Restauração de imagens baseada em múltiplos quadros

Abstract

The restoration of images with defects is currently carried out predominantly by generative networks for inpainting. In some cases, the image may end up merely replicating patterns from neighboring pixels around the defect, as there is no way to recover themissing information. In the case of images printed on a large scale, such as product packaging, book covers, and discs, where multiple copies of these packages still exist, the restoration of these images can benefit from this larger quantity of copies to combine the information contained in each copy to fill in the gaps caused by image degradation. Based on this principle, solutions from a similar domain that leverages information from multiple images for super-resolution, known as Multi-Frame Super Resolution (MFSR), have been investigated. The aim is to create or adapt a solution that combines information from multiple observations of a widely distributed printed image to restore the original image. Given the MFSR pipeline composed of feature extraction, alignment, fusion, and image reconstruction, a form of image alignment has been proposed without the need to use deep neural networks, as there is no need for such powerful tools to align images of a flat object. A method for merging the aligned images to generate the restored image is still to be proposed. Key-words: Machine Learning. Super Resolution. Computer Vision.