Contextual Residual Aggregation for Ultra High-Resolution Image Inpainting

Abstract

Fig. 1: Inpainting results on ultra high-resolution images.

Automatic Object Removal With Obstructed Façades Completion Using Semantic Segmentation and Generative Adversarial Inpainting

“…Previous image inpainting methods, such as exemplar-based inpainting, can find the approximate nearest neighbor to match image patches [16]. With the rise of learning-based image inpainting methods, the inpainting results have become more realistic and reasonable [24].…”

“…• Combining the Two: The third class of approaches attempted to combine the two to overcome the limitations of replication methods or modeling methods. Not only do these methods learned to build image distributions in a data-driven manner, but they were also designed to explicitly borrow patches or features from background regions [16]. However, when the training dataset and the content of the processed images do not match, the generated image quality is not satisfactory.…”

“…Visual authenticity and semantic correctness are essential indicators for evaluating the quality of synthetic images. Recent encouraging advances in data-driven image inpainting methods, which are more effective than classical methods in handling object removal from complex scenes and large occlusion rate images, have attracted the interest of researchers [16].…”

Grave Artifacts in Image Inpainting: Investigating the Causes and Untangling the Factors

“…At each level of the decoder, these attention maps are concatenated with the decoder features coming from the encoder path. Finally, in [34], the authors adopt most of the techniques described above, including a coarse-to-fine architecture, attention and gated convolutions, with an approach aimed at performing inpainting on images of arbitrary size. This is achieved by downsampling the image and performing inpainting, and subsequently upscaling just the inpainted region to the original resolution, and pasting it on top of the original image.…”

Synthesizing photorealistic images with deep generative learning

“…It can be used for restoring damaged paintings, removing unwanted objects, and generating new content for incomplete scenes. Many approaches have been proposed for this non-trivial task, including diffusion-based methods [8,13,14,113], patch-based methods [10,32,33,89]) and learning-based methods [86,123,139,146,208,214]. While these approaches rapidly improve the completion results, they produce only one "optimal" result for a given masked image and do not have the capacity to generate a variety of semantically meaningful results.…”