人工智能培训

搜索

机器学习论文:AUTO-DEEPLAB:分层神经架构搜索语义图像分割(Auto-DeepLab: Hierarchical Neural Architectur

[复制链接]
liwei906666 发表于 2019-1-11 10:27:19 | 显示全部楼层 |阅读模式
liwei906666 2019-1-11 10:27:19 449 0 显示全部楼层
机器学习论文:AUTO-DEEPLAB:分层神经架构搜索语义图像分割(Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image  Segmentation)最近,神经网络搜索(NAS)已经成功地识别出在大规模图像分类问题上超出人类设计的神经网络架构。在本文中,我们研究NAS用于语义图像分割,这是一个重要的计算机视觉任务,它为图像中的每个像素分配一个语义标签。现有的工作通常集中在搜索可重复的单元结构,同时手工设计控制空间分辨率变化的外部网络结构。这种选择简化了搜索空间,但是对于密集图像预测而言,这种选择变得越来越困难,因为它们需要更多的网络级架构变体。因此,我们建议除了单元级结构之外还搜索网络级结构,这构成了分层体系结构搜索空间。我们提供包含许多流行设计的网络级搜索空间,并开发一种允许有效的基于梯度的架构搜索的配置(在Cityscapes图像上为3 P100GPU天)。我们证明了所提出的方法在具有挑战性的Cityscapes,PASCAL VOC 2012和ADE20Kdatasets上的有效性。在没有任何ImageNet预训练的情况下,我们的架构专门针对语义图像分割进行了搜索,从而获得了最先进的性能。
Recently, Neural Architecture Search (NAS) has successfully identified neuralnetwork architectures that exceed human designed ones on large-scale imageclassification problems.In this paper, we study NAS for semantic imagesegmentation, an important computer vision task that assigns a semantic labelto every pixel in an image.Existing works often focus on searching therepeatable cell structure, while hand-designing the outer network structurethat controls the spatial resolution changes.This choice simplifies the searchspace, but becomes increasingly problematic for dense image prediction whichexhibits a lot more network level architectural variations.Therefore, wepropose to search the network level structure in addition to the cell levelstructure, which forms a hierarchical architecture search space.We present anetwork level search space that includes many popular designs, and develop aformulation that allows efficient gradient-based architecture search (3 P100GPU days on Cityscapes images).We demonstrate the effectiveness of theproposed method on the challenging Cityscapes, PASCAL VOC 2012, and ADE20Kdatasets.Without any ImageNet pretraining, our architecture searchedspecifically for semantic image segmentation attains state-of-the-artperformance.机器学习论文:AUTO-DEEPLAB:分层神经架构搜索语义图像分割(Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image  Segmentation) ZdsiDICCojZS4hmD.jpg
URL地址:https://arxiv.org/abs/1901.02985     ----pdf下载地址:https://arxiv.org/pdf/1901.02985    ----机器学习论文:AUTO-DEEPLAB:分层神经架构搜索语义图像分割(Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image  Segmentation)
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

liwei906666当前离线
新手上路

查看:449 | 回复:0

快速回复 返回顶部 返回列表