首页|Multi-level augmented inpainting network using spatial similarity
Multi-level augmented inpainting network using spatial similarity
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
Recently, multi-scale neural networks have shown promising improvements in image inpainting. However, most of them adopt the progressive way, in which the errors on lower scales may be propagated on higher scales. Addressing this issue, we propose a multi-level augmented inpainting network (MLA Net) to rationally harmonize the inter-and intra-level contexts. Here, a pyramid reconstruction structure (PRS) with three parallel levels is designed to establish the inter-level relationship, which can boost the representation of the features by integrating the texture details into semantics. Then, we propose a novel spatial similarity based attention mechanism (SSA) to ensure the intra-level local continuity between the holes and related available patches. In SSA, in order to focus on the important textures and structures rather than calculating each pixel of the feature equally, a spatial map is utilized to highlight the corresponding spatial locations during the similarity computation. The experiments are evaluated on multiple challenging datasets, which demonstrate that MLA-Net can generate accurate results with better visual quality compared with the state-of-the-art methods. For the 256 x 256 Places2 dataset, PSNR increases 1.02 dB, while FID decreases 0.075. For the 256 x 256 CelebA-HQ dataset, there are 0.22 dB and 0.613 improvements in PSNR and FID. (c) 2022 Elsevier Ltd. All rights reserved.
NSTL
Elsevier
