文章目录
VGG
2014
Very Deep Convolutional Networks for Large-Scale Image Recognition
ResNet
2015
Deep Residual Learning for Image Recognition
- Residual Representations / Shortcut Connections
PreAct-ResNet
2016
Identity Mappings in Deep Residual Networks
- 为了构造identity mapping f(y) = y,因此作者对activation functions(BN和reLU)进行更改.那么在forward或者backward的时候,信号都能直接propagate from 一个unit to other unit。
GoogLeNet
Inception V1
2014
Going deeper with convolutions
- 利用1x1的卷积解决维度爆炸
Inception V2
2015
v2:Batch Normalization: Accelerating Deep Network Training by ReducingInternal Covariate Shift
- Batch Normalization
- 用 2 个 3x3 的 conv 替代 Inception v1 模块中的5x5
Inception V3
2015
v3:Rethinking the InceptionArchitecture for Computer Vision
- Asymmetric Convolutions
将7x7分解成两个一维的卷积(1x7,7x1),3x3也是一样(1x3,3x1) - 优化v1的auxiliary classifiers
- 新的pooling层
- Label smooth
Inception V4
2016
v4:Inception-v4,Inception-ResNet and the Impact of Residual Connections on Learning
- Inception模块结合ResNet
Inception module来替换resnet shortcut中的bootlenect
Xception
2017
Xception: DeepLearning with Depthwise Separable Convolutions
Xception就是在 spatial dimensions , channel dimension 这2个变换上做文章。
-
depth-wise convolution
<img src=”https://img-blog.csdnimg.cn/20190924094637463.png"> -
借鉴(非采用)depth-wise convolution 改进 Inception V3(卷积的时候要将通道的卷积与空间的卷积进行分离)
-
原版 Depth-wise convolution,先逐通道 3×3 卷积,再 1×1 卷积
-
而 Xception 是反过来,先 1*1 卷积,再逐通道卷积.
ResNeXt
2017
Aggregated ResidualTransformations for Deep Neural Networks
MobileNet
MobileNet V1
2017
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
Depthwise Separable Convolution
MobileNet V2
Inverted residuals
Linear bottlenecks
MobileNet V3
2019 CVPR
Searching for MobileNetV3
优化激活函数(可用于其他网络结构)
引入的基于squeeze and excitation结构的轻量级注意力模型
ShuffleNet
ShuffleNet V1
2017
ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
- 借鉴ResNet单元
- channel shuffle解决了多个group convolution叠加出现的边界效应
- pointwise group convolution 和 depthwise separable convolution主要减少了计算量。
ShuffleNet V2
2018
ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
- 弃用了1x1的group convolution
- Channel Split:把特征图分成两组A和B
- A组 认为是short-cut;B组经过 bottleneck 输入输出channel一样
- 最后concat A和B
- concat后进行Channel Shuffle
DenseNet
2017
Densely Connected Convolutional Networks
DPN
2017
Dual Path Networks
High Order RNN结构(HORNN)