A Review on Image Inpainting Techniques and Datasets

• David J. A. Barrientos UPE
• Bruno Fernandes UPE
• Sergio Fernandes UPE

Resumo

Image inpainting is a process that allows filling in target regions with alternative contents by estimating the suitable information from auxiliary data, either from surrounding areas or external sources. Digital image inpainting techniques are classified in traditional techniques and Deep Learning techniques. Traditional techniques are able to produce accurate high-quality results when the missing areas are small, however none of them are able to generate novel objects not found in the source image neither to produce semantically consistent results. Deep Learning techniques have greatly improved the quality on image inpainting delivering promising results by generating semantic hole filling and novel objects not found in the original image. However, there is still a lot of room for improvement, specially on arbitrary image sizes, arbitrary masks, high resolution texture synthesis, reduction of computation resources and reduction of training time. This work classifies and orders chronologically the most prominent techniques, providing an overall explanation on its operation. It presents, as well, the most used datasets and evaluation metrics across all the works reviewed.

Palavras-chave: convolution based, dataset, deep learning, diffusion based, inpainting, patch based, reconstruction