文章目錄
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
- 爲了構造identity mapping f(y) = y,因此作者對activation functions(BN和reLU)進行更改.那麼在forward或者backward的時候,信號都能直接propagate from 一個unit to other unit。
GoogLeNet
Inception V1
2014
Going deeper with convolutions
- 利用1x1的卷積解決維度爆炸
Inception V2
2015
v2:Batch Normalization: Accelerating Deep Network Training by ReducingInternal Covariate Shift
- Batch Normalization
- 用 2 個 3x3 的 conv 替代 Inception v1 模塊中的5x5
Inception V3
2015
v3:Rethinking the InceptionArchitecture for Computer Vision
- Asymmetric Convolutions
將7x7分解成兩個一維的卷積(1x7,7x1),3x3也是一樣(1x3,3x1) - 優化v1的auxiliary classifiers
- 新的pooling層
- Label smooth
Inception V4
2016
v4:Inception-v4,Inception-ResNet and the Impact of Residual Connections on Learning
- Inception模塊結合ResNet
Inception module來替換resnet shortcut中的bootlenect
Xception
2017
Xception: DeepLearning with Depthwise Separable Convolutions
Xception就是在 spatial dimensions , channel dimension 這2個變換上做文章。
-
depth-wise convolution
<img src=”https://img-blog.csdnimg.cn/20190924094637463.png"> -
借鑑(非採用)depth-wise convolution 改進 Inception V3(卷積的時候要將通道的卷積與空間的卷積進行分離)
-
原版 Depth-wise convolution,先逐通道 3×3 卷積,再 1×1 卷積
-
而 Xception 是反過來,先 1*1 卷積,再逐通道卷積.
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
優化激活函數(可用於其他網絡結構)
引入的基於squeeze and excitation結構的輕量級注意力模型
ShuffleNet
ShuffleNet V1
2017
ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
- 借鑑ResNet單元
- channel shuffle解決了多個group convolution疊加出現的邊界效應
- pointwise group convolution 和 depthwise separable convolution主要減少了計算量。
ShuffleNet V2
2018
ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
- 棄用了1x1的group convolution
- Channel Split:把特徵圖分成兩組A和B
- A組 認爲是short-cut;B組經過 bottleneck 輸入輸出channel一樣
- 最後concat A和B
- concat後進行Channel Shuffle
DenseNet
2017
Densely Connected Convolutional Networks
DPN
2017
Dual Path Networks
High Order RNN結構(HORNN)