High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis

TL;DR

This paper introduces a multi-scale neural patch synthesis method for high-resolution image inpainting that preserves details and structures, outperforming previous techniques especially on large images.

Contribution

The paper presents a novel multi-scale neural patch synthesis approach that effectively handles high-resolution images for inpainting, overcoming memory and quality limitations of prior methods.

Findings

Achieved state-of-the-art inpainting accuracy on ImageNet and Paris Streetview datasets.

Produced sharper, more coherent inpainted images compared to prior methods.

Effectively preserves high-frequency details and contextual structures in high-resolution images.

Abstract

Recent advances in deep learning have shown exciting promise in filling large holes in natural images with semantically plausible and context aware details, impacting fundamental image manipulation tasks such as object removal. While these learning-based methods are significantly more effective in capturing high-level features than prior techniques, they can only handle very low-resolution inputs due to memory limitations and difficulty in training. Even for slightly larger images, the inpainted regions would appear blurry and unpleasant boundaries become visible. We propose a multi-scale neural patch synthesis approach based on joint optimization of image content and texture constraints, which not only preserves contextual structures but also produces high-frequency details by matching and adapting patches with the most similar mid-layer feature correlations of a deep classification network. We evaluate our method on the ImageNet and Paris Streetview datasets and achieved state-of-the-art inpainting accuracy. We show our approach produces sharper and more coherent results than prior methods, especially for high-resolution images.