Inpainting of Irregular Holes in a Manuscript using UNet and Partial Convolution

Abstract

Image inpainting of irregular holes in the old handwritten manuscripts is a challenging task. It requires finding the missing part in the manuscript and then filling it with an appropriate pixel. It needs to fill it in such a way that words or letters in that manuscript can be recovered. There are many degraded manuscripts available in museums and libraries all across the globe. This degradation can be due to human mishandling or certain environmental factors (fire, heat, water, microorganisms, pests, and other vermin). Even if it is tired very hard to protect our written heritage from all this, since it is written on natural materials (wood, paper), it is prone to inherent vice. A novel method is presented to inpaint the irregular holes in manuscripts with appropriate pixels. Deep learning, which tends to solve many complex problems like image classification, object detection, image segmentation, helps in achieving image inpainting as well. UNet is proposed to generate the mask of an irregular hole part in a manuscript image and partial convolution, which takes two input masks and the image for inpainting.