brakuta / Summary-and-Practice-of-Semantic-Segmentation

This is a survey and practice of semantic segmentation network. Taking the satellite image segmentation as the research task, several classic neural networks based on deep learning are tested.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Summary-and-Practice-of-Semantic-Segmentation

This is a survey and practice of semantic segmentation network. Taking the satellite image segmentation as the research task, several classic neural networks based on deep learning are tested.

调研了目前较为常用的语义分割网络,并在一套卫星图像语义分割数据集上做训练和测试。代码基于 Keras 框架编写,网络结构根据实际情况做了一些调整。

数据集下载地址:https://pan.baidu.com/s/1i6oMukH (yqj2)

预训练模型:目前预训练模型暂不公开。

SegNet

最为基础的语义分割网络之一,结构简单。

Sample

UNet

同样是较为经典的语义分割网络,结构简单。

结构方面的细节推荐查看:https://www.cnblogs.com/fanhaha/p/7242758.html

Sample

FCN

全卷积神经网络是比较经典的分割网络。根据转置卷积上采样倍数的区别,FCN family 包括 FCN-8,FCN-16, 和 FCN-32。其中 FCN-8 的效果最佳。

Sample

ENet

设计了非对称的编码解码模块,旨在减少参数和计算量。

使用了可分离卷积、空洞卷积、跳跃连接等结构。

Sample

DeepLab_v3

参考:https://www.jianshu.com/p/755b001bfe38

Sample

PSPNet

使用了金字塔池化和插值缩放方法。

参考:https://blog.csdn.net/u011974639/article/details/78985130

Sample

ERFNet

核心操作是 residual connections 和 factorized convolutions(空间可分离卷积)

参考:https://blog.csdn.net/baidu_27643275/article/details/98187098

Sample

About

This is a survey and practice of semantic segmentation network. Taking the satellite image segmentation as the research task, several classic neural networks based on deep learning are tested.


Languages

Language:Python 100.0%