residual learning introduced in ResNet.