Semantic Inpainting of Images using Deep Learning

Computer Vision enables computers to retrieve information from digital images and use the inferred data to perform the required task. Image inpainting, a computer vision technique, helps to reconstruct damaged images by refilling the missing pixels, called holes, using the relevant and known pixels, so that repaired image looks very natural and realistic. Traditional inpainting methods generally fill the holes by matching the most similar pixels in the surrounding known regions, focusing to reconstruct the exact ground truth image, leaving behind the texture and quality. Currently, many deep learning methods produced drastic improvements in visual quality and texture and also look for the semantic context of the image. However, achieving success on high resolution images with complex structures remains challenging. This paper imparts an intensive vision of the existing Inpainting methods by providing a comprehensive description of the methods used, datasets and evaluation metrics for all the analyzed techniques.