语义分割--DeconvNet--Learning Deconvolution Network for Semantic Segmentation

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://cloud.tencent.com/developer/article/1437716

Learning Deconvolution Network for Semantic Segmentation

ICCV2015

http://cvlab.postech.ac.kr/research/deconvnet/

https://devhub.io/zh/repos/myungsub-DeconvNet

本文提出了一个 deconvolution network 用于语义分割，还是针对 FCN 改善的。在pooling operation 记录 the locations of maximum activations ，将这些位置用于 unpooling

这一点和 SegNet 是一样的。

首先来说说FCN 有什么问题？

FCN 的 limitations：

1）FCN 因为其固定尺寸receptive field只能解决单尺度的semantics ，对于过大过小的目标分割都有可能有问题

the network can handle only a single scale semantics within image due to the fixed-size receptive field. Therefore, the object that is substantially larger or smaller than the receptive field may be fragmented or mislabeled.

2）FCN 的 deconvolution procedure 太粗糙太简单，FCN 的 deconvolution procedure输入尺寸只有16 × 16，将这个尺寸通过 bilinear interpolation 放大到输入图像尺寸。目标很多细节信息丢失

3 System Architecture

我们的网络包括两个部分：convolution and deconvolution networks

卷积网络用于提取特征，使用 VGG 16-layer net

convolution network corresponds to feature extractor

反卷积网络根据特征产生分割结果

deconvolution network is a shape generator that produces object segmentation from the feature extracted from the convolution network.

3.2. Deconvolution Network for Segmentation

反卷积网络中主要有两个操作步骤： unpooling and deconvolution

3.2.1 Unpooling

Pooling 会产生降采样的效果，导致一些细节丢失。 spatial information within a receptive field is lost during pooling，这些信息对于语义分割很重要。为了解决这个问题，我们在反卷积网络中使用 unpooling layers， It

records the locations of maximum activations selected during pooling operation in switch variables, which are employed to place each activation back to its original pooled location.

3.2.2 Deconvolution

unpooling layers 得到一个放大的但是稀疏的响应特征图， 这里通过deconvolution layers 来将稀疏的特征变为稠密的特征

The deconvolution layers densify the sparse activations obtained by unpooling through convolution-like operations with multiple learned filters.

The learned filters in deconvolutional layers correspond to bases to reconstruct shape of an input object.

Unpooling captures example-specific structures

learned filters in deconvolutional layers tend to capture class-specific shapes

与FCN 效果对比：

PASCAL VOC 2012 test set

instance-wise prediction 的好处

本文算法较FCN的细节要好些

FCN 的全局观要好一些

与FCN的结合