SR: Severely Defective Image Inpainting with Structural Reconstruction
Li Yuelong, Peng Wang, Yue Xing, Jianming Wang · 2021
Image inpainting especially inpainting large defects regions in images is a promising yet challenging task. However, existing approaches can suffer from generating distorted structures and blurry textures in large defects images. In this paper, a framework for severe defects of inpainting images has been presented by using image structure features. When a large defect in the image appears, the image structure is already severely damaged. Firstly, aiming at the problem of image structure defects, an iterative repair method based on large defect areas is proposed to guide the repair. Specifically, we integrally use depth encoding and decoding structures to generate models. Owing to deeper feature maps represent the structural features of the images in convolutional neural networks. We used deep features to guide the inpainting and repurposed deep features for the decoding process during the inpainting process. Utilizing feature map as feedback progressively fills the hole by trusting only the high confidence pixels within the hole at each iteration and focuses on the remaining pixels at the next iteration. Meanwhile, an AC block has been proposed to enhance contextual inference. Taking advantage of previously iterated partial as known pixels to predict other pixels, this process gradually improves the results. The evaluations for quantitative and qualitative are superior to existing methods by our experiments. Keywords-large defective image inpainting, deep learning, struct reconstruction