Text Recovery via Deep CNN-BiLSTM Recognition and Bayesian Inference

Image inpainting is an essential process of semantically filling the missing holes in a corrupt image. However, concurrent methods cannot semantically recover some self-described objects, such as a text instance. In this paper, we focus on the recovery of a missing character in a detected corrupt text instance on an image and propose a procedure to semantically recover the text. Specifically, the corrupt text instance is first recognized with a pre-trained CNN-BiLSTM architecture and the missing character is inferred by the statistical Bayesian posterior probability. The horizontal coordinate of the missing character is estimated by the image histogram. The obtained candidate character and possible position enable the synthesis of the corrupt image and the latent character, and the text information, meanwhile, is semantically recovered. Experiments and corresponding results demonstrate that our procedure is able to predict the missing character and synthesize the images of the corrupt context and the candidate character; besides, it can be employed in the recovery of natural text images.