# CNN

![](https://352802547-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-M7DcNFZmOW8qzXTiAei%2Fsync%2F76b613ab82ead3771cf6c662e7894188ad500f16.jpg?generation=1589383540124854\&alt=media)

> 注意， s2到c3采用了16个 6X5X5 的kernel，生成了16个feature map

## Alexnet

![](https://352802547-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-M7DcNFZmOW8qzXTiAei%2Fsync%2Fdf2782d3c87369535c710156cc28173bbb0aa1a1.png?generation=1589383542021076\&alt=media)

## 卷积

一维卷积：\
$$f(x)\*g(x) = \int\_{-\infty}^{\infty} f(\tau)g(x-\tau) d\tau \tag{1}\label{1}$$

二维卷积：\
![](https://352802547-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-M7DcNFZmOW8qzXTiAei%2Fsync%2F9164a61bbea3676436f46857f4e1689169b110c4.png?generation=1589383541702300\&alt=media)\
对kernel有个反转操作，所以有的会生成对称的kernel，这样直接运算了。

### 参考佳文

[四大经典CNN网络技术原理](https://mp.weixin.qq.com/s?__biz=MzI5NTIxNTg0OA==\&mid=2247485440\&idx=1\&sn=054105f9731120426f6b4c8ca17a4b6f)

[CMU研究者探索新卷积方法：在实验中可媲美基准CNN（附实验代码](https://mp.weixin.qq.com/s/ybI8kJPRn7sH-hJbc5uqnw)

<https://arxiv.org/abs/1512.07108> Recent Advances in Convolutional Neural Networks

[Convolutional Neural Networks (CNNs): An Illustrated Explanation](http://xrds.acm.org/blog/2016/06/convolutional-neural-networks-cnns-illustrated-explanation/)\
[Convolutional Neural Networks (LeNet)](http://deeplearning.net/tutorial/lenet.html)\
[卷积神经网络（CNN）](http://www.cnblogs.com/charleshuang/p/3651843.html)\
[Deep learning](http://neuralnetworksanddeeplearning.com/chap6.html)\
[Convolution Neural Network (CNN) 原理与实现](http://blog.csdn.net/abcjennifer/article/details/25912675?utm_source=tuicool\&utm_medium=referral)\
[Deep Learning论文笔记之（四）CNN卷积神经网络推导和实现](http://blog.csdn.net/zouxy09/article/details/9993371)

[变形卷积核、可分离卷积？卷积神经网络中十大拍案叫绝的操作](https://zhuanlan.zhihu.com/p/28749411)\
[Deep Learning for Computer Vision – Introduction to Convolution Neural Networks](http://www.analyticsvidhya.com/blog/2016/04/deep-learning-computer-vision-introduction-convolution-neural-networks/)

[PCANet训练过程](http://www.voidcn.com/blog/u014365862/article/p-5785537.html)\
[【目标检测】Fast RCNN算法详解](http://blog.csdn.net/shenxiaolu1984/article/details/51036677)

[解读|Facebook 何凯明发大招：Mask R-CNN 狙击目标实例分割](https://mp.weixin.qq.com/s?__biz=MzA4NzE1NzYyMw==\&mid=2247488392\&idx=2\&sn=7c8e41aef37c370d6155607283d776ef)

[Focal Loss for Dense Object Detection](https://mp.weixin.qq.com/s/mj14otbEw_E1S23fiwdR8w)

理解卷积：\
[图像卷积与滤波的一些知识点](http://blog.csdn.net/zouxy09/article/details/49080029) 此文有用的信息量巨大\
[Lode's Computer Graphics Tutorial - Image Filtering](http://lodev.org/cgtutor/filtering.html)\
[卷积](https://zh.wikipedia.org/wiki/%E5%8D%B7%E7%A7%AF)\
[Understanding Convolutions](http://colah.github.io/posts/2014-07-Understanding-Convolutions/)\
[我对卷积的理解](http://mengqi92.github.io/2015/10/06/convolution/) 这个挺不错的

[\[CV\] 通俗理解『卷积』——从傅里叶变换到滤波器](https://zhuanlan.zhihu.com/p/28478034)\
[Conv Nets: A Modular Perspective](http://colah.github.io/posts/2014-07-Conv-Nets-Modular/)

[CNN 中， 1X1卷积核到底有什么作用呢？](http://blog.csdn.net/u014114990/article/details/50767786)\
[多通道(比如RGB三通道)卷积过程](http://blog.csdn.net/u014114990/article/details/51125776)

[Inception in CNN](http://blog.csdn.net/stdcoutzyx/article/details/51052847)

[Deep Residual Network 深度残差网络](https://zhuanlan.zhihu.com/p/22447440)\
[秒懂！何凯明的深度残差网络PPT是这样的](http://www.leiphone.com/news/201608/vhqwt5eWmUsLBcnv.html)

<http://geek.csdn.net/news/detail/129128>

<https://zhuanlan.zhihu.com/p/24774302> SPPNet-引入空间金字塔池化改进RCNN

[How do these "neural network style transfer" tools work?](https://jvns.ca/blog/2017/02/12/neural-style/)

[UC伯克利大学AI实验室用一张单色图像生成高质量3D几何结构](https://mp.weixin.qq.com/s/hOFbVBU_FJ4bkD4HWh-Lrg)