DL之DeconvNet：DeconvNet算法的简介(论文介绍)、架构详解、案例应用等配图集合之详细攻略

DL之DeconvNet：DeconvNet算法的简介(论文介绍)、架构详解、案例应用等配图集合之详细攻略

 

 

 

 

目录

DeconvNet算法的简介(论文介绍)

0、实验结果

DeconvNet算法的架构详解

DeconvNet算法的案例应用

---

 

 

 

 

 

 

 

相关文章
DL之DeconvNet：DeconvNet算法的简介(论文介绍)、架构详解、案例应用等配图集合之详细攻略
DL之DeconvNet：DeconvNet算法的架构详解

 

DeconvNet算法的简介(论文介绍)

        DeconvNet网络架构，是由Convolution network、Deconvolution network两种架构组成。

• Convolution network：feature extractor，采用VGG-16提取特征；
• Deconvolution network：shape generator，通过上采样，计算像素的类别得分图。
   

Abstract  
       We propose a novel semantic segmentation algorithm by  learning a deconvolution network. We learn the network  on top of the convolutional layers adopted from VGG 16-  layer net. The deconvolution network is composed of deconvolution  and unpooling layers, which identify pixel-wise  class labels and predict segmentation masks. We apply the  trained network to each proposal in an input image, and  construct the final semantic segmentation map by combining  the results from all proposals in a simple manner. The  proposed algorithm mitigates the limitations of the existing  methods based on fully convolutional networks by integrating  deep deconvolution network and proposal-wise  prediction; our segmentation method typically identifies detailed  structures and handles objects in multiple scales naturally.  Our network demonstrates outstanding performance  in PASCAL VOC 2012 dataset, and we achieve the best accuracy  (72.5%) among the methods trained with no external  data through ensemble with the fully convolutional network.
       本文提出了一种新的基于反卷积网络的语义分割算法。我们学习了VGG 16层网在卷积层之上的网络。反卷积网络由反褶积层和反池层组成，它们识别像素级标签并预测分割掩码。我们将训练好的网络应用于输入图像中的每个提案，并将所有提案的结果以一种简单的方式结合起来，构造出最终的语义分割图。该算法将深度反卷积网络与建议预测相结合，克服了现有全卷积网络方法的局限性;我们的分割方法通常识别详细的结构和处理对象在多个尺度自然。我们的网络在PASCAL VOC 2012数据集中表现出色，通过全卷积网络集成，在没有外部数据训练的方法中，我们的准确率最高(72.5%)。
Conclusion  
       We proposed a novel semantic segmentation algorithm  by learning a deconvolution network. The proposed deconvolution  network is suitable to generate dense and pre-cise object segmentation masks since coarse-to-fine structures  of an object is reconstructed progressively through  a sequence of deconvolution operations. Our algorithm  based on instance-wise prediction is advantageous to handle  object scale variations by eliminating the limitation  of fixed-size receptive field in the fully convolutional network.  We further proposed an ensemble approach, which  combines the outputs of the proposed algorithm and FCNbased  method, and achieved substantially better performance  thanks to complementary characteristics of both algorithms.  Our network demonstrated the state-of-the-art  performance in PASCAL VOC 2012 segmentation benchmark  among the methods trained with no external data.
       本文提出了一种新的基于反卷积网络的语义分割算法。该反褶积网络通过一系列的反卷积操作，逐步重构出由粗到细的目标结构，适用于生成密集的预分割掩码。我们的基于实例预测的算法消除了全卷积网络中固定大小接受域的限制，有利于处理对象尺度变化。我们进一步提出了一种集成方法，将所提算法的输出与基于FCN的方法相结合，由于两种算法的互补特性，取得了较好的性能。在没有外部数据训练的方法中，我们的网络在PASCAL VOC 2012分割基准测试中展示了最先进的性能。

 

论文
Hyeonwoo Noh, SeunghoonHong, BohyungHan.
Learning deconvolution network for semantic segmentation, ICLR, 2015.
https://arxiv.org/abs/1505.04366

 

0、实验结果

1、PASCAL VOC 2012验证图像的语义分割结果实例
Example of semantic segmentation results on PASCAL VOC 2012 validation images
PASCAL VOC 2012上获得的mean IoU=72.5%

(a) Examples that our method produces better results than FCN	GT框是人工标定框、FCN算法、DeconvNet算法、EDeconvNet算法(FCN和EDeconvNet集成学习后)、EDeconvNet+CRF后处理的算法(效果更好！)
(b) Examples that FCN produces better results than our method 所提出的方法和FCN具有用于语义分割的互补特性，并且两种方法的组合通过集成学习提高了准确性。	下图展示了FCN算法优于DeconvNet算法(因为多出了一些细节)，同样的，EDeconvNet+CRF后处理的算法(效果更好！)
集成学习 (c) Examples that inaccurate predictions from our method and FCN are improved by ensemble

 

 

DeconvNet算法的架构详解

更新……

DL之DeconvNet：DeconvNet算法的架构详解

 

 

 

 

 

DeconvNet算法的案例应用

更新……