人工智能培训

搜索

深度学习论文:通过实时样式转移语义网络净化自然图像(Purifying Naturalistic Images through a Real-time Styl

[复制链接]
tetra 发表于 2019-3-15 12:26:15 | 显示全部楼层 |阅读模式
tetra 2019-3-15 12:26:15 856 0 显示全部楼层
深度学习论文:通过实时样式转移语义网络净化自然图像(Purifying Naturalistic Images through a Real-time Style Transfer  Semantics Network)最近,综合学习的进展提出了合成图像的训练模型,可以有效地降低人力和物质资源的成本。然而,由于与真实图像相比合成图像的不同分布,所期望的性能仍然不能实现。真实图像由多种形式的光取向组成,而合成图像由均匀的光取向组成。这些特征分别被认为是室外和室内场景的特征。解决这个问题,前一种方法学会了一种模型来改善合成图像的真实感。与以往的方法不同,本文采用净化真实图像的第一步。通过风格转移任务,将室外真实图像的分布转换为室内合成图像,从而减少光的影响。因此,本文提出了一种区域时间风格转移网络,其保留了输入图像(真实图像)的图像内容信息(例如,注视方向,瞳孔中心位置),同时推断了风格图像的风格信息(例如,图像颜色结构,语义特征)。合成图像)。此外,网络加速了模型的收敛速度,并适应多尺度图像。使用混合研究(定性和定量)方法进行实验,以证明在复杂方向上纯化真实图像的可能性。定性地,将所提出的方法与LPW数据集的一系列室内和室外场景中的可用方法进行比较。在定量计中,它通过在交叉数据集上训练凝视估计模型来评估纯化图像。结果显示,与原始实际图像相比,基线方法有显着改进。
Recently, the progress of learning-by-synthesis has proposed a training modelfor synthetic images, which can effectively reduce the cost of human andmaterial resources.However, due to the different distribution of syntheticimages compared to real images, the desired performance cannot still beachieved.Real images consist of multiple forms of light orientation, whilesynthetic images consist of a uniform light orientation.These features areconsidered to be characteristic of outdoor and indoor scenes, respectively.Tosolve this problem, the previous method learned a model to improve the realismof the synthetic image.Different from the previous methods, this paper takesthe first step to purify real images.Through the style transfer task, thedistribution of outdoor real images is converted into indoor synthetic images,thereby reducing the influence of light.Therefore, this paper proposes areal-time style transfer network that preserves image content information (eg,gaze direction, pupil center position) of an input image (real image) whileinferring style information (eg, image color structure, semantic features) ofstyle image (synthetic image).In addition, the network accelerates theconvergence speed of the model and adapts to multi-scale images.Experimentswere performed using mixed studies (qualitative and quantitative) methods todemonstrate the possibility of purifying real images in complex directions.Qualitatively, it compares the proposed method with the available methods in aseries of indoor and outdoor scenarios of the LPW dataset.In quantitativeterms, it evaluates the purified image by training a gaze estimation model onthe cross data set.The results show a significant improvement over thebaseline method compared to the raw real image.深度学习论文:通过实时样式转移语义网络净化自然图像(Purifying Naturalistic Images through a Real-time Style Transfer  Semantics Network) anmNygN6MMURYWdX.jpg
URL地址:https://arxiv.org/abs/1903.05820     ----pdf下载地址:https://arxiv.org/pdf/1903.05820    ----深度学习论文:通过实时样式转移语义网络净化自然图像(Purifying Naturalistic Images through a Real-time Style Transfer  Semantics Network)
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

tetra当前离线
新手上路

查看:856 | 回复:0

快速回复 返回顶部 返回列表