COLOR IMAGE DEMOSAICKING VIA DEEP RESIDUAL LEARNING
COLOR IMAGE DEMOSAICKING VIA DEEP RESIDUAL LEARNING
复制标题
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
Run-xu Tan;K. Zhang;W. Zuo;Lei Zhang
中科院分区：
文献类型：
--
作者：
Run-xu Tan;K. Zhang;W. Zuo;Lei Zhang
颜色去马赛克在使用滤色器阵列的数字成像中起着关键作用。大多数现有的去马赛克方法都是基于手工制作的先验知识，这在困难的情况下可能会显示出令人不快的视觉伪影(例如，具有高颜色饱和度和尖锐颜色过渡的区域)。为了解决彩色图像的去马赛克问题，提出了一种端到端训练的定制卷积神经网络。具体地，利用残差学习策略学习去马赛克先验，采用两阶段结构：第一阶段旨在恢复G通道的中间结果作为指导先验，第二阶段利用中间G通道信息来指导最终颜色去马赛克结果的重建。我们在广泛使用的Kodak和McMaster数据集和一个新的数据集上的实验结果表明，所提出的CNN模型不仅在数量和质量上都优于现有的去马赛克算法，而且在GPU计算中具有较快的去马赛克速度。
Color demosaicking plays a key role in digital imaging with a color filter array. Most existing demosaicking methods are based on hand-crafted priors, which may exhibit unpleasant visual artifacts in hard cases (e.g., regions with high color saturation and sharp color transition). This paper presents a customized convolutional neural network (CNN), which is trained in an end-to-end manner from natural color images to address the color demosaicking problem. Specifically, by utilizing the residual learning strategy, our network learns the demosaicking prior with a two-stage architecture: the first stage aims to recover an intermediate result of the G channel as guidance prior, while the second stage uses the intermediate G channel information to guide the reconstruction of final color demosaicking result. Our experimental results on the widely-used Kodak and McMaster datasets and a new dataset demonstrate that the proposed CNN model not only yields superior results to state-of-the-art demosaicking algorithms both quantitatively and qualitatively, but also enjoys a fast demosaicking speed by GPU computation.