ICCV_2009_What is the best multi-stage architecture for object recognition?
个人认为这是一篇很好的分析型文章,思路浅显易懂,实验验证也很充分。
本文主要在于研究三个问题:1. 滤波器后面的非线性对识别率有什么影响?
2. 有监督学习与无监督学习,以及hard-wired与random的filter有什么影响?
3. 两层结构是否优于单层结构?
本文用PSD无监督学习方法,对几个因素进行组合比较,主要有一下几种组合:
卷积层filter bank layer:
![](http://image.sciencenet.cn/album/201407/24/050656bpbtw6gpebzqbbe4.png)
修正层rectification layer:
![](http://image.sciencenet.cn/album/201407/24/050656ufp353t6zut66mzz.png)
局部对比度归一层local contrast normalization layer:
![](http://image.sciencenet.cn/album/201407/24/050656p6cvz83hqq3qclpl.png)
最大值池化或均值池化max-pooling or subsampling layer:
![](http://image.sciencenet.cn/album/201407/24/0506570xxxsoixtzslre6l.png)
训练方法:
Random Features and Supervised Classifier: R and RR
Unsupervised Features, Supervised Classifier: U and UU
Random Features, Global Supervised Refinement: R+ and R+R+
Unsupervised Feature, Global Supervised Refinement: U+ and U+U+
实验数据:Caltech-101
实验对比结果图:
On NORB dataset:
https://blog.sciencenet.cn/blog-1583812-814174.html
上一篇:
2010_Tutorial_Feature Learning for Image Classification笔记下一篇:
An analysis of single-layer network in unsupervised...