arXiv reaDer
PHNet: Patch-based Normalization for Portrait Harmonization
A common problem for composite images is the incompatibility of their foreground and background components. Image harmonization aims to solve this problem, making the whole image look more authentic and coherent. Most existing solutions predict lookup tables (LUTs) or reconstruct images, utilizing various attributes of composite images. Recent approaches have primarily focused on employing global transformations like normalization and color curve rendering to achieve visual consistency, and they often overlook the importance of local visual coherence. We present a patch-based harmonization network consisting of novel Patch-based normalization (PN) blocks and a feature extractor based on statistical color transfer. Extensive experiments demonstrate the network's high generalization capability for different domains. Our network achieves state-of-the-art results on the iHarmony4 dataset. Also, we created a new human portrait harmonization dataset based on FFHQ and checked the proposed method to show the generalization ability by achieving the best metrics on it. The benchmark experiments confirm that the suggested patch-based normalization block and feature extractor effectively improve the network's capability to harmonize portraits. Our code and model baselines are publicly available.
updated: Mon Sep 30 2024 08:24:09 GMT+0000 (UTC)
published: Tue Feb 27 2024 14:59:48 GMT+0000 (UTC)
参考文献 (このサイトで利用可能なもの) / References (only if available on this site)
被参照文献 (このサイトで利用可能なものを新しい順に) / Citations (only if available on this site, in order of most recent)
Amazon.co.jpアソシエイト