House Number Image Inpainting is a project that seeks to bridge the gap between substandard images and opticalcharacter recognition, or OCR. Currently, OCR works well when the characters in the images are explicit, but itsperformance diminishes rapidly in situations where the character may be blurred, obscured, or ruined in some manner.Our project seeks to use machine learning to repair the images through a technique called image inpainting. Imageinpainting works by filling in the damaged sections of images using information from the rest of the image. Anautoencoder is well suited for this application because it can learn the most important features of the images bycompressing an image into an embedding, reconstructing the image, and comparing it to its original form.
See our paper here.