Super-attention for exemplar-based image colorization
Abstract
In image colorization, exemplar-based methods use a reference color image to guide the colorization of a target grayscale image. In this article, we present a deep learning framework for exemplar-based image colorization which relies on attention layers to capture robust correspondences between high-resolution deep features from pairs of images. To avoid the quadratic scaling problem from classic attention, we rely on a novel attention block computed from superpixel features, which we call super-attention. Super-attention blocks can learn to transfer semantically related color characteristics from a reference image at different scales of a deep network. Our experimental validations highlight the interest of this approach for exemplar-based colorization. We obtain promising results, achieving visually appealing colorization and outperforming state-of-theart methods on different quantitative metrics.
Fichier principal
Super-attention_for_exemplar-based_image_colorization.pdf (22.43 Mo)
Télécharger le fichier
Super_attention_for_exemplar_based_image_colorization_supplementary_materials.pdf (37.23 Mo)
Télécharger le fichier
Origin : Files produced by the author(s)