人工智能培训

搜索

人工智能论文:用于非均匀单图像去模糊的多尺度深度神经网络中的学习内核缩减(Down-Scaling with Learned Kernels in Multi-

[复制链接]
犀牛 发表于 2019-3-26 11:58:28 | 显示全部楼层 |阅读模式
犀牛 2019-3-26 11:58:28 607 0 显示全部楼层
人工智能论文:用于非均匀单图像去模糊的多尺度深度神经网络中的学习内核缩减(Down-Scaling with Learned Kernels in Multi-Scale Deep Neural Networks  for Non-Uniform Single Image Deblurring)多尺度方法已被用于盲图像/视频去模糊问题,以产生用于传统和最近基于深度学习的最先进方法的优异性能。双立方下采样是多尺度方法的非典型选择,可以在使用固定内核进行滤波后减小空间维数。然而,这个固定的内核可能是次优的,因为它可能破坏重要的信息,以便可靠地去除这种强大的边缘。我们提出了基于卷积神经网络(CNN)的下尺度方法,用于基于多尺度深度学习的非均匀单图像去模糊。我们认为,基于CNN的缩小有效地减少了原始图像的空间维度,而具有多通道的学习内核可以很好地保留去模糊任务的必要细节。对于每个规模,我们采用RCAN(残留信道注意网络)作为骨干网络,以进一步提高性能。我们提出的方法大大提高了GoPro数据集的最新性能。我们提出的方法能够比道先生最先进的方法获得高出2.59dB的PSNR。我们提出的基于CNN的缩减是这种优异性能的关键因素,因为我们的网络性能没有降低1.98dB。在大型Su数据集上也评估了使用GoPro集训练的相同网络,并且我们提出的方法比Tao的方法产生了1.15dB的PSNR。 Lai数据集的定性比较还证实了我们提出的方法优于其他方法的优越性能。
Multi-scale approach has been used for blind image / video deblurringproblems to yield excellent performance for both conventional and recentdeep-learning-based state-of-the-art methods.Bicubic down-sampling is atypical choice for multi-scale approach to reduce spatial dimension afterfiltering with a fixed kernel.However, this fixed kernel may be sub-optimalsince it may destroy important information for reliable deblurring such asstrong edges.We propose convolutional neural network (CNN)-based down-scalemethods for multi-scale deep-learning-based non-uniform single imagedeblurring.We argue that our CNN-based down-scaling effectively reduces thespatial dimension of the original image, while learned kernels with multiplechannels may well-preserve necessary details for deblurring tasks.For eachscale, we adopt to use RCAN (Residual Channel Attention Networks) as a backbonenetwork to further improve performance.Our proposed method yieldedstate-of-the-art performance on GoPro dataset by large margin.Our proposedmethod was able to achieve 2.59dB higher PSNR than the current state-of-the-artmethod by Tao.Our proposed CNN-based down-scaling was the key factor for thisexcellent performance since the performance of our network without it wasdecreased by 1.98dB.The same networks trained with GoPro set were alsoevaluated on large-scale Su dataset and our proposed method yielded 1.15dBbetter PSNR than the Tao's method.Qualitative comparisons on Lai dataset alsoconfirmed the superior performance of our proposed method over otherstate-of-the-art methods.人工智能论文:用于非均匀单图像去模糊的多尺度深度神经网络中的学习内核缩减(Down-Scaling with Learned Kernels in Multi-Scale Deep Neural Networks  for Non-Uniform Single Image Deblurring) wSjPaO0TpjXlK523.jpg
URL地址:https://arxiv.org/abs/1903.10157     ----pdf下载地址:https://arxiv.org/pdf/1903.10157    ----人工智能论文:用于非均匀单图像去模糊的多尺度深度神经网络中的学习内核缩减(Down-Scaling with Learned Kernels in Multi-Scale Deep Neural Networks  for Non-Uniform Single Image Deblurring)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

犀牛当前离线
新手上路

查看:607 | 回复:0

快速回复 返回顶部 返回列表