Binary Inscription Character Inpainting Based on Improved Context Encoders

Due to the limitations of textual datasets, there are currently few studies on character inpainting. Second, unlike image inpainting, character inpainting cannot replicate the results of image inpainting, and some methods are even inaccessible. Finally, the extraction of character features in the existing model is insufficient, resulting in the model being unable to reconstruct characters based on complete and accurate character features. In this paper, we first threshold the obtained inscription dataset to obtain a visually better-binarized inscription dataset. Second, we improve the Context encoders to design BCEs(Binary Context Encoders) and add dilated convolution to learn the structural features of the character. It has been experimentally proven that BCEs is not only slightly better than methods in the same field, but can also restore real-life inscription characters with missing strokes.