作者：Zhanghan Ke Yuhao Liu Lei Zhu Nanxuan Zhao Rynson W. H. Lau
在本文中，我们提出了一种神经预设技术来解决现有颜色风格转换方法的局限性，包括视觉伪影、巨大的内存需求和缓慢的风格切换速度。我们的方法基于两个核心设计。首先，我们提出了确定性神经颜色映射（DNCM），通过图像自适应颜色映射矩阵对每个像素进行一致操作，避免了伪影，并支持小内存占用的高分辨率输入。其次，我们开发了一个两阶段的流水线，将任务分为颜色规范化和风格化，通过提取颜色样式作为预设并在规范化的输入图像上重新使用它们来实现有效的样式切换。由于成对数据集的不可用性，我们描述了如何通过自我监督策略训练神经预设。通过综合评估，展示了神经预设相对于现有方法的各种优势。此外，我们还表明，我们训练的模型可以自然地支持多个应用程序
In this paper, we present a Neural Preset technique to address thelimitations of existing color style transfer methods, including visualartifacts, vast memory requirement, and slow style switching speed. Our methodis based on two core designs. First, we propose Deterministic Neural ColorMapping (DNCM) to consistently operate on each pixel via an image-adaptivecolor mapping matrix, avoiding artifacts and supporting high-resolution inputswith a small memory footprint. Second, we develop a two-stage pipeline bydividing the task into color normalization and stylization, which allowsefficient style switching by extracting color styles as presets and reusingthem on normalized input images. Due to the unavailability of pairwisedatasets, we describe how to train Neural Preset via a self-supervisedstrategy. Various advantages of Neural Preset over existing methods aredemonstrated through comprehensive evaluations. Besides, we show that ourtrained model can naturally support multiple applications without fine-tuning,including low-light image enhancement, underwater image correction, imagedehazing, and image harmonization.
论文链接：http://arxiv.org/pdf/2303.13511v1
更多计算机论文：http://cspaper.cn/